Model review audits
Compare · Silent Carrier

Head to head

Same brief, same rubric, same scale. Every model on the task side by side, or just the ones you pick: dimension scores, how many of their claims held up, and which issues each review raised.

Models 2 of 16
All 16 models →
Claude Opus 5 F D C B A A 92 / 100
rank 1 of 16 · graded 17 Sep 2026

It measured what everyone argued about.

Clean room: a fresh copy of the repository (46eebda with its uncommitted changes, its own docs and the HR-C6000 manual) with no earlier reviews, reviewed by one model with no network tools, no skills, no plugins, no MCP servers and no subagents.

 GPT-6 Astra F D C B A B 89 / 100
rank 2 of 16 · graded 17 Sep 2026

Reproduced, not asserted.

Clean room: a fresh copy of the repository (46eebda with its uncommitted changes, its own docs and the HR-C6000 manual) with no earlier reviews, reviewed by one model through an isolated agent harness with no network, skills or memory.
From the audits

The first to find all four, and the only one that measured them

Sixteen models have now reviewed this repository. The fourteen before this one all missed the vocoder CPU wall; this one profiled it in an emulator and tied it to the parser reset that ends the call. It also found both halves of the analog chain, the sample source and the frame-clock rule. GPT-6 Astra remains the best of the ranked audits and the one that found what this review skipped: the instruments.

From the Claude Opus 5 audit, 17 September 2026.

Weighted score

Out of 100, on the grade scale.

1 Claude Opus 5Anthropic · graded 17 Sep 2026 92A 2 GPT-6 AstraOpenAI · graded 17 Sep 2026 89B
Where the points went

Earned and lost, by rubric slot

Each bar is a model’s 100-point grade split into the rubric’s weighted slots; filled is earned, empty is lost. The slots line up, so a gap in one bar against a full slot in another is exactly where the models part ways.

Claude Opus 5
92/ 100
GPT-6 Astra
89/ 100
  • 1Accuracy & evidence30 pts
  • 2Coverage of decode problems25 pts
  • 3Root cause & prioritisation15 pts
  • 4Fix plan & acceptance gates15 pts
  • 5Originality & attribution10 pts
  • 6Clarity & calibration5 pts
Dimension profile

Scores on each rubric dimension

One panel per model: its score on each rubric dimension and the weighted total over the grade zones, in its own colour, with every other model on this task as a grey dot for scale. Hover a row for every model’s score and the auditor’s reasoning.

  • The panel’s model, joined down the dimensions
  • Every other model on this task

Claude Opus 592 A

AccuracyCoverageRoot causeFix planOriginalityClarityTotalFDCBA AccuracyCoverageRoot causeFix planOriginalityClarityTotalFDCBA

GPT-6 Astra89 B

AccuracyCoverageRoot causeFix planOriginalityClarityTotalFDCBA AccuracyCoverageRoot causeFix planOriginalityClarityTotalFDCBA
Scores and weighted points
Dimension Weight Claude Opus 5GPT-6 AstraGPT-5.6 SolClaude Fable 5.1HY4 PreviewGPT-5.6 LunaUNIONALPHAGrok 4.6Muse Spark 1.3 ContributorDeepSeek V4.1 FlashGLM 5.3Gemini 3.8 FlashQwen3.8 MaxQwen3.8 FlashGLM 5.3 FlashDeepSeek V4 Pro Difference
Accuracy & evidence 30% 94 · 28.2 pts 92 · 27.6 pts 91 · 27.3 pts 85 · 25.5 pts 85 · 25.5 pts 82 · 24.6 pts 90 · 27.0 pts 74 · 22.2 pts 74 · 22.2 pts 73 · 21.9 pts 68 · 20.4 pts 65 · 19.5 pts 70 · 21.0 pts 62 · 18.6 pts 64 · 19.2 pts 61 · 18.3 pts +2
Coverage of decode problems 25% 86 · 21.5 pts 84 · 21.0 pts 70 · 17.5 pts 62 · 15.5 pts 56 · 14.0 pts 58 · 14.5 pts 65 · 16.3 pts 59 · 14.8 pts 54 · 13.5 pts 49 · 12.3 pts 52 · 13.0 pts 46 · 11.5 pts 39 · 9.8 pts 40 · 10.0 pts 34 · 8.5 pts 33 · 8.3 pts +2
Root cause & prioritisation 15% 96 · 14.4 pts 88 · 13.2 pts 84 · 12.6 pts 84 · 12.6 pts 80 · 12.0 pts 80 · 12.0 pts 60 · 9.0 pts 70 · 10.5 pts 72 · 10.8 pts 66 · 9.9 pts 74 · 11.1 pts 74 · 11.1 pts 71 · 10.7 pts 68 · 10.2 pts 62 · 9.3 pts 64 · 9.6 pts +8
Fix plan & acceptance gates 15% 93 · 14.0 pts 92 · 13.8 pts 90 · 13.5 pts 86 · 12.9 pts 84 · 12.6 pts 80 · 12.0 pts 70 · 10.5 pts 75 · 11.3 pts 71 · 10.7 pts 70 · 10.5 pts 68 · 10.2 pts 70 · 10.5 pts 71 · 10.7 pts 70 · 10.5 pts 62 · 9.3 pts 60 · 9.0 pts +1
Originality & attribution 10% 95 · 9.5 pts 90 · 9.0 pts 84 · 8.4 pts 84 · 8.4 pts 82 · 8.2 pts 72 · 7.2 pts 76 · 7.6 pts 72 · 7.2 pts 70 · 7.0 pts 74 · 7.4 pts 70 · 7.0 pts 70 · 7.0 pts 66 · 6.6 pts 64 · 6.4 pts 58 · 5.8 pts 62 · 6.2 pts +5
Clarity & calibration 5% 92 · 4.6 pts 86 · 4.3 pts 86 · 4.3 pts 88 · 4.4 pts 88 · 4.4 pts 84 · 4.2 pts 80 · 4.0 pts 78 · 3.9 pts 80 · 4.0 pts 74 · 3.7 pts 72 · 3.6 pts 84 · 4.2 pts 78 · 3.9 pts 70 · 3.5 pts 68 · 3.4 pts 48 · 2.4 pts +6
Weighted total 100% 92 · A89 · B84 · B79 · C+77 · C+75 · C+74 · C70 · C68 · C−66 · C−65 · C−64 · D63 · D59 · D56 · D54 · D +3
Claim check

Every claim, checked

One cell per claim in each audit’s claim table, checked at the lines it cites. Hover a cell for the claim and the auditor’s note. The reviews made different numbers of claims, so the bar under each grid gives the shares.

Claude Opus 5’s claim table · GPT-6 Astra’s claim table

  • Holds
  • Qualifiedoverstated, miscounted or doubtful
  • Wrong
Claude Opus 521 claims checked

20 hold · 1 qualified · 0 wrong · 95% hold

GPT-6 Astra29 claims checked

27 hold · 2 qualified · 0 wrong · 93% hold

Issue coverage

What each review raised

Claude Opus 5 found 4 of the 4 decode-critical issues; GPT-6 Astra found 3 of the 4 decode-critical issues. Every tracked issue below, with the reviews that were already in the repository for reference.

Issue signatures

Each review as a trace across the tracked issues. Where traces part is where the reviews differ; hover a column for every review’s entry. Column numbers match the table below.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 Project docs earlier repo review · in the repository GPT-6 Astra model · Sep 17 Claude Opus 5 model · Sep 17
  • raised
  • partly
  • absent
  • Decode-critical
Issues raised by each review of P25 voice decoding on the DM-1701
# Issue Project docsEarlier repo review · in the repository GPT-6 AstraModel · Sep 17 Claude Opus 5Model · Sep 17
1 Capture export stages the wrong I2S halfwords absent raisedF4, reproduced on the host absent
2 Capture parser certifies an incomplete stream absent partlyasks for tests at the callback absent
3 pcm_starve never increments absentdocumented as working raisedF9, reproduced absent
4 No static RAM margin partlymargins still to measure raisedF8, from its own link map raisedwith the overlay buffers named
5 The I2S stream is most likely microphone audio Decode-critical partlyopen, leaning sceptical raisedF2, as far as the evidence goes raisedproved from 0xE0=0x89 and sound.c
6 HR-C6000 de-emphasis on the capture path Decode-critical absent raisedF2, 0x34=0x3C raisedmeasured at 48.9% SER
7 AT1846S FM filters, low-frequency bit, 25 kHz Decode-critical partly“require characterization” raisedF2, the FM settings table raisedall three, with the DMR fix
8 0x10=0x6E hybrid state; 0x36 dual role partlybring-up clock rules partly0x36 and 0x10 via the squelch path partly0x36 and 0xE0, not 0x10
9 Manual: I2S frame clock “must be 8KHz” Decode-criticalmissed or ruled out absent raisedF7, with the divisor arithmetic raisedH2
10 One-layer 4FSK test mode as a P25 tap missed or ruled out partlystock BER-test block only partlypoints at the layer architecture raisedRoute A, with an acceptance test
11 ±10% health gate versus ±1% timing clamp absent raisedF7 absent
12 Fail-closed muting at LDU cadence absent raisedF6, the late-entry half partlyvia the MFID path
13 Non-standard MFID mutes clear calls absent absent raisedM1, with the Motorola case
14 Test waveform shares the receiver’s RRC filter partly“synthetic RRC/AWGN” caveat raisedF10, with why the tests missed F1/F4 raisedtested, called harmless
15 MCU runs at 72 MHz raised raisedF9, the target for measurement raisedand measured against it
16 Vocoder needs 11–16× the 72 MHz CPU Decode-criticalmissed or ruled out partlydecode timing unmeasured partlycounted the filter, not the vocoder raisedmeasured 6.5× on the demo build
17 Direct discriminator tap (M17 mod) raised raisedF2, ADC or another interface raisedRoute B, with the ADC channels checked
18 Phase 2 architecture and scope partlynot implemented raisedthe most complete of the fourteen raisedincluding the descrambler seed
19 Two unverified SPI writes per decoded 20 ms frame absent partlythe sink is cited, not the writes absent
20 Capture sessions lack epochs absent raisedF4, segment on retune and clock change absent
21 Ring and tick real-time budget partlydeadlines unproven raisedF9, all four figures exact raisedthe 21 ms ring against a 55–190 ms stall
22 Clock config 3 assumes 12,288 Hz; the codec formula gives 12,000 absent partlycomputes config 2 instead absent
23 Clock-config writes bypass the verified SPI writer partlySPI retry note partlyasks that writes be verified partlycalls the heal machinery fragile
24 Stock squelch re-arms FM audio (0x10=0x80) during monitoring absentassumes it can’t re-arm raisedF1, reproduced on the host absent
25 Stale clear-call state releases a new call’s first frames absent partlyvia identity, not the 1,120 samples absent
26 No frequency tracking; ad-hoc timing loop gains partlya code comment calls the DC estimate biased absent raisedDC fit once per window, with the fix
27 Unknown talkgroup opens audio (fail-open gating) absent raisedF6, new and reproduced absent

From the provenance table in the Claude Fable 5.1 audit (17 Sep 2026), the newest audit on this task. Shaded columns are reviews that were already in the repository.

Run facts

How each run went

Descriptive, not graded, and in grade order rather than ranked: one run per model. Run time is wall-clock, so it includes the provider’s speed, rate-limit waits and retries, and token counts vary with each model’s tokenizer. Every run followed Clean Room 1.0, which is what makes them comparable.

Run facts for each compared model, in grade order. Descriptive only: not graded or ranked.
Model Run time Agent steps Tool calls Output tokens Route Launch
Claude Opus 5 25 min 120 54 185K Claude Code CLI, subscription one-shot
GPT-6 Astra 15 min 29 42 21K OpenAI subscription one-shot
Every run figure
Model Started (UTC) Output tokens Reasoning tokens Context tokens Messages Sessions Retry budget
Claude Opus 5 18 Sep 2026, 02:44 185,228 118,897 17,469,122 – 1 –
GPT-6 Astra 18 Sep 2026, 01:06 21,399 3,777 3,081,430 72 1 10

Context tokens are input plus cache reads and writes: mostly the conversation re-sent at every step, so they grow with steps × context size. The retry budget is how many API retries the harness allowed.