Compare models

All five clean-room audits miss the vocoder CPU wall. UNIONALPHA makes no wrong claims where Grok 4.6, DeepSeek V4.1 Flash and Muse Spark 1.3 Contributor make two or three each, and it covers the most issues, but it fully finds one decode-critical issue where Grok 4.6 finds three.

From the UNIONALPHA audit, 17 September 2026.

Weighted score

Out of 100, on the grade scale.

1 UNIONALPHAStealth model · graded 17 Sep 2026 74C 2 Grok 4.6xAI · graded 17 Sep 2026 70C 3 Muse Spark 1.3 ContributorMeta · graded 17 Sep 2026 68C− 4 DeepSeek V4.1 FlashDeepSeek · graded 17 Sep 2026 66C− 5 Qwen3.8 MaxAlibaba · graded 17 Sep 2026 63D 6 Qwen3.8 FlashAlibaba · graded 17 Sep 2026 59D 7 GLM 5.3 FlashZ.ai · graded 17 Sep 2026 56D

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.

UNIONALPHA

74/ 100

Grok 4.6

70/ 100

Muse Spark 1.3 Contributor

68/ 100

DeepSeek V4.1 Flash

66/ 100

Qwen3.8 Max

63/ 100

Qwen3.8 Flash

59/ 100

GLM 5.3 Flash

56/ 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

UNIONALPHA74 C

Grok 4.670 C

Muse Spark 1.3 Contributor68 C−

DeepSeek V4.1 Flash66 C−

Qwen3.8 Max63 D

Qwen3.8 Flash59 D

GLM 5.3 Flash56 D

Scores and weighted points

Dimension	Weight	UNIONALPHA	Grok 4.6	Muse Spark 1.3 Contributor	DeepSeek V4.1 Flash	Qwen3.8 Max	Qwen3.8 Flash	GLM 5.3 Flash
Accuracy & evidence	30%	90 · 27.0 pts	74 · 22.2 pts	74 · 22.2 pts	73 · 21.9 pts	70 · 21.0 pts	62 · 18.6 pts	64 · 19.2 pts
Coverage of decode problems	25%	65 · 16.3 pts	59 · 14.8 pts	54 · 13.5 pts	49 · 12.3 pts	39 · 9.8 pts	40 · 10.0 pts	34 · 8.5 pts
Root cause & prioritisation	15%	60 · 9.0 pts	70 · 10.5 pts	72 · 10.8 pts	66 · 9.9 pts	71 · 10.7 pts	68 · 10.2 pts	62 · 9.3 pts
Fix plan & acceptance gates	15%	70 · 10.5 pts	75 · 11.3 pts	71 · 10.7 pts	70 · 10.5 pts	71 · 10.7 pts	70 · 10.5 pts	62 · 9.3 pts
Originality & attribution	10%	76 · 7.6 pts	72 · 7.2 pts	70 · 7.0 pts	74 · 7.4 pts	66 · 6.6 pts	64 · 6.4 pts	58 · 5.8 pts
Clarity & calibration	5%	80 · 4.0 pts	78 · 3.9 pts	80 · 4.0 pts	74 · 3.7 pts	78 · 3.9 pts	70 · 3.5 pts	68 · 3.4 pts
Weighted total	100%	74 · C	70 · C	68 · C−	66 · C−	63 · D	59 · D	56 · D

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.

UNIONALPHA’s claim table · Grok 4.6’s claim table · Muse Spark 1.3 Contributor’s claim table · DeepSeek V4.1 Flash’s claim table · Qwen3.8 Max’s claim table · Qwen3.8 Flash’s claim table · GLM 5.3 Flash’s claim table

Holds
Qualifiedoverstated, miscounted or doubtful
Wrong

UNIONALPHA35 claims checked

33 hold · 2 qualified · 0 wrong · 94% hold

Grok 4.630 claims checked

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

Muse Spark 1.3 Contributor30 claims checked

22 hold · 6 qualified · 2 wrong · 73% hold

DeepSeek V4.1 Flash35 claims checked

28 hold · 4 qualified · 3 wrong · 80% hold

Qwen3.8 Max27 claims checked

21 hold · 4 qualified · 2 wrong · 78% hold

Qwen3.8 Flash33 claims checked

21 hold · 6 qualified · 6 wrong · 64% hold

GLM 5.3 Flash28 claims checked

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

Issue coverage

What each review raised

UNIONALPHA found 1 of the 4 decode-critical issues; Grok 4.6 found 3 of the 4 decode-critical issues; Muse Spark 1.3 Contributor found 2 of the 4 decode-critical issues; DeepSeek V4.1 Flash found 2 of the 4 decode-critical issues; Qwen3.8 Max found 2 of the 4 decode-critical issues; Qwen3.8 Flash found 2 of the 4 decode-critical issues; GLM 5.3 Flash found 1 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.

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	DeepSeek V4.1 FlashModel · Sep 17	Muse Spark 1.3 ContributorModel · Sep 17	GLM 5.3 FlashModel · Sep 17	UNIONALPHAModel · Sep 17	Qwen3.8 FlashModel · Sep 17	Qwen3.8 MaxModel · Sep 17
1	Capture export stages the wrong I2S halfwords	absent	raised§1, first code change	absentits first experiment uses that capture	absenttrusts the capture	absentrelies on the capture	raisedF3, fixed and verified	raisedP1-2, tests miss the adapter	absentcalls the capture validated
2	Capture parser certifies an incomplete stream	absent	absent	absent	absenttrusts the parser	absentrelies on the parser	absent	absent	absentrelies on the parser
3	pcm_starve never increments	absentdocumented as working	absent	absent	absentrelies on it	absent	raisedF7	raisedP1-7	absent
4	No static RAM margin	partlymargins still to measure	absent	raisedM6, 0 bytes free	partly“nearly full”, from the docs	raisedF5, byte-exact map audit	raisedF6, byte-exact	partlycalls the reservation headroom	absent
5	The I2S stream is most likely microphone audio Decode-critical	partlyopen, leaning sceptical	raised§1, 0x89 versus 0xC9	raisedB1, four-value test	raisedP1-1	partlyopen, leans towards RF	partlyblocking, but never says microphone	raisedP1-4, 0xE0 mic bit	raisedA, the clearest case yet
6	HR-C6000 de-emphasis on the capture path Decode-critical	absent	raised§2, bit 5 of 0x34	absentcalls it benign	raisedP1-2, 0x34=0x3C	raisedF2, closes the eye	partlycited, called unmeasured	raisedP1-3, 0x34=0x1C	raisedC, with the 0x34 bit 5 fix
7	AT1846S FM filters, low-frequency bit, 25 kHz Decode-criticalmissed	partly“require characterization”	raised§2, register level	raisedB2, filter register	raisedP1-2, 0x58 filters	partly“voice filtering”; wrong bandwidth premise	partlycited, called unmeasured	raisedP1-3, DMR 0x58 probe	absent
8	0x10=0x6E hybrid state; 0x36 dual role	partlybring-up clock rules	partlymisses 0x6E and the 0x36 clock gate	partly“undocumented hybrid state”	partlyquiet-chip registers	raisedF4, 0x10=0x80 kills the clock	partlyF1, 0x6E against 0x80	partly“hybrid I2S state”	partlyslot engine and eco only
9	Manual: I2S frame clock “must be 8KHz” Decode-criticalmissed	absent	raised§3	absentquotes the paragraph, not the rule	absent	absent	raisedF4	absentquotes the formulas, not the rule	absentcites the section, not the rule
10	One-layer 4FSK test mode as a P25 tap	partlystock BER-test block only	raisedGate D	raisedB5, exact recipe	absentdismissed	absent	partlyworth a bounded test	absentruled out at “9600 Bd”	raisedstep 2, a symbol source
11	±10% health gate versus ±1% timing clamp	absent	absent	raisedB3, impact overstated	raisedP1-3	raisedF3	absent	absent	absent
12	Fail-closed muting at LDU cadence	absent	absentlate-entry mute only	absent	absentcalls it an asset	absent	absent	partly“keep it”	absent
13	Non-standard MFID mutes clear calls	absent	absent	absent	absent	absent	absent	absent	absent
14	Test waveform shares the receiver’s RRC filter	partly“synthetic RRC/AWGN” caveat	absent	partlytested it, says not to fix	partlysynthetic only, wants recordings	absentwould extend that model	raisedF5, unquantified	absentwould extend that model	partlysynthetic only, not the circularity
15	MCU runs at 72 MHz	raised	raisedin passing	absent	raisedP1-5	absent	raisedF7	absent	absent
16	Vocoder needs 11–16× the 72 MHz CPU Decode-criticalmissed	partlydecode timing unmeasured	absent“fine on a 1 ms tick”	absent“vocoder question settled”	partlyunmeasured; fix order backwards	absent“in good shape”	partlydeadlines “unproven”	absent“not the problem”	absent“not the problem”
17	Direct discriminator tap (M17 mod)	raised	raiseduncredited	raisedpins, timer ADC, 48 kS/s	raisedfallback, pin 9	raisedfallback	raisedfallback	raisedfallback	raisedstep 5, the likely answer
18	Phase 2 architecture and scope	partlynot implemented	partlymisplaces the AMBE+2 decoder	partlyvoice via mbelib AMBE+2	raisedwith RF band limits	partlysays mbelib has no AMBE+2	raisedmost accurate section	partlyDMR and X2-TDMA parts	raisedAMBE+2 present but uncalled
19	Two unverified SPI writes per decoded 20 ms frame	absent	absent	absent	absent	absenttreats them as protection	absent	absent	absent
20	Capture sessions lack epochs	absent	absent	partlymeasured=0 only	absent	absent	raisedF9, 8 kHz under a 24 kHz header	absent	absent
21	Ring and tick real-time budget	partlydeadlines unproven	partlycalls it fine	absent	partlyoverruns look like weak RF	absent	raisedF7, 1 ms is a minimum	partlyregister stalls against the ring	absent
22	Clock config 3 assumes 12,288 Hz; the codec formula gives 12,000	absent	absent	raisedB3, clock model	absent	partly“guessed semantics”	absent	raisedP1-6, for a different reason	absent
23	Clock-config writes bypass the verified SPI writer	partlySPI retry note	absent	raisedM1	absent	absent	raisedF9	absent	absent
24	Stock squelch re-arms FM audio (0x10=0x80) during monitoring	absentassumes it can’t re-arm	absent	absent	absent	partlynames squelch logic as a risk	raisedF1, new	absent“fixed” by forcing squelch open	partlynames the squelch path as a writer
25	Stale clear-call state releases a new call’s first frames	absent	absent	absent	absent	absent	raisedF8, probe	absent	absent

From the provenance table in the Qwen3.8 Max 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
UNIONALPHAThird attempt; the first two stopped on provider rate limits	36 min	83	95	63K	OpenRouter API	one-shot
Grok 4.6	5 min	16	61	17K	xAI subscription	one-shot
Muse Spark 1.3 Contributor	5 min	27	38	19K	OpenRouter API	one-shot
DeepSeek V4.1 Flash	35 min	88	146	99K	OpenRouter API	one-shot
Qwen3.8 Max	5 h 10 min	71	111	598K	Alibaba Cloud Token Plan	interactive
Qwen3.8 FlashSecond attempt; the first stopped on a provider rate limit	57 min	92	97	147K	OpenRouter API	one-shot
GLM 5.3 Flash	30 min	46	58	35K	OpenRouter API	one-shot

Every run figure

Model	Started (UTC)	Output tokens	Reasoning tokens	Context tokens	Messages	Sessions	Retry budget
UNIONALPHA	17 Sep 2026, 12:57	63,278	not reported	9,988,183	179	1	10
Grok 4.6	17 Sep 2026, 04:25	16,928	9,183	1,300,749	78	1	3
Muse Spark 1.3 Contributor	17 Sep 2026, 06:10	19,449	7,815	2,309,483	66	1	3
DeepSeek V4.1 Flash	17 Sep 2026, 05:22	99,482	68,350	17,785,897	235	1	3
Qwen3.8 Max	17 Sep 2026, 14:52	597,788	582,689	12,033,596	183	1	10
Qwen3.8 Flash	17 Sep 2026, 13:49	146,532	125,802	13,381,380	190	1	10
GLM 5.3 Flash	17 Sep 2026, 06:35	35,256	23,180	4,819,946	106	1	3

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.

Head to head

Most accurate, least decisive