OKFS: G-VOICE-1 conversational ear legible+live

Author: nihilistau

memory-okf/LUT.md

@@ -205,3 +205,4 @@ Lookup before you build. `python tools/okf_mem.py lookup --root <root> <kw>`
 | fdebbd495a530e25 | agent | KAI-4,latent voice,GNA ear,inject_frames,audio placeholder 258881,Audex teardown,discrete token reject,voice head,GAN vocoder,V_sub vocab,G-VOICE-0,voice channel GUI,voxtral reuse,E=3840,WSL no GNA MMIO | Sovereign voice: KAI-3 GNA ear (silicon-gated) wired to mic + /v1/voice + console -> inject_frames -> 12B (G-VOICE-0 parity 8/8, live e2e proven). Audex: steal unified-input + 2-stage output split, reject discrete-token output. P0=plumbing GREEN; P1 vocab scale-up makes free speech legible; P3 voice head = text bypass. E=3840 not 1024; GNA is frozen INT16 conv-only leverage. | GREEN | sum/fdebbd495a530e25.md |
 | 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 |

memory-okf/full/94283c3b803270c7.md

@@ -0,0 +1,19 @@
+---
+type: memory
+title: ADR-KAI4 G-VOICE-1: conversational ear legible+live (0.796, V=217); hard-select fix; wake head; POT deferred to WSL
+description: 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).
+timestamp: 2026-07-10T18:13:11Z
+resource: kairos c1527b4
+sp_status: ACTIVE
+sp_gate: ear held-out 0.796 tok/0.754 exact live; >=0.85 + GNA POT pending
+sp_commit: kairos c1527b4
+sp_repro: voice_bake style: corpus/render/frames + voice_train --gpu (daemon down) + voice_export_ir; ear.hear full-eval
+mem_kind: agent
+mem_addr: 94283c3b803270c7
+mem_verified: unverified
+mem_lifecycle: active
+tags: [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, agent, tier-2]
+mem_tier: full
+---
+
+ADR-KAI4 P1 / G-VOICE-1 (2026-07-11): conversational ear SCALED + LEGIBLE + LIVE. Corpus 83->343->1893 sentences (rich templates, bounded word bank -> vsub 32->217, under the 508 GNA cap). ★BATCH TOKENIZE: one sp_tok_enc call over newline-joined sentences, split output on the newline token (107) - was 1 subprocess/sentence (minutes -> instant). SAPI render 6-voice x 3-rate = 11358 wavs (~40s, CPU, 2060 free). GPU BAKE (stopped the daemon to free the 2060; cuda-training contends): unregularized run OVERFIT (peak 0.749 ep20 -> decline); regularized (dropout 0.15 + weight_decay 1e-4) HELD 0.796 tok / 0.754 exact held-out-BY-SENTENCE (unseen compositions), up from 0.21 - the KAI-3 data-scale law confirmed (61->924 samples = 0.44->0.87). ★ear.hear now HARD-SELECTS the argmax embed row per kept frame (the softmax(tau=0.2) blend is too soft at V=217 -> blurry injected vectors; SP_VOICE_SOFT=1 restores KAI-3 V=32 soft behaviour). LIVE: /v1/voice/status V=217; ear.hear on the FULL 1134-eval reproduces training EXACTLY (0.796) - faithful CPU/OV deployment. voice_export_ir (ckpt->ONNX->OV IR; MUST mirror Dropout in the export arch for state_dict indices; weights_only=False; onnx pkg). ★★DEBUG LESSON: a 300-sample eval HEAD cluster read 0.46 and nearly sent me chasing a phantom CUDA-vs-CPU bug (logits actually identical 2.3e-5) - ALWAYS eval the FULL held-out set. WAKE HEAD (voice_wake.py): tiny GNA-legal binary 'shannon' classifier, 100% recall but 10% FAR at raw argmax -> use TWO-STAGE (cheap wake proposes, ear-decode-contains-shannon confirms) to kill false trips. POT GNA i16 (silicon leg) DEFERRED: openvino-dev pot import throws WinError 6714 (Windows filesystem-transaction minifilter on the import dir scan) - this is exactly why the KAI-3 POT ran in WSL; the FP32 IR runs the ear on CPU in ms so the interactive channel is unblocked; GNA i16 matters for P4 always-on wake. REMAINING G-VOICE-1: scale corpus more for >=0.85; POT via WSL; two-stage wake wiring. Then P2 vocoder, P3 voice head.

memory-okf/sum/94283c3b803270c7.md

@@ -0,0 +1,24 @@
+---
+type: memory
+title: ADR-KAI4 G-VOICE-1: conversational ear legible+live (0.796, V=217); hard-select fix; wake head; POT deferred to WSL
+description: 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).
+timestamp: 2026-07-10T18:13:11Z
+resource: kairos c1527b4
+sp_status: ACTIVE
+sp_gate: ear held-out 0.796 tok/0.754 exact live; >=0.85 + GNA POT pending
+sp_commit: kairos c1527b4
+sp_repro: voice_bake style: corpus/render/frames + voice_train --gpu (daemon down) + voice_export_ir; ear.hear full-eval
+mem_kind: agent
+mem_addr: 94283c3b803270c7
+mem_verified: unverified
+mem_lifecycle: active
+tags: [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, agent, tier-1]
+mem_tier: summary
+mem_full: 94283c3b803270c7
+---
+
+# ADR-KAI4 G-VOICE-1: conversational ear legible+live (0.796, V=217); hard-select fix; wake head; POT deferred to WSL
+
+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).
+
+Full context: [full/94283c3b803270c7.md](../full/94283c3b803270c7.md)