Coverage

producing harness is absent from 1 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run.

Blocks

An artefact that names no harness cannot be traced to the thing that wrote it, so a reader has to take the file's word for its own origin.

Owner

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

Exact change

Change scripts/cotenancy/mig-h200-probe.py to write harness into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harness is null, undefined or empty in 1 of the suite's 2 normalised run(s)

Affected artefacts (1)

mig-h200-2026-08-03.json

measurement instant is absent from 2 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

date_utc is a calendar day, not a measurement instant.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write measuredAt into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

measuredAt is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

gpu-partition-mechanism@1 requires 2 mutually comparable run(s) and the check could not run at all: 2 of 2 run(s) carry no measurement instant, so ZERO pairs were compared.

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Any statement that this result reproduces, and equally any statement that it does not. Nothing was compared, so the answer here is unknown rather than no.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write a measurement instant into every artefact. Until it does, this row can never say more than "not checked", and re-running the harness more times will not change that.

How this row was computed

runs with measuredAt !== null = 0; pairs formed = 0; compareRuns() was never called

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

host is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

The machine behind the number is unrecoverable from the file.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

harness revision is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

The producing revision is unrecoverable from the file.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

measurement instant is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

This artefact is a bare array with no wrapper object; there is nothing to hang a timestamp on.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write measuredAt into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

measuredAt is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

ldpc-step-load@1 requires 2 mutually comparable run(s) and the check could not run at all: 2 of 2 run(s) carry no measurement instant, so ZERO pairs were compared.

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

Any statement that this result reproduces, and equally any statement that it does not. Nothing was compared, so the answer here is unknown rather than no.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write a measurement instant into every artefact. Until it does, this row can never say more than "not checked", and re-running the harness more times will not change that.

How this row was computed

runs with measuredAt !== null = 0; pairs formed = 0; compareRuns() was never called

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

producing harness is absent from 1 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Without it the analysis is not re-runnable.

Owner

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/score-detectors.py to write harness into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harness is null, undefined or empty in 1 of the suite's 2 normalised run(s)

Affected artefacts (1)

detector-scores-2026-08-03.json

measurement instant is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

This family writes no timestamp at all; the date in the filename is a filename.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write measuredAt into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

measuredAt is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

rf-channel-quality@1 requires 2 mutually comparable run(s) and the check could not run at all: 2 of 2 run(s) carry no measurement instant, so ZERO pairs were compared.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Any statement that this result reproduces, and equally any statement that it does not. Nothing was compared, so the answer here is unknown rather than no.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write a measurement instant into every artefact. Until it does, this row can never say more than "not checked", and re-running the harness more times will not change that.

How this row was computed

runs with measuredAt !== null = 0; pairs formed = 0; compareRuns() was never called

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

generator sha256 is absent from 14 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

The generator IS the test distribution here.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write identity.generatorSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.generatorSha256 is null, undefined or empty in 14 of the suite's 15 normalised run(s)

Affected artefacts (14)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

model label is absent from 14 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

At schema 2 the arm identity survives only in the filename; measured_against records the ENDPOINT, not the model.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write identity.modelLabel into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.modelLabel is null, undefined or empty in 14 of the suite's 15 normalised run(s)

Affected artefacts (14)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

sensing-generalization@1 requires 2 mutually comparable run(s). 105 pair(s) were compared and 0 run(s) came out with a partner.

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Any statement that this result reproduces. A single occasion is a single occasion however many files describe it.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to emit the identity fields the preflight needs, starting with dataset.hash, which blocked 105 of the 105 pair(s) checked, then re-run it 2 times. Repetition without identity does not count: the preflight is what decides, not the file count.

How this row was computed

compareRuns() over all 105 pair(s) of runs with a measurement instant; comparable requires every field in BASE_REQUIREMENTS present and equal; top blocker dataset.hash (absent, 105 pair(s))

Affected artefacts (15)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-DEPLOYED-seed4242.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

13 run(s) carry their seed only in the file name.

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Any statement about independent seeds. A file name is metadata a copy can change; it is not what the producer ran under.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write the seed, and the arm identity where the file name is currently carrying it, into the artefact body.

How this row was computed

identity.seedSource === 'filename' on 13 of 15 run(s)

Affected artefacts (13)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-DEPLOYED-seed4242.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json

split id is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without it the two accuracies may be over different held-out splits.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write facts.dataset.split_id into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

facts.dataset.split_id is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

preprocessing id is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without it the two accuracies may be computed after different preprocessing.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write facts.preprocessing.id into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

facts.preprocessing.id is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

producing harness is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

This family declares no harness; the runner exists only in a sibling markdown file.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write harness into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harness is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

checkpoint sha256 is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without it the artefact names no weights.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

dataset hash is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without it two runs at the same seed and sample count cannot be shown to have seen the same data. This project generates its data in-process, so the generator can move underneath a fixed seed.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

generator sha256 is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

The generator IS the data here; without its hash the distribution is unpinned.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write identity.generatorSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.generatorSha256 is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

measurement instant is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

This family writes no timestamp at all.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write measuredAt into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

measuredAt is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

heldout-accuracy@1 requires 3 mutually comparable run(s) and the check could not run at all: 6 of 6 run(s) carry no measurement instant, so ZERO pairs were compared.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Any statement that this result reproduces, and equally any statement that it does not. Nothing was compared, so the answer here is unknown rather than no.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write a measurement instant into every artefact. Until it does, this row can never say more than "not checked", and re-running the harness more times will not change that.

How this row was computed

runs with measuredAt !== null = 0; pairs formed = 0; compareRuns() was never called

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

tvws-selection-calibration@1 requires 2 mutually comparable run(s). 6 pair(s) were compared and 0 run(s) came out with a partner.

What is in the way
Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set. Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

Any statement that this result reproduces. A single occasion is a single occasion however many files describe it.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Change external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py to emit the identity fields the preflight needs, starting with dataset.hash, which blocked 6 of the 6 pair(s) checked, then re-run it 2 times. Repetition without identity does not count: the preflight is what decides, not the file count.

How this row was computed

compareRuns() over all 6 pair(s) of runs with a measurement instant; comparable requires every field in BASE_REQUIREMENTS present and equal; top blocker dataset.hash (absent, 6 pair(s))

Affected artefacts (4)

tvws-retrain-metrics-seed42-2026-08-03.json, tvws-retrain-metrics-seed7-2026-08-03.json, tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

tvws-ota-occupancy@1 requires 2 mutually comparable run(s). 3 pair(s) were compared and 0 run(s) came out with a partner.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

Any statement that this result reproduces. A single occasion is a single occasion however many files describe it.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to emit the identity fields the preflight needs, starting with dataset.hash, which blocked 3 of the 3 pair(s) checked, then re-run it 2 times. Repetition without identity does not count: the preflight is what decides, not the file count.

How this row was computed

compareRuns() over all 3 pair(s) of runs with a measurement instant; comparable requires every field in BASE_REQUIREMENTS present and equal; top blocker dataset.hash (absent, 3 pair(s))

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

training seed is absent from 2 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

The seed is not recorded in this artefact; anything shown came from the filename, which a rename loses.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write identity.seed into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.seed is null, undefined or empty in 2 of the suite's 4 normalised run(s)

Affected artefacts (2)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json

tvws-ota-per-epoch@1 requires 2 mutually comparable run(s). 6 pair(s) were compared and 0 run(s) came out with a partner.

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Any statement that this result reproduces. A single occasion is a single occasion however many files describe it.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to emit the identity fields the preflight needs, starting with dataset.hash, which blocked 6 of the 6 pair(s) checked, then re-run it 2 times. Repetition without identity does not count: the preflight is what decides, not the file count.

How this row was computed

compareRuns() over all 6 pair(s) of runs with a measurement instant; comparable requires every field in BASE_REQUIREMENTS present and equal; top blocker dataset.hash (absent, 6 pair(s))

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

2 run(s) carry their seed only in the file name.

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Any statement about independent seeds. A file name is metadata a copy can change; it is not what the producer ran under.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write the seed, and the arm identity where the file name is currently carrying it, into the artefact body.

How this row was computed

identity.seedSource === 'filename' on 2 of 4 run(s)

Affected artefacts (2)

ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

checkpoint sha256 is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Nothing binds the curve to the weights that shipped.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

seed is absent from 1 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

The arm exists only in the filename, which a rename loses.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write identity.seed into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.seed is null, undefined or empty in 1 of the suite's 3 normalised run(s)

Affected artefacts (1)

fixC-training-history-2026-07-30.json

measurement instant is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

These files carry none; they are hand-renamed copies of a checkpoint directory.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write measuredAt into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

measuredAt is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

training-history@1 requires 2 mutually comparable run(s) and the check could not run at all: 3 of 3 run(s) carry no measurement instant, so ZERO pairs were compared.

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Any statement that this result reproduces, and equally any statement that it does not. Nothing was compared, so the answer here is unknown rather than no.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write a measurement instant into every artefact. Until it does, this row can never say more than "not checked", and re-running the harness more times will not change that.

How this row was computed

runs with measuredAt !== null = 0; pairs formed = 0; compareRuns() was never called

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

2 run(s) carry their seed only in the file name.

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Any statement about independent seeds. A file name is metadata a copy can change; it is not what the producer ran under.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write the seed, and the arm identity where the file name is currently carrying it, into the artefact body.

How this row was computed

identity.seedSource === 'filename' on 2 of 3 run(s)

Affected artefacts (2)

perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

GPU is absent from 1 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Without a device block the GPU is unknowable from the file, so the number is unattributable to a machine.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write environment.gpu into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.gpu is null, undefined or empty in 1 of the suite's 2 normalised run(s)

Affected artefacts (1)

cuphy-per-slot-20260729-005456.json

cuphy-ldpc-decode@1 requires 3 mutually comparable run(s). 1 pair(s) were compared and 0 run(s) came out with a partner.

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Any statement that this result reproduces. A single occasion is a single occasion however many files describe it.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to emit the identity fields the preflight needs, starting with environment.profile, which blocked 1 of the 1 pair(s) checked, then re-run it 3 times. Repetition without identity does not count: the preflight is what decides, not the file count.

How this row was computed

compareRuns() over all 1 pair(s) of runs with a measurement instant; comparable requires every field in BASE_REQUIREMENTS present and equal; top blocker environment.profile (absent, 1 pair(s))

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

mps-cotenancy@1 requires 3 mutually comparable run(s) and the check could not run at all: 0 of 1 run(s) carry no measurement instant, so ZERO pairs were compared.

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Any statement that this result reproduces, and equally any statement that it does not. Nothing was compared, so the answer here is unknown rather than no.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Change scripts/cuphy-mps-sweep.py to write a measurement instant into every artefact. Until it does, this row can never say more than "not checked", and re-running the harness more times will not change that.

How this row was computed

runs with measuredAt !== null = 1; pairs formed = 0; compareRuns() was never called

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a GPU, a radio or an RF capture that GitHub-hosted runners do not have, so the evidence for this suite can only be produced by hand on a host that has them.

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Add a job that runs scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

harness revision is absent from 2 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

checkpoint sha256 is absent from 2 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

dataset hash is absent from 2 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

repo commit is absent from 2 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

workload hash is absent from 2 of 2 run(s).

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

Co-tenancy was never collected: none of the 2 artefact(s) carries the field at all.

What is in the way
Needs an H200 with MIG enabled and an operator willing to partition and tear it down again on a device this repository does not own and cannot schedule. This is a privileged one-way hardware reconfiguration, not a run. Needs to start its own MPS control daemon on the GPU. That is device-wide, so it cannot share the box with the running demo however careful the script is about its own processes.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owners

scripts/cotenancy/mig-h200-probe.py ✓ verified

Regenerates the observable half of this family and refuses the destructive half unless an operator explicitly asks and the GPU is free.

scripts/mps-cap-test.sh ✓ verified

Runs both clients inside a private MPS server to test whether an active-thread cap partitions the GB10.

Exact change

Change scripts/cotenancy/mig-h200-probe.py and scripts/mps-cap-test.sh to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 2 null

Affected artefacts (2)

mig-h200-2026-08-03.json, mps-cap-result.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a GPU, a radio or an RF capture that GitHub-hosted runners do not have, so the evidence for this suite can only be produced by hand on a host that has them.

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Add a job that runs scripts/cuphy-cotenancy-sweep.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

checkpoint sha256 is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

dataset hash is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

repo commit is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

workload hash is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

Co-tenancy was never collected: none of the 2 artefact(s) carries the field at all.

What is in the way
Needs the GPU to itself while a stepped AI load runs against cuPHY. The two artefacts are also a bare JSON array with no envelope, so the producer has to grow a wrapper object before a re-run is comparable with anything.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

scripts/cuphy-cotenancy-sweep.py ✓ verified

Steps an AI load against cuPHY LDPC and records one object per load level.

Exact change

Change scripts/cuphy-cotenancy-sweep.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 2 null

Affected artefacts (2)

window-clean-gpu-runA.json, window-clean-gpu-runB.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It reads an IQ capture that is not in the repository. No workflow in .github/workflows runs this producer, and nothing about the producer stops one: it is a CPU-only numpy job that finishes in about 13 seconds. What it needs is the 5 IQ captures under benchmarks/datasets/ota/, roughly 400 MB that this repository does not carry, so a GitHub runner has nothing to score.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Add a job that runs scripts/analyze-618mhz.py and scripts/score-detectors.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Without a host the number is unattributable to a machine.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

harness revision is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

checkpoint sha256 is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

dataset hash is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

repo commit is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

workload hash is absent from 2 of 2 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

Co-tenancy was never collected: none of the 2 artefact(s) carries the field at all.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture directory is in this tree and the producer runs here on CPU. What no re-run can widen: this is one channel at one site on one day, through an antenna that is not matched to the band. It is a CPU-only numpy job of about 13 seconds against captures already in this tree, and it is safe to run while the demo is up. Proven on 2026-08-04.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owners

scripts/analyze-618mhz.py ✓ verified

Analyses the 618 MHz capture: quality stages, a synthetic SNR ladder and a bandlimit sweep.

scripts/score-detectors.py ✓ verified

Runs the detector scorer and writes the artefact with a full provenance envelope.

Exact change

Change scripts/analyze-618mhz.py and scripts/score-detectors.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 2 null

Affected artefacts (2)

ch39-618mhz-quality-2026-07-31.json, detector-scores-2026-08-03.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs the sensing service running and a model checkpoint that is not in the repository.

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Add a job that runs scripts/sensing-generalization.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 15 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Without a host the number is unattributable to a machine.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 15 of the suite's 15 normalised run(s)

Affected artefacts (15)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-DEPLOYED-seed4242.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

harness revision is absent from 14 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 14 of the suite's 15 normalised run(s)

Affected artefacts (14)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

checkpoint sha256 is absent from 14 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 14 of the suite's 15 normalised run(s)

Affected artefacts (14)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

dataset hash is absent from 15 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 15 of the suite's 15 normalised run(s)

Affected artefacts (15)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-DEPLOYED-seed4242.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

repo commit is absent from 14 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 14 of the suite's 15 normalised run(s)

Affected artefacts (14)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

workload hash is absent from 15 of 15 run(s).

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 15 of the suite's 15 normalised run(s)

Affected artefacts (15)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-DEPLOYED-seed4242.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

Co-tenancy was never collected: none of the 15 artefact(s) carries the field at all.

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Change scripts/sensing-generalization.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 15 null

Affected artefacts (15)

sensing-generalization-new-control-seed4242.json, sensing-generalization-new-fixA-seed4242.json, sensing-generalization-new-fixB-seed4242.json, sensing-generalization-new-legacy-seed4242.json, sensing-generalization-old-control-seed4242.json, sensing-generalization-old-control-seed9137.json, sensing-generalization-old-fixA-seed4242.json, sensing-generalization-old-fixA-seed9137.json, sensing-generalization-old-fixB-DEPLOYED-seed4242.json, sensing-generalization-old-fixB-seed4242.json, sensing-generalization-old-fixB-seed9137.json, sensing-generalization-old-legacy-seed4242.json, sensing-generalization-old-legacy-seed9137.json, sensing-generalization-runA.json, sensing-generalization-runB.json

sensing-generalization@1 declares "Synthetic" a valid evidence tier and has no run at it.

What is in the way
Needs the sensing service, and on this box that endpoint is the one serving the live demo. A sweep against it is load on a running service, so it wants a second instance or a quiet window.

Blocks

Any claim that needs synthetic evidence. Measured on generated inputs. It can only ever describe the generator: a flaw the generator and the test share is invisible to it by construction.

Owner

scripts/sensing-generalization.py ✓ verified

Generates per-class samples and scores them through the deployed sensing endpoint.

Exact change

Run scripts/sensing-generalization.py under conditions that make the result synthetic evidence, and record the tier-defining conditions in the artefact.

How this row was computed

contract.validEvidenceTiers includes 'synthetic' AND no run in the suite has tier === 'synthetic'

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a model checkpoint, which is not in the repository.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Add a job that runs external/tvws-sensing/scripts/evaluate_heldout.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without a host the number is unattributable to a machine.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

harness revision is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

repo commit is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

workload hash is absent from 6 of 6 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 6 of the suite's 6 normalised run(s)

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

Co-tenancy was never collected: none of the 6 artefact(s) carries the field at all.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. torch 2.13.0+cpu is on this box and the checkpoint is in the tree at external/tvws-sensing/models/checkpoints/best.pt, so the producer runs here. It is a slow CPU run rather than a blocked one. What re-running still cannot give is a real held-out set: the split is regenerated from the seed at run time, so a re-run proves the recipe and never the data.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

external/tvws-sensing/scripts/evaluate_heldout.py ✓ verified

Scores a checkpoint against a freshly generated held-out split.

Exact change

Change external/tvws-sensing/scripts/evaluate_heldout.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 6 null

Affected artefacts (6)

heldout-committed-8eca022f-2026-07-31.json, heldout-committed-8eca022f-IMAGE-APP-SRC-trap-2026-07-31.json, heldout-committed-8eca022f-prefixgen-17517e57-2026-07-31.json, heldout-v1-828dc46e-gen-9dc4f892-2026-07-31.json, heldout-v1-828dc46e-headgen-c23711ab-2026-07-31.json, heldout-v1-828dc46e-prefixgen-17517e57-2026-07-31.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a GPU, a radio or an RF capture that GitHub-hosted runners do not have, so the evidence for this suite can only be produced by hand on a host that has them. No workflow in .github/workflows runs this producer. It needs a checkpoint directory and the OTA capture set.

What is in the way
Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs. Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Add a job that runs external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 4 of 4 run(s).

What is in the way
Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set. Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

Without a host the number is unattributable to a machine.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Change external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

tvws-retrain-metrics-seed42-2026-08-03.json, tvws-retrain-metrics-seed7-2026-08-03.json, tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

harness revision is absent from 4 of 4 run(s).

What is in the way
Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set. Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Change external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

tvws-retrain-metrics-seed42-2026-08-03.json, tvws-retrain-metrics-seed7-2026-08-03.json, tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

checkpoint sha256 is absent from 2 of 4 run(s).

What is in the way
Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 2 of the suite's 4 normalised run(s)

Affected artefacts (2)

tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

dataset hash is absent from 4 of 4 run(s).

What is in the way
Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set. Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Change external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

tvws-retrain-metrics-seed42-2026-08-03.json, tvws-retrain-metrics-seed7-2026-08-03.json, tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

repo commit is absent from 4 of 4 run(s).

What is in the way
Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set. Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Change external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

tvws-retrain-metrics-seed42-2026-08-03.json, tvws-retrain-metrics-seed7-2026-08-03.json, tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

Co-tenancy was never collected: none of the 4 artefact(s) carries the field at all.

What is in the way
Needs the checkpoint directory the artefact names in its own args.checkpoint_dir, which sits outside this repository. The set it reports is a SELECTION set, so no amount of re-running turns it into a test set. Needs a full training run on a GPU per seed. The two artefacts differ only by seed, so they are a variance probe rather than two measurements of one thing, and more seeds cost more training jobs.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owners

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN for a retrain round and writes the run record.

scripts/sensing-select-calibrate-ood.py ✓ verified

Selects a checkpoint, calibrates it, scores OOD behaviour and re-scores against the OTA captures.

Exact change

Change external/tvws-sensing/scripts/train_model.py and scripts/sensing-select-calibrate-ood.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 4 null

Affected artefacts (4)

tvws-retrain-metrics-seed42-2026-08-03.json, tvws-retrain-metrics-seed7-2026-08-03.json, tvws-retrain-seed42-2026-08-03.json, tvws-retrain-seed7-2026-08-03.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs the OTA capture set and a checkpoint, neither of which is on a GitHub runner.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Add a job that runs scripts/eval-ota-checkpoint.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 3 of 3 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

Without a host the number is unattributable to a machine.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

harness revision is absent from 3 of 3 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

The exact code that produced the number is unpinned.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

dataset hash is absent from 3 of 3 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

repo commit is absent from 3 of 3 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

workload hash is absent from 3 of 3 run(s).

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

Co-tenancy was never collected: none of the 3 artefact(s) carries the field at all.

What it takes
Nothing outside this repository is in the way. The producer is in this tree and runs on this box, so this gap closes by changing the producer and running it. The capture set and a checkpoint are both in this tree and the producer runs here on CPU torch. What no re-run can fix: every capture in the set is labelled occupied, so a false-vacant count is measurable and a false-occupied count is not.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

scripts/eval-ota-checkpoint.py ✓ verified

Scores a checkpoint against the real over-the-air captures using the service's own inference path.

Exact change

Change scripts/eval-ota-checkpoint.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 3 null

Affected artefacts (3)

ota-eval-fixB.json, ota-eval-fixC-final.json, ota-eval-fixC.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs the OTA capture set and a directory of checkpoints.

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Add a job that runs scripts/ota-per-epoch-sweep.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 4 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Without a host the number is unattributable to a machine.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

harness revision is absent from 4 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

checkpoint sha256 is absent from 4 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

dataset hash is absent from 4 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

repo commit is absent from 4 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

workload hash is absent from 4 of 4 run(s).

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 4 of the suite's 4 normalised run(s)

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

Co-tenancy was never collected: none of the 4 artefact(s) carries the field at all.

What is in the way
Needs a directory of per-epoch checkpoints that is not in this repository. Underneath that, the captures came through a SenseCAP 860-930 MHz antenna used at 470-700 MHz. A matched UHF antenna is a purchase and a new capture session, not a run, and until then every VACANT-CHANNEL and DYNAMIC-RANGE number in this family carries the mismatch. It does not carry the occupied-channel numbers: on the 2026-08-04 UHF survey the two strongest carriers in the band, 42 dB over the session floor, were classified Noise on 20 of 20 records at 0.969 and 0.957, so a matched antenna would not have recovered them. The occupied-direction failure is the scale-invariant input normalisation (docs/SENSING-OCCUPANCY-ROOT-CAUSE-2026-08-05.md), which a new antenna cannot touch.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

scripts/ota-per-epoch-sweep.py ✓ verified

Runs the OTA eval across every checkpoint in a directory, one row per epoch.

Exact change

Change scripts/ota-per-epoch-sweep.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 4 null

Affected artefacts (4)

ota-per-epoch-2026-07-30.json, ota-per-epoch-packed-2026-07-30.json, ota-per-epoch-packed-seed7-2026-07-30.json, ota-per-epoch-seed7-2026-07-30.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a GPU, a radio or an RF capture that GitHub-hosted runners do not have, so the evidence for this suite can only be produced by hand on a host that has them.

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Add a job that runs external/tvws-sensing/scripts/train_model.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Without a host the number is unattributable to a machine.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

harness revision is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

dataset hash is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

repo commit is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

workload hash is absent from 3 of 3 run(s).

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 3 of the suite's 3 normalised run(s)

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

Co-tenancy was never collected: none of the 3 artefact(s) carries the field at all.

What is in the way
Needs a full training run on a GPU. These are training curves rather than a benchmark, so re-running costs a training job and still yields a selection signal, not a test result.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

external/tvws-sensing/scripts/train_model.py ✓ verified

Trains the sensing CNN and dumps the per-epoch history.

Exact change

Change external/tvws-sensing/scripts/train_model.py to write an environment.co_tenants list, even an empty one, together with a note saying where it was collected from. Until it writes the field, absence and idleness are the same value here, and they are not the same thing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 0 non-empty, 3 null

Affected artefacts (3)

fixC-training-history-2026-07-30.json, perepoch-history-seed42-2026-07-30.json, perepoch-history-seed7-2026-07-30.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a GPU, a radio or an RF capture that GitHub-hosted runners do not have, so the evidence for this suite can only be produced by hand on a host that has them.

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Add a job that runs scripts/cuphy-per-slot-latency.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

host is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Without a host the number is unattributable to a machine.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write environment.host into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

environment.host is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

harness revision is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

The producing code is unpinned; a change in it is unattributable.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

checkpoint sha256 is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

dataset hash is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

repo commit is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

workload hash is absent from 2 of 2 run(s).

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 2 of the suite's 2 normalised run(s)

Affected artefacts (2)

cuphy-per-slot-20260729-005456.json, cuphy-per-slot-20260730-080736.json

No verifiable clean-host run. 1 run(s) report an empty co-tenant list that the artefact itself says it could not see past.

What is in the way
Needs the GPU to itself. The harness re-execs into the Aerial cuBB container and times one LDPC decode at a time, so anything else on the device moves the numbers. On this box that means draining the demo first and taking a scheduled window, not squeezing a run in beside it.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

scripts/cuphy-per-slot-latency.py ✓ verified

Times one pyaerial LdpcDecoder.decode() call per sample with cudaEvent, inside the Aerial container, and writes the per-slot artefact to benchmarks/reports/.

Exact change

Change scripts/cuphy-per-slot-latency.py to collect co-tenancy from the HOST, with nvidia-smi outside the container rather than inside it, and to record what the collection could and could not see. An empty list taken from inside a container does not mean the machine was idle, and this suite's own artefacts say so.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 1 empty but qualified, 0 non-empty, 1 null

Affected artefacts (1)

cuphy-per-slot-20260730-080736.json

No workflow in .github/workflows produces this suite's evidence. No workflow in .github/workflows runs this producer. It needs a GPU, a radio or an RF capture that GitHub-hosted runners do not have, so the evidence for this suite can only be produced by hand on a host that has them.

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Any statement that this result is continuously verified. It is produced by hand, on a host with the hardware, when somebody remembers.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Add a job that runs scripts/cuphy-mps-sweep.py on a runner that has what it needs, or state in the contract that this suite is hand-produced by design.

How this row was computed

every family in this suite has PRODUCER_MAP[family].ci === null. The map is hand-maintained, so this row is the map's claim about .github/workflows and not a scan of it — the workflow files are in the tree and can be read directly.

harness revision is absent from 1 of 1 run(s).

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Without the producing revision, a difference between two runs may be a change in the measurement rather than a change in the thing measured.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Change scripts/cuphy-mps-sweep.py to write harnessRevision into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

harnessRevision is null, undefined or empty in 1 of the suite's 1 normalised run(s)

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json

checkpoint sha256 is absent from 1 of 1 run(s).

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Without it the weights under test are unidentified. With it, the file is pinned — and still not bound to any endpoint.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Change scripts/cuphy-mps-sweep.py to write identity.checkpointSha256 into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.checkpointSha256 is null, undefined or empty in 1 of the suite's 1 normalised run(s)

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json

dataset hash is absent from 1 of 1 run(s).

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

The same seed and the same sample count are not the same data. Without a hash, two runs are asserted to share inputs, never shown to.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Change scripts/cuphy-mps-sweep.py to write identity.datasetHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.datasetHash is null, undefined or empty in 1 of the suite's 1 normalised run(s)

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json

repo commit is absent from 1 of 1 run(s).

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Without a commit, the code that produced the number cannot be recovered.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Change scripts/cuphy-mps-sweep.py to write identity.repoCommit into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.repoCommit is null, undefined or empty in 1 of the suite's 1 normalised run(s)

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json

workload hash is absent from 1 of 1 run(s).

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Without a workload hash the two runs are asserted to answer the same question.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Change scripts/cuphy-mps-sweep.py to write identity.workloadHash into every artefact it produces, then re-run. Backfilling it by hand into the existing files would be a fabrication: the value was never recorded.

How this row was computed

identity.workloadHash is null, undefined or empty in 1 of the suite's 1 normalised run(s)

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json

Every one of the 1 run(s) here records other tenants; none records an idle machine.

What is in the way
Needs a scheduled clean-GPU window and an MPS control daemon, which is a device-wide object rather than a process you can start beside the demo. The producer's own docstring asks for the window in those words.

Blocks

Any statement that the number was measured without interference from other tenants on the machine.

Owner

scripts/cuphy-mps-sweep.py ✓ verified

Sweeps an MPS active-thread cap on an AI client while cuPHY LDPC runs uncapped.

Exact change

Run scripts/cuphy-mps-sweep.py once on a drained machine and keep that artefact beside these, so the co-tenancy numbers have an idle baseline to be read against. Nothing in the existing files needs editing.

How this row was computed

environment.coTenants across the suite: 0 empty and unqualified, 0 empty but qualified, 1 non-empty, 0 null

Affected artefacts (1)

mps-sweep-2026-08-03-livebox.json