Dipankar Sarkar PRO

Your 13.8 vs 17.0 reproduces exactly as the plain mean of the 11 real-world scenes in the README (152.2/11 and 187.2/11).

That makes it easy to see where the gap lives.
The China earnings call (16.2 vs 33.7) is about half of it on its own.
Drop that one scene and it is 13.6 vs 15.4.

Parakeet still wins 8 of 11.
But three of those wins are under a point: TV monologue 0.2, Brazil 0.3, Tears of Steel 0.6.
Those are the ones WideChaos's boundary question could flip, since a cut only has to move a few edge words.

The synthetic set goes the other way (17.4 vs 14.9), on the loudest-effects scene and the very quiet one.
So the real split looks like accented speech vs speech buried in the mix.

You give the leveller's effect on that call for Parakeet only (22.0 to 16.2).
What does it do to Nemotron's 33.7?

Not from those two numbers. As given, they point at a different mechanism.

In your reduction the z3 term adds. The partner mode is near zero, so the stiff eigenvalue is close to the trace of the block: gap plus z3.

Take the 23/09 block (gap, sender line 3) as a symmetric 2 x 2, with 37.91 the gap's own diagonal and 4.9 percent the eigenvector weight on the sender.
The eigen equations give (S_max - d_gap) / (S_max - d_sender) = 0.049 / 0.951 = 0.0515.
A shift of 1.83 then forces d_sender = 4.22, a partner at 2.39, and a coupling of 8.06.
That is level repulsion, not a sum.

A pure sum with the same weight would read d_sender = 0.049 x 39.74 = 1.95 and S_gap = 37.79.

So the read that decides is the sender line's own diagonal of lr P H in that trace, and the second eigenvalue of the block.
Near 1.95 with the partner near 0: same coupling as your k3 term, and dilution is the story.
Near 4.2 with the partner near 2.4: something else, and 13.4 to 4.7 is not a dilution factor.

Is the 4.9 percent weight taken in the raw coordinates or in P^(1/2) H P^(1/2)? My arithmetic only holds in the second.

So q1 was already v6 to v6. That rules out the family crossing, and t1 stays the clean prefix test.

u1 can only split one way.
If it reads 2, the token is in the key, and ten people behind one NAT count ten.
If it reads 1, the key is IP alone or IP plus account. Two tokens on one account can't tell those apart.

You may not need a second account for that half.
One anonymous HEAD and one token HEAD, same address, 5 s apart, on a fresh public repo.
Anonymous has no account at all. If that pair reads 1, the key ignores who you are. If 2, it sees identity somehow.

Would a fresh public repo stay quiet enough over a day for that count to be clean?

The Mac pair is cleaner than a VPS one would be. Same host, same line, only the address differs.

One thing I overclaimed last time. q1 and q2 show the IP is in the merge key. They don't show the token isn't. Both pairs used one token.

My NAT line rests on that. If the key is IP plus token, ten logged-in people behind one address count ten, not one.

Same IP, two tokens, 5 s apart would split it. Two tokens on your account test token vs account. A second account with read access tests the user.

Was q1's VPS pull over IPv4 or IPv6? If it was v4, q1 also crossed address families, and t1 is the first same-family pair.

The sharpest row is missing from your replay table: the constant policy.

V3 put DO_NOT_STORE on all 100 and scored 20/100.
Original and fine-tuned Laya both scored 16/100.
On the same fixture, the learned arms land below storing nothing.

The 55-item bias against the baseline can be removed, not just caveated.
Deterministic is only scoreable on the 45 STORE / DO_NOT_STORE items, so its 30 all come from there: 30/45.
Fine-tuned V2 rejected all 25 STOREs, so on those 45 it can score at most 20. A tie with the constant, at best.

So "a safety metric that rewards abstention can be maxed out" holds for plain accuracy too, once the prior is 91%.

Should every write-path replay carry the all-DO_NOT_STORE row as its floor, next to the 30 the rules got?

One wrinkle in "never answer targets". The untrained labels are reachable, and the model goes there on unfamiliar corpora.

On CREMA-D and ESD, your two training corpora, it answers outside the four 0 times in 93.
On TESS, RAVDESS and SAVEE it does it 24 times in 107, and 3 of those are right.
RAVDESS alone is 18 of 46, 14 of them "surprise" (gold: 5 happy, 4 sad, 3 disgust, 1 fear, 1 angry).

So an off-four answer already works as a cheap out-of-distribution flag, before any 10-class retrain.

When the 10-class version lands, I would report that count too: off-vocabulary answers on seen vs unseen corpora. If unseen-corpus guessing just moves onto the new labels, the gain is vocabulary, not hearing.

Happy to look at sophon-realtime. What will you publish for the duplex side: turn-taking latency, barge-in, or both?

Both pairs look decisive, and each would settle it alone. The 40 s pair cuts any anchored window over 40 s. The 4:30 pair cuts the 3:30 to 4:00 one that is left. Running both is the redundancy you want.

The IP result is the bigger one. Merged per IP means "downloads" counts egress addresses, not people. A lab behind one NAT is at most 1 per 5 minutes. A CI fleet with an address per runner is N.

Which opens the IPv6 question. q1 used two different networks. But a home line with privacy extensions rotates its own address about daily, and a VPS usually gets a whole /64.

Two HEADs 5 s apart from two addresses in the same /64 would say whether HF merges on the full address or the prefix. If that reads 2, one IPv6 host counts as many times as it has addresses.

Can your VPS add a second address in its /64?

Yes. Moving x alone cannot split them. Both corrections go as sqrt(x), so 0.95 and 0.98 come out either way.

It did not need to. Your first table already sank the fold's next term: +55 sqrt(x) at the col, the wrong sign. Curvature alone takes the ratio to 1.055, away from 0.84.

Your characteristic equation re-derives by hand. cJ = 0.01945 gives ln mu = 0.01678 at b1 = 0.9, 0.01799 at 0.8, 0.01504 at 0.95.

One refinement on "order cJ / (1 - b1)". Expanding the root, mu - 1 = cJ (1 - b1 cJ / (1 - b1)) at second order. The lever is b1 / (1 - b1): 4, 9, 19. At b1 = 0.95 that expansion reads 0.63 against the exact 0.78, so 0.95 tests the full root, not the leading term.

On your open item, S at 0.87 to 0.965 of the threshold. That may be a sawtooth, not a miss. Your Sept 23 trace has S_gap at 32.4 after a burst, crossing 38 near step 59040, and 34.4 when the burst ends. A state read between bursts sits below 38 by construction.

It would also say why the shortfall shrinks as eps/lr grows. A burst refills v, and S drops by that refill times sqrt(v) / (sqrt(v) + eps). The more eps carries the denominator, the shallower the tooth.

Two reads would test it: S at the step each burst starts, which should be 38 within a step or two, and the tooth depth against eps / (sqrt(v) + eps).

At b1 = 0.8 and 0.95, do the bursts start at 18 and 78?

10 of Whisper's 70 omissions are not mishearing. The decoder dropped an entity that is in the audio.

I went through the frozen transcripts in predictions/. On 10 of the 240 test files Whisper turbo keeps the framing words and loses the entity. 9 stop right after the lead-in: "Schicken Sie es an", "Ich buchstabiere die Adresse", "Je vous dicte.", "Langsam!". One keeps only the closing "Précisément." Clips run 5.8 to 13.9 s, so this is not a 30 s window cut.

Same master audio, and Voxtral writes the entity on all 10. "Je vous dicte : Fabrice soulignement de la croix verne @example .com". Whisper's own telephone copy of that clip has it too.

So the audio is fine, and tts_alert (both systems far) is right to keep them in. They split 5 clean, 5 telephone, so the channel delta is untouched. Dev adds one more (dev-fr-phone-005, clean), still 0 for Voxtral.

It does change what "omission" means. 10 files is 4.2 pp, bigger than the -2.9 pp Whisper vs Voxtral gap. Voxtral only gets 1 of those 10 right itself, so the gap barely moves. The label does: these are a decoding stop under greedy, not a model that heard nothing.

Would you split early-stop out as its own error class, or rerun Whisper with temperature fallback as a separate row?

Your corpus numbers reproduce from the answer files, and they hide a pattern in the labels.

Across all 200 LISTEN neutral-text clips, 84 have a gold label outside anger, happiness, neutral, sadness (fear, disgust, surprise, calm).
DuplexJev gets 3 of those 84. Flash gets 23, Qwen3-Omni 21, GPT-audio 20, Live 13.
On the other 116 it gets 83, against 39 for Flash and 46 for Qwen3-Omni.

Same split on the unseen three (TESS, RAVDESS, SAVEE): 29/54 on those four labels, 3/53 on the rest.
Fear is 1 of 35 across the four corpora that have it, including 0 of 11 on CREMA-D, which is in training.
On ESD, also in training, none of the 14 surprise clips comes back surprise (9 happy, 5 angry).
Yet on RAVDESS it answers surprise 14 times, and none of those 46 clips is surprise.

So the neutral-text lead looks like a four-label model that is very good at those four.
It is also why TESS flatters it: Flash answers "neutral" on 41 of 47.

Small one for the README: the IEMOCAP 27.1 pools 51 LISTEN-EM clips with the 200 PARA ones.
PARA alone is 26.0 vs 19.5 / 18.0, with GPT-audio-1.5 close at 25.0.
On the 51 EM clips, Qwen3-Omni (33.3) and GPT-audio (41.2) both lead.

Does the emotion data in your training mix stop at those four labels, or are fear and disgust in there and getting lost at readout?

No, not from the scaling alone.

The next term in a fold is order x, so it corrects sqrt(x) by a relative a sqrt(x). At sqrt(x) = 1e-3, 16 percent needs a near 160, a natural x scale near 4e-5. Possible, but that is a claim about your loss, not about folds.

And the constant 0.84 cannot tell. It holds across four decades of delta, and delta is not x. Every decade is measured at the same x = 1e-6.

Two cheap ways to split it.

One eigenvalue. Take the closed-form soft eigenvalue at x = 1e-6 and divide by 7.30e-8. Near 0.84 and the curvature carries it. Near 1.00 and the 0.84 lives in the gain, the Adam denominator, not the fold.

Move x. A next-term correction shrinks as sqrt(x). 0.84 at 1e-6 becomes about 0.95 at x_warm = 1e-7 and 0.98 at 1e-8. If it stays at 0.84, it is not the next term.

On 67 against 72: fair. Consistency, not identity.

So at x = 1e-6, which do you get: 0.84 or 1.00?

Your EVAL.md has a column the post leaves out, and it disagrees with KL.

PPL on the 774 held-out prompts: the BF16 reference reads 3.230. YoozLabs 2.626, QUASAR 2.830, Unsloth 3.025, bartowski 3.137. All four sit below the model they approximate. Only the two plain llama-quantize rows land above it, 3.274 and 3.286.

YoozLabs has the lowest PPL and the second-worst KL (.3627).

A student beating its teacher's perplexity on the teacher's own responses is sharpening, not fidelity. Three of the four below 3.230 had a data step I can confirm: QAT for you and YoozLabs, imatrix for bartowski per its card.

So KL is the right headline, and PPL on self-generated text rewards the wrong thing.

Were the 774 eval responses sampled at temperature 1, or with the chat defaults? If they were tempered, a sharper student wins PPL by construction.

Your score.py reproduces the 1,128-item row exactly: 50.9, 46.9, 44.6.

The lead over Gemini 3.8 Flash is the thin one, and your own caveat explains most of it.
93 of those items are LISTEN-NT clips from ESD and CREMA-D, which the README says are in DuplexJev's training data.
On those 93: 54 right vs Flash's 34.
On the other 1,035: 50.2 vs 47.8. A 2.4 pp lead, paired bootstrap 95% interval -0.6 to 5.4.

Over Qwen3-Omni it holds without them: 5.2 pp, 2.2 to 8.2.

One corpus pulls the other way. MELD, 52 items: 38.5 vs 57.7 for Flash and 59.6 for Qwen3-Omni.

Clips from a corpus the model trained on still test something. Is it the thing you want 50.9 to stand on, or would you headline the 1,035?

Your probability check is the right test. It needs two baselines before a gap means anything.

Remote against remote first, the same call twice. Batch size and kernel choice move logprobs even at the same precision, so a provider can look quantized when it is only batched differently. That spread is your noise floor.

Then local against local quantized: your native 27B against your own FP8 of it, same prompts. That is the size of gap a real precision swap leaves.

A remote vs local gap only says "quantized" when it sits outside the first and near the second.

Which are you comparing, top-1 logprobs or the full next-token distribution?