OKFS: P1.5 sync/output fixes + P1.6 real-speech loop + voxtral augment

Author: nihilistau

memory-okf/LUT.md

@@ -206,3 +206,4 @@ Lookup before you build. `python tools/okf_mem.py lookup --root <root> <kw>`
 | 05e1cca71bf544b5 | agent | KAI-4 P1,ear bake,voice corpus,SAPI render,CPU render,vsub 170,held-out by sentence,data-scale ceiling,safetensors pagefile,cuda contention CPU default,voice_train,voice_bake,G-VOICE-1,hey shannon wake | Full ear-scaling pipeline built+automated CPU-only (2060 stays Gemma's): voice_corpus (templated + wake -> sp_tok_enc -> vsub<=508), voice_render_sapi (6-voice CPU, resumable, 2058 wavs/40s), voice_frames, voice_train (CPU default; NO safetensors mmap - blows pagefile; cuda contends with daemon). Vocab 32->170; unseen-sentence exact recog 6->22%. Held-out-by-sentence is data-bound (KAI-3 pattern). Bake running -> voice_ctc.pt. Remaining: scale corpus + ONNX/POT/GNA re-gate (G-VOICE-1) + re-point ear.py + wake head. | ACTIVE | sum/05e1cca71bf544b5.md |
 | c768dec51af383b7 | agent | KAI-4 P1 live,ear V170 serving,voice_export_ir,ONNX OV IR,ear auto-select,acoustic 0.73,unseen 0.21,data-scale bake,G-VOICE-1 remaining,onnx weights_only | 343-sentence bake done (acoustic 0.73/unseen 0.21). ckpt->ONNX->OV IR path built; ear.py auto-selects voice_ctc.xml; /v1/voice/status now V=170 SERVING. Full chain automated corpus->live inject. Legibility = pure corpus-size bake (KAI-3 pattern). Remaining G-VOICE-1: scale corpus + GNA POT re-gate + wake head. | ACTIVE | sum/c768dec51af383b7.md |
 | 94283c3b803270c7 | agent | G-VOICE-1,ear V217,0.796 legible,batch tokenize newline 107,GPU bake overfit,dropout weight decay,hard-select embed row,full-eval lesson,300-sample phantom,wake head two-stage,POT WinError 6714 WSL,voice live | Corpus 1893 sentences/V=217; batch-tokenize (split on 107); regularized GPU bake 0.21->0.796 held-out-by-sentence (KAI-3 law); ear.hear hard-selects embed rows (soft blend too soft at V217); live V=217 faithful (full-eval 0.796). Wake head 100% recall (two-stage confirm for FAR). POT blocked by WinError 6714 -> WSL like KAI-3. LESSON: eval the FULL held-out set (300-head cluster read 0.46, phantom cuda/cpu bug). | ACTIVE | sum/94283c3b803270c7.md |
+| 792fb441b2cc9f07 | agent | P1.5,P1.6,voice sync 48k 16k,control token strip,graceful framing,real-mic OOD,voice_train capture page,real speech loop,REAL_REPEAT,voxtral augment,iGPU cubecl panic,roleplay feature,max_tokens 256 | Live-play: 'not understanding' = AudioContext 48k vs 16k-trained ear (fixed: console resamples to 16k). + output 96->256, control-token strip, graceful framing (stops confabulation), moderate aug. Real mic is OOD for the SAPI ear -> P1.6 real-speech loop: /voice_train.html capture -> upweighted 8x in the bake + voxtral TTS subset (iGPU panics -> daemon-down 2060 bake). Roleplay banked as harness feature. | ACTIVE | sum/792fb441b2cc9f07.md |

memory-okf/full/792fb441b2cc9f07.md

@@ -0,0 +1,23 @@
+---
+type: memory
+title: ADR-KAI4 P1.5/P1.6: sync fix (48k->16k) + real-speech capture loop + voxtral augment; real mic was OOD for the SAPI ear
+description: Live-play: 'not understanding' = AudioContext 48k vs 16k-trained ear (fixed: console resamples to 16k). + output 96->256, control-token strip, graceful framing (stops confabulation), moderate aug. Real mic is OOD for the SAPI ear -> P1.6 real-speech loop: /voice_train.html capture -> upweighted 8x in the bake + voxtral TTS subset (iGPU panics -> daemon-down 2060 bake). Roleplay banked as harness feature.
+timestamp: 2026-07-10T19:16:45Z
+resource: kairos 45c7431
+sp_status: ACTIVE
+sp_gate: sync+output+strip shipped live; combined real+voxtral+SAPI bake pending operator recordings
+sp_commit: kairos 45c7431
+sp_repro: open /voice_train.html, read sentences; voice_render_voxtral.ps1 (daemon down); voice_frames; voice_train --gpu
+mem_kind: agent
+mem_addr: 792fb441b2cc9f07
+mem_verified: unverified
+mem_lifecycle: active
+tags: [P1.5, P1.6, voice sync 48k 16k, control token strip, graceful framing, real-mic OOD, voice_train capture page, real speech loop, REAL_REPEAT, voxtral augment, iGPU cubecl panic, roleplay feature, max_tokens 256, agent, tier-2]
+mem_tier: full
+---
+
+ADR-KAI4 P1.5/P1.6 (2026-07-11 live-play): first real-mic voice session exposed the real gaps + fixes.
+P1.5 FIXES (shipped kairos 758d229): (1) ★SYNC = the 'getting out of sync / not understanding' = AudioContext runs at 48k even when 16k requested -> audio 3x time-warped -> the 16k-trained ear mis-hears everything. FIX: console resamples the captured utterance to EXACTLY 16k before send (linear resample from ctx.sampleRate). HIGHEST-probability single fix. (2) OUTPUT length: voice max_tokens 96->256 (replies truncated mid-sentence, e.g. Alpha-Centauri scenario cut at 'B'). (3) CONTROL-TOKEN leak on the inject-frame path (<0x0D>/`/[audio]) -> voice service _clean() strips them + a graceful-framing system note ('answer what you can make out, ASK, do NOT invent') stops the confabulation failure mode (the model invented a gaming backstory + a different GPU spec each turn on garbled input). (4) moderate acoustic augmentation in voice_frames (reverb/mic-EQ/noise@20-40dB/gain, --aug_copies; the first attempt at 8dB SNR + heavy reverb UNDERFIT at 0.38 - real quiet-room mic is milder).
+★ROOT UNDERSTANDING GAP: real human mic speech is OUT OF DISTRIBUTION for the SAPI-only-trained ear (0.796 on SAPI, confabulates on real mic = mostly-blank CTC + random tokens). The true fix = REAL + more-natural audio.
+P1.6 REAL-SPEECH LOOP (shipped kairos 45c7431): console /voice_train.html capture page (reads corpus sentences, records real voice, resamples 16k, POSTs /v1/voice/record); gateway endpoints record/corpus/record-status; harness/voice/record.py saves wav+manifest; voice_frames folds real recordings UPWEIGHTED 8x (REAL_REPEAT) into the CTC bake. + voxtral TTS augmentation (operator suggestion - SAPI isn't human enough, and a human can't produce enough samples): voice_render_voxtral.ps1 renders a corpus SUBSET with OUR voxtral (20 natural voices, --euler-steps 3). ★voxtral iGPU 'integrated' path PANICS (cubecl-runtime client.rs:291); auto/discrete needs the 2060 = DAEMON-DOWN bake (RTF~4, subset only). COMBINED NEXT BAKE = operator real samples + voxtral-subset + SAPI + moderate aug -> retrain -> export -> personalized+robust ear.
+ROLEPLAY: banked as a harness feature (emergent strength - the 12B GMs scenarios beautifully unprompted; formalize scene-state persistence like CosySim, don't replace the improvisation).

memory-okf/sum/792fb441b2cc9f07.md

@@ -0,0 +1,24 @@
+---
+type: memory
+title: ADR-KAI4 P1.5/P1.6: sync fix (48k->16k) + real-speech capture loop + voxtral augment; real mic was OOD for the SAPI ear
+description: Live-play: 'not understanding' = AudioContext 48k vs 16k-trained ear (fixed: console resamples to 16k). + output 96->256, control-token strip, graceful framing (stops confabulation), moderate aug. Real mic is OOD for the SAPI ear -> P1.6 real-speech loop: /voice_train.html capture -> upweighted 8x in the bake + voxtral TTS subset (iGPU panics -> daemon-down 2060 bake). Roleplay banked as harness feature.
+timestamp: 2026-07-10T19:16:45Z
+resource: kairos 45c7431
+sp_status: ACTIVE
+sp_gate: sync+output+strip shipped live; combined real+voxtral+SAPI bake pending operator recordings
+sp_commit: kairos 45c7431
+sp_repro: open /voice_train.html, read sentences; voice_render_voxtral.ps1 (daemon down); voice_frames; voice_train --gpu
+mem_kind: agent
+mem_addr: 792fb441b2cc9f07
+mem_verified: unverified
+mem_lifecycle: active
+tags: [P1.5, P1.6, voice sync 48k 16k, control token strip, graceful framing, real-mic OOD, voice_train capture page, real speech loop, REAL_REPEAT, voxtral augment, iGPU cubecl panic, roleplay feature, max_tokens 256, agent, tier-1]
+mem_tier: summary
+mem_full: 792fb441b2cc9f07
+---
+
+# ADR-KAI4 P1.5/P1.6: sync fix (48k->16k) + real-speech capture loop + voxtral augment; real mic was OOD for the SAPI ear
+
+Live-play: 'not understanding' = AudioContext 48k vs 16k-trained ear (fixed: console resamples to 16k). + output 96->256, control-token strip, graceful framing (stops confabulation), moderate aug. Real mic is OOD for the SAPI ear -> P1.6 real-speech loop: /voice_train.html capture -> upweighted 8x in the bake + voxtral TTS subset (iGPU panics -> daemon-down 2060 bake). Roleplay banked as harness feature.
+
+Full context: [full/792fb441b2cc9f07.md](../full/792fb441b2cc9f07.md)