Model review audits
Compare · P25 firmware review

Head to head

Same brief, same rubric, same scale. Put up to three models side by side: dimension scores, how many of their claims held up, and which issues each review raised.

A GLM 5.3 Flash F D C B A D 56 / 100
rank 4 of 4 · graded 17 Sep 2026

Audits the memory to the byte. Misreads the receive path.

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.

 B Grok 4.6 F D C B A C 70 / 100
rank 1 of 4 · graded 17 Sep 2026

Finds what erases the signal. Misses what stalls the voice.

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

Best on memory, last on the rest

All four clean-room audits miss the vocoder CPU wall. GLM 5.3 Flash is the only one to audit memory in detail, and one of three to name de-emphasis, but it finds one decode-critical issue where Grok 4.6 finds three and DeepSeek V4.1 Flash and Muse Spark 1.3 Contributor two each, and it makes more wrong claims about the receive path than any of them.

From the GLM 5.3 Flash audit, 17 September 2026.

Weighted score

Out of 100, on the grade scale.

4 GLM 5.3 FlashZ.ai · graded 17 Sep 2026 56D 1 Grok 4.6xAI · graded 17 Sep 2026 70C
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.

GLM 5.3 Flash
56/ 100
Grok 4.6
70/ 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

Each dimension out of 100, then the weighted total, over the grade zones. The compared models are the coloured dots, each joined down the dimensions by a thin line; the other models on this task are the small grey ones. Hover or focus a row for every score and the auditor’s reasoning.

  • AGLM 5.3 Flash
  • BGrok 4.6
  • 2 other models on this task
Accuracy & evidence30%Coverage of decode problems25%Root cause & prioritisation15%Fix plan & acceptance gates15%Originality & attribution10%Clarity & calibration5%Weighted total100%FDCBA050658090100 Accuracy & evidence30%Coverage of decode problems25%Root cause & prioritisation15%Fix plan & acceptance gates15%Originality & attribution10%Clarity & calibration5%Weighted total100%FDCBA
Scores and weighted points
Dimension Weight A · GLM 5.3 FlashB · Grok 4.6Muse Spark 1.3 ContributorDeepSeek V4.1 Flash A − B
Accuracy & evidence 30% 64 · 19.2 pts 74 · 22.2 pts 74 · 22.2 pts 73 · 21.9 pts −10
Coverage of decode problems 25% 34 · 8.5 pts 59 · 14.8 pts 54 · 13.5 pts 49 · 12.3 pts −25
Root cause & prioritisation 15% 62 · 9.3 pts 70 · 10.5 pts 72 · 10.8 pts 66 · 9.9 pts −8
Fix plan & acceptance gates 15% 62 · 9.3 pts 75 · 11.3 pts 71 · 10.7 pts 70 · 10.5 pts −13
Originality & attribution 10% 58 · 5.8 pts 72 · 7.2 pts 70 · 7.0 pts 74 · 7.4 pts −14
Clarity & calibration 5% 68 · 3.4 pts 78 · 3.9 pts 80 · 4.0 pts 74 · 3.7 pts −10
Weighted total 100% 56 · D70 · C68 · C−66 · C− −14
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.

GLM 5.3 Flash’s claim table · Grok 4.6’s claim table

  • Holds
  • Qualifiedoverstated, miscounted or doubtful
  • Wrong
GLM 5.3 Flash28 claims checked

17 hold · 7 qualified · 4 wrong · 61% hold

Grok 4.630 claims checked

21 hold · 7 qualified · 2 wrong · 70% hold

Issue coverage

What each review raised

GLM 5.3 Flash found 1 of the 4 decode-critical issues; Grok 4.6 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 Project docs earlier repo review · in the repository Grok 4.6 model · Sep 17 GLM 5.3 Flash 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 Grok 4.6Model · Sep 17 GLM 5.3 FlashModel · Sep 17
1 Capture export stages the wrong I2S halfwords absent raised§1, first code change absentrelies on the capture
2 Capture parser certifies an incomplete stream absent absent absentrelies on the parser
3 pcm_starve never increments absentdocumented as working absent absent
4 No static RAM margin partlymargins still to measure absent raisedF5, byte-exact map audit
5 The I2S stream is most likely microphone audio Decode-criticalmissed partlyopen, leaning sceptical raised§1, 0x89 versus 0xC9 partlyopen, leans towards RF
6 HR-C6000 de-emphasis on the capture path Decode-critical absent raised§2, bit 5 of 0x34 raisedF2, closes the eye
7 AT1846S FM filters, low-frequency bit, 25 kHz Decode-critical partly“require characterization” raised§2, register level partly“voice filtering”; wrong bandwidth premise
8 0x10=0x6E hybrid state; 0x36 dual role partlybring-up clock rules partlymisses 0x6E and the 0x36 clock gate raisedF4, 0x10=0x80 kills the clock
9 Manual: I2S frame clock “must be 8KHz” Decode-criticalmissed absent raised§3 absent
10 One-layer 4FSK test mode as a P25 tap partlystock BER-test block only raisedGate D absent
11 ±10% health gate versus ±1% timing clamp absent absent raisedF3
12 Fail-closed muting at LDU cadence absent absentlate-entry mute only absent
13 Non-standard MFID mutes clear calls absent absent absent
14 Test waveform shares the receiver’s RRC filter partly“synthetic RRC/AWGN” caveat absent absentwould extend that model
15 MCU runs at 72 MHz raised raisedin passing absent
16 Vocoder needs 11–16× the 72 MHz CPU Decode-criticalmissed partlydecode timing unmeasured absent“fine on a 1 ms tick” absent“in good shape”
17 Direct discriminator tap (M17 mod) raised raiseduncredited raisedfallback
18 Phase 2 architecture and scope partlynot implemented partlymisplaces the AMBE+2 decoder partlysays mbelib has no AMBE+2
19 Two unverified SPI writes per decoded 20 ms frame absent absent absenttreats them as protection
20 Capture sessions lack epochs absent absent absent
21 Ring and tick real-time budget partlydeadlines unproven partlycalls it fine absent
22 Clock config 3 assumes 12,288 Hz; the codec formula gives 12,000 absent absent partly“guessed semantics”
23 Clock-config writes bypass the verified SPI writer partlySPI retry note absent absent

From the provenance table in the GLM 5.3 Flash audit (17 Sep 2026), the newest audit on this task. Shaded columns are reviews that were already in the repository.