THOR RECORD / J-012 / 02 SEPTEMBER 2026
The voice path now runs end to end. The acceptance gate still does not pass.
J-011 ended with separate realtime components and an integration plan. Five days later, the latest physically exercised Thor path reaches from the XVF3800 microphone through turn formation, transcript admission, Core-owned conversation state, Qwen3.8, Scylla speech, and one physical playback owner.
In the latest eight-response run, median detected speech-end to first generated PCM was 1.3705 seconds. That is a material improvement from the roughly 2.2–2.6-second fresh response range at the 31 August CP01 freeze, but it is not a reliability result and it is not the same as an independently measured acoustic hearing timestamp.
Latest physical checkpoint
What the 1.37-second result actually measures.
These statistics come from the physical CP09-I run at commit 64df168f4a8711a07def4f598e91f30f35f9e46b. Different metrics have different sample sets and should not be added as if they were one deterministic trace.
- Speech end → first PCM · P50
- 1.3705 sEight physical responses
- Speech end → first PCM · P95
- 1.7971 sFails the ≤1.3 s gate
- Qwen TTFT · P50
- 211.8 msLocal Qwen3.8 / SGLang
- Later-turn TTFT · P50
- 212.5 msNo old cache cliff in this run
- Hard commit · P50
- 499.3 msEleven turn decisions
- First token → speech unit · P50
- 399.8 msLocal safe-phrase framing
- 01End the turnHard commit P50 · 499.3 ms
- 02Start cognitionQwen TTFT P50 · 211.8 ms
- 03Form safe speechPost-token P50 · 399.8 ms
- 04Synthesize PCMScylla examples · about 0.28–0.38 s
Short, naturally complete replies reached about 1.02–1.19 seconds. Normal and longer replies commonly remained around 1.5–1.8 seconds because useful phrase formation and speech onset—not Qwen's first token alone—now dominate the variable part of the path.
29 August → 2 September
Five days of controlled integration, failure, and simplification.
The result did not come from one model swap. Each checkpoint isolated a different failure layer and retained the useful evidence even when the candidate was rejected.
Persistent recognition, a duplex latency harness, turn/playback reliability repairs, and startup-health checks turned the modular plan into a measurable physical path.
ConversationLedger took canonical ownership of spoken history, generation leases, WAIT, CONTINUE, STOP, stale output, and generated-but-unheard exclusion. Spain-Spanish controls worked in the final run, with one retained Espera → Espero miss.
Decode graphs lifted normal Qwen output to about 13.3 tok/s. Natural chunk probes rejected arbitrary timed cuts, and direct measurement exposed severe GPU contention between continuing Qwen decode and Magpie/NanoCodec continuation synthesis.
Smart Turn overlapped Cohere Transcribe without surrendering Core turn authority: 15/15 speculative requests were reused with zero wait at commit. The same 15-turn run exposed later-session Qwen TTFT growth.
Front-sliding model history destroyed SGLang prefix-cache reuse. Core-side open-session memory was separated from a bounded append-only Qwen working-set epoch with explicit prewarm.
The cache-stable architecture entered the live head. Completeness, startup, bilingual-quality and staged-response experiments exposed new problems; CP09-I removed the model-facing speech protocol and returned to one natural Qwen stream.
Latest physically tested path
One governed response, one natural generation, one playback owner.
The latest checkpoint is no longer the Nemotron/DualTurn/UltraVAD/Magpie component plan described in J-011. The tested path changed as measurement eliminated weak candidates.
- 01Physical audio boundaryXVF3800 hardware AEC + Creative Pebble V3Thor fixture
LEFT carries near-end/barge evidence; RIGHT carries semantic speech. The standalone Thor AEC fixture remains the accepted hardware basis.
- 02Turn formationNear-end acoustic evidence + Smart TurnPhysical path
Deterministic local timing retains floor authority. Smart Turn advises whether the utterance appears complete; it does not commit a user turn by itself.
- 03TranscriptionSpeculative Cohere Transcribe overlapExperimental
Recognition starts before final commit and is reused only after bounded admission. Correction, 5 September: Cohere Transcribe runs locally on Thor through a resident service; the earlier wording incorrectly described this ASR path as nonlocal.
- 04Conversation authorityEvopien Core + ConversationLedgerCore-owned
Core owns user admission, spoken-only history, WAIT/CONTINUE/STOP, cancellation, checkpoints, audibility, and generation leases.
- 05Session contextDerived open-session memory + bounded Qwen epochCache-stable
Canonical session evidence remains outside the model cache. Qwen receives a bounded append-only working set with prewarmed rollover instead of front-sliding history.
- 06Local cognitionQwen3.8-27B R8.1 normal no-thinking decodeThor exercised
The retained SGLang control keeps fast first-token behavior. Alternative runtimes did not displace it for first speech.
- 07Speech onsetLocal 4–6-word safe-prefix framingOne stream
A naturally complete short answer may release immediately; otherwise the speech layer—not Qwen—waits for a small syntactically safe prefix while the same generation continues.
- 08Speech + playbackScylla INT8 CPU + one ordered ownerPhysical path
CPU speech avoids the measured Qwen/GPU-codec contention mechanism. Unheard text and stale PCM must lose authority on cancellation.
The long-session failure
A bigger chat window only moved the cliff.
CP08 began near 228.6 ms Qwen TTFT, then reached 681.0 ms after turn five. CP09-A tied the jump to the first front eviction from model-visible history: cached-prefix reuse collapsed and the whole prompt had to be prefetched again.
Raising the window from four to 32 exchanges did not solve it. It postponed the same failure to turn 34, where TTFT rose to 1.849 seconds and later turns held around 1.685 seconds P50.
Core lifecycle freeze
Only what was actually heard becomes conversation truth.
Latency work was not allowed to bypass the conversation ledger. A faster answer is not acceptable if cancelled or unheard text contaminates later context.
WAIT creates a real resumable checkpoint. CONTINUE receives a fresh governed lease. STOP clears without filler. Assistant units become canonical only after audibility, and stale provider output is rejected after ownership changes.
On a non-WAIT interruption, canonical memory keeps the exact spoken prefix. Only the disposable model working set may annotate that prefix as interrupted, preventing Qwen from mistaking a cut-off answer for a deliberately complete one.What the experiments eliminated
The active path is narrower because alternatives were measured.
Rejected here means rejected for the current first-speech role under the tested Thor constraints. It does not mean the work has no future research value.
Decode graphs improved token time by roughly 35–38%, reaching about 13.3 tok/s with hot TTFT in the 233–240 ms P95 class.
NanoCodec RTF P95 degraded from 0.188 idle to 1.552 during Qwen decode—an 8.236× collapse in the tested contention case.
Moving NanoCodec to eight CPU cores avoided the measured GPU contention and made it the best Magpie continuation candidate, but first-response timing remained worse and the practical Qwen + Scylla path displaced it.
Avoids the specific GPU-codec contention path and produced ordinary short-chunk first PCM in roughly the 0.3–0.4-second class.
English and Spanish synthesis worked, but warmed usable time-to-first-audio remained multi-second and some streams emitted only fragments.
Built and ran on Thor with strong prefill, but about 12.3 tok/s decode did not improve the active conversation path.
Reached 24.62 tok/s on a long Spanish generation, but conversational TTFT and first-sentence latency were worse than the retained first-speech runtime.
Model-owned markers, abort fences, rescue prompts, and second requests added complexity and still produced compliance, factual, and proportionality failures.
Why the branch is not green
The median improved. The conversation still breaks in important ways.
In the latest physical run, a false acoustic barge cancelled a valid assistant answer. Cohere then classified the candidate interruption as single-word noise, so the user got neither the original answer nor a valid replacement turn.
The session also ended with three temporary generated units still present. Only eight Qwen/context samples existed, below the ten-sample sustained-history gate, and no clean admitted interruption physically proved the new disposable-history annotation.
Rollback and development state
The measured anchor is preserved; newer repository work is not requalified.
The physical run anchor is deliberately separate from newer repository commits. Newer code and experiment scaffolding are not stronger evidence until the full physical path runs again.
The eleven post-anchor CP09-I commits implement a provider-EOS guard for short onset, checkpoint retention until interruption, cancellation cleanup with resume metadata, bounded fail-soft handling for long Cohere and PCM behavior, and sentence-boundary and continuation framing with regression tests. Those are implemented changes, not a newer physical qualification.
- Physical run anchor
64df168f4a8711a07def4f598e91f30f35f9e46b
Source of the measured CP09-I latency and failure record on this page.- CP09-I branch head
b4d7212f827006578c628c380f10435757777f31
Numeric-freeze documentation at the reviewed remote branch pointer; it does not replace the physical evidence anchor.- Latest reviewed repository head
ms08/cp09j-qwen-decode-frontier478699e8b9c2c9ecdbb45d8d9d8b15ce2957b812defines the next decode-frontier experiment. No CP09-J result is claimed.- Evidence ceiling
- CP09-I is Thor-exercised with known failures. Later repairs and CP09-J are implemented or documented repository work; no newer physical qualification exists for them.
Current evidence boundary
Where Evopien stands on 2 September.
Microphone-to-physical-playback operation exists at the exact frozen checkpoint and stack described above.
Exact median was 1.3705 s speech-end to first PCM; P95 was 1.7971 s. This is not a guaranteed or acoustic-audibility figure.
The 100,000-turn Core-session/cache architecture passed its declared text-only Thor corpus; physical sustained qualification remains short of sample count.
English and Spain Spanish have bounded subpath evidence, but the paired physical exit corpus and lifecycle gates have not passed.
Later CP09-I repairs and the CP09-J decode-frontier scaffold are present in the reviewed repository history; no newer physical qualification is claimed.
Standalone C920 capture works. Current Qwen3.8 video, bounded scene state, coexistence, and freshness handling remain MS-09 work.
Foundation v1.2 remains controlling; privacy, real-user, recovery, safety, independent-isolation, and deployment gates remain open.