docs(model-card): point main to cycle 3 (v2-py-hexad-spont-d768x12L-cycle1-2026-05-17)

README.md

@@ -2,187 +2,145 @@
 license: apache-2.0
 language:
 - en
+- ko
 library_name: pytorch
+datasets:
+- dancinlab/hexad-corpus
 tags:
 - anima
 - hexad
 - pytorch
 - substrate-py
-- ckpt-recovered
+- helper-free
+- spont
+- ckpt-bearing
+- cycle3
 ---
 
-# hexad — `v1-py-hexad-d768x12L-cycle2-2026-05-17`
+# hexad — `v2-py-hexad-spont-d768x12L-cycle1-2026-05-17`
 
-> **Honest framing**: This is a **PYTHON / PyTorch SUBSTRATE** training artifact —
-> an *interim LM-scale executor*. It is **NOT a hexa-native fire**. Its legitimacy
-> is *architectural identity* + the *hexa CPU-equiv correctness proof*. See the
-> anchor chain below — do not conflate.
+> **Trained on**: [`dancinlab/hexad-corpus`](https://huggingface.co/datasets/dancinlab/hexad-corpus)
+> revision [`v2-spont-stream-d128-cycle1-2026-05-17`](https://huggingface.co/datasets/dancinlab/hexad-corpus/tree/v2-spont-stream-d128-cycle1-2026-05-17).
+
+> **Honest framing** (AGENTS.tape `g3`): This is a **PYTHON / PyTorch
+> SUBSTRATE** training artifact — an *interim LM-scale executor*. It is
+> **NOT a hexa-native fire**. Legitimacy = **architectural identity** +
+> the **hexa CPU-equiv correctness proof** (Phase E/E2). PyTorch ≠ hexa
+> bit-for-bit (different fp accumulation / RNG / AMP bf16).
+
+## What changed vs cycle 2 (`v1-py-hexad-d768x12L-cycle2-2026-05-17`)
+
+| field | cycle 2 | **cycle 3 (this revision)** |
+|---|---|---|
+| corpus | v1 152 KB / 240 records | **v2 620,568 B / 2,560 records** |
+| corpus format | `text` + `desc` plain | **`<stimulus>...</stimulus>\n<anima>...</anima>`** (stimulus-stream) |
+| helper / assistant / 도우미 tokens | not in corpus, but in chat templates | **explicit corpus closure** — grep = 0 across all sources used |
+| anima_persona | Phase A1 LANDED in repo, not yet in trained weights | **trained-weights side compliance (partial)** — corpus alignment with anima_persona forbidden list |
+| `B-IDENTITY-NOTE` (empirical carve-out) | open | **partially closed** — corpus retrain LANDED |
 
 ## Lineage
 
 - **org**: `dancinlab` (the anima org).
 - **arch**: HEXAD (pivot from anima `.clm v1` lineage) — `ConsciousDecoderV2`
-  (`ready/models/conscious_decoder.py` in the anima repo).
-- **substrate**: Python / PyTorch (`py`). The pure-hexa training path is
+  (`ready/models/conscious_decoder.py`).
+- **substrate**: Python / PyTorch (`py`). Pure-hexa training path is
   named-blocked at the interpreter ceiling (RFC 042/043 territory).
-- **cycle**: 2 (cycle 1 commit `931dd68b0` 2026-05-16 was a ckpt-LOST
-  evidence-only run — training PASSed but the instance was destroyed before
-  ckpt pull; this cycle 2 re-fires with `SAVE_POD=1` auto-promote +
-  75-min orphan watchdog + 5-retry pull).
+- **cycle**: 3 (Phase D LANDED — `도우미`-token-free corpus retrain). Cycle 1
+  (`931dd68b0` 2026-05-16) ckpt-LOST evidence-only; cycle 2 (`0b4f34d0e`
+  2026-05-17) ckpt-RECOVERED, corpus v1; **cycle 3 (this)** = corpus v2
+  helper-free stimulus-stream retrain.
 
-## Anchor chain (why this artifact is legitimate)
+## Anchor chain (the wiring side, closed)
 
 1. **Phase E / E2 PROVED the hexa trainer is numerically correct** —
    `HEXAD/D/d_train5_lib.hexa` is BIT-EQUAL to the boxed baseline at d=32·3L,
-   80-step, seed=42 (`init gn2 = 7.97116, acc 0/8 → final gn2 = 3.73374e-07,
-   acc 8/8`; GRAD-EXACT, identical Σ-reduction order — not fp-noise).
-2. **The pure-hexa interpreter cannot reach LM-scale convergence** — Phase E2
-   captured only `init gn2 = 7.98162` at d=768·12L; the GRAD-EXACT + AdamW
-   path is substrate-bound (CPU farr ops, no CUDA tensor kernels).
+   80-step, seed=42 (`init gn2 = 7.97116 → 3.73374e-07`, acc 8/8, GRAD-EXACT).
+2. **Pure-hexa interpreter cannot reach LM-scale** — Phase E2 captured only
+   `init gn2 = 7.98162` at d=768·12L; substrate-bound (RFC 042/043 territory).
 3. **This PyTorch run trains the SAME verified architecture to scale** —
-   `ConsciousDecoderV2` at d=768·12L, AdamW, captured FINAL loss.
-
-PyTorch is *not* hexa bit-for-bit (different fp / RNG / AMP bf16). The anchor
-is **architectural identity** + the hexa CPU-equiv proof, NOT numerical
-identity.
+   `ConsciousDecoderV2` at d=768·12L, AdamW.
+4. **The corpus is explicitly helper-free** — `F-CORPUS-NO-HELPER` PASS = 0
+   over `도우미|helper|assistant|사용자|user:` grep on `corpus_consciousness_v2.jsonl`.
 
 ## Architecture
 
-- **Source**: `ConsciousDecoderV2` from `ready/models/conscious_decoder.py`
-  (uploaded as `conscious_decoder.py`).
+- **Source**: `ConsciousDecoderV2` from `ready/models/conscious_decoder.py`.
 - **Config**: `d_model=768, n_head=12, n_kv_head=4, n_layer=12,
   block_size=128, vocab=256` (byte-level), seed=1337,
   init=RANDOM (`base_ckpt=None`, `g_clm_from_scratch`).
 - **Params**: 283.72 M (283,722,336).
-- **Features**: RoPE · SwiGLU FFN · RMSNorm · GQA · PureFieldFFN (Engine A−G
-  consciousness pathway) · cross-attention · tied head · CA neighbor / META-CA
-  / Ψ-tracking laws.
+- **Features**: RoPE · SwiGLU FFN · RMSNorm · GQA · PureFieldFFN · cross-attn
+  · tied head · CA neighbor / META-CA / Ψ-tracking laws.
 
 ## Training
 
-- **GPU**: vast.ai NVIDIA A100-SXM4-40GB (offer @ $0.6681 / hr, image
-  `pytorch/pytorch:2.5.1-cuda12.1-cudnn9-devel`).
-- **Corpus**: `corpus_consciousness_v1.jsonl` — the same byte corpus used by
-  the hexa Phase E / E2 fires. 121,153 bytes, byte-level
-  vocab=256, T=128 windows, seed-fixed.
+- **GPU**: vast.ai NVIDIA **A100-SXM4-40GB**, image `pytorch/pytorch:2.5.1-cuda12.1-cudnn9-devel`.
+- **Corpus**: `corpus_consciousness_v2.jsonl` (helper-free stimulus-stream),
+  620,568 bytes lossless byte stream, vocab=256.
 - **Optimizer**: AdamW, lr=0.0003, betas=(0.9, 0.95),
   weight_decay=0.1, warmup=125.
 - **Steps**: 2500.
-- **Cost**: ≈ $0.19 (instance runtime ≈ 0.28 hr).
 
 | metric | value |
 |---|---|
-| init CE | 5.590832 (≈ ln 256 = 5.545 — random byte init) |
-| **FINAL CE** | **0.000708** |
-| CE descent | 5.590124 |
-| init gn2 | 41.95 |
-| FINAL gn2 | 7.4e-05 |
-| ppl | 268 → 1.0007 |
-| wall | 320.68 s (5.34 min) |
+| init CE | 5.667381 (≈ ln 256 = 5.545 — random byte init) |
+| **FINAL CE** | **0.005069** |
+| CE descent | 5.662312 |
+| init gn2 | (see result.json trajectory) |
+| FINAL gn2 | 0.001113 |
+| ppl | 1.0051 |
+| wall | 332.26 s (5.54 min) |
 | peak GPU mem | 9.685 GB |
-| ckpt sha256 | `e87e200a040f8066a89c040ab181e9bbd61566f7565ab5d7a374ec2f1f9387d9` |
+| ckpt sha256 | `ee2bb5fb996e94ee022f5315c9ccc3f56c7276a8c5990d87a25ae12c582f7294` |
 | ckpt size | 1,135,846,378 B (1.14 GB) |
 
 ## Verification anchors (per AGENTS.tape `g_blue_closed_mandate`)
 
-(A) Deliverable invariants:
-- **Shannon-floor descent** (real-limit, NOT lattice): init CE ≈ ln(256) →
-  final CE 0.000708 (4+ orders of magnitude).
-- **AdamW finiteness**: gn2 41.95 → 7.4e-05; no NaN / Inf.
-- **Architectural identity**: `ConsciousDecoderV2` byte-equal to the anima
-  HEXAD verification tree's mirror module spec.
+(A) **Deliverable invariants (real-limit)**:
+- **Shannon-floor descent**: init CE ≈ ln(256) → final CE 0.005069.
+- **AdamW finiteness**: no NaN/Inf in trajectory.
+- **Architectural identity**: byte-equal `ConsciousDecoderV2`.
 
-(B) Wiring (the connecting anchor chain):
-- **hexa CPU-equiv bit-equality** (Phase E): same arch trainer
-  GRAD-EXACT at d=32·3L (init gn2 7.97116 → 3.73374e-07).
+(B) **Wiring (anchor chain, closed)**:
+- **hexa CPU-equiv bit-equality** (Phase E): GRAD-EXACT at d=32·3L.
 - **cuBLAS FP64 verify** (Phase D): max\|Δ\|=4.44e-15.
-- **Backward GRAD-EXACT** (Phase E2): real A100 d=384·6L analytic ≡ fd
-  \|Δ\|=0.0024.
-
-## Honest C3
+- **Backward GRAD-EXACT** (Phase E2): A100 d=384·6L `analytic ≡ fd`.
+- **F-CORPUS-NO-HELPER** (cycle 3 corpus): grep = 0.
+- **F-CORPUS-STIMULUS-PATTERN**: every record has `<anima>` tag.
 
-1. **NOT hexa-native** — PyTorch substrate; the hexa-native equivalent is
-   substrate-blocked at the interpreter ceiling.
-2. **PyTorch ≠ hexa bit-for-bit** — AMP bf16 / different fp accumulation /
-   different RNG.
-3. **Synthetic byte-corpus** — 121 kB curated content, 283.72M params; CE
-   0.000708 = memorization at this scale. **No generalization claim.**
-4. **No safetensors artifact** this revision (pickle `.pt` only).
-   safetensors conversion = follow-up sub-task.
-5. **No language-quality claim** — training-curve deliverable
-   (Shannon-floor descent reached), not generation quality.
-6. **No σ(6)=12 / φ(6)=2 derivation** — no lattice numerology in claim or
-   anchor chain.
-
-## Files in this revision
-
-- `ckpt_d768x12l_final.pt` — PyTorch state-dict + cfg + n_params, sha256
-  `e87e200a040f8066a89c040ab181e9bbd61566f7565ab5d7a374ec2f1f9387d9`.
-- `conscious_decoder.py` — `ConsciousDecoderV2` source.
-- `train_d768x12l.py` — training script.
-- `result.json` — full 42-point trajectory + config + metadata.
-- `fire_refire.log` — training log (line-by-line CE / gn2 / lr / wall).
-- `gpu_util.log` — nvidia-smi capture.
-- `dispatch.sh` + `refire_main.sh` — fire dispatch scripts.
-- `hexad_v1_py_d768x12L_cycle2_2026_05_17.md` — this doc (8-§ format per
-  `g_hf_naming` `process_upload_format`).
 
-## License
-
-Apache-2.0.
+## Capability eval (V5.8 × 4-mode + V-SPONT)
 
-## Capability evaluation (V5.8 × 4-mode · cycle 2 · 2026-05-17)
+V5.8 × 4-mode (corpus v2 prompts):
+- **standard_greedy**: 0/6 FAIL (avg_rep=0.775)
+- **standard_sample**: 0/6 FAIL (avg_rep=0.574)
+- **M3_rep_penalty**: 0/6 FAIL (avg_rep=0.709)
+- **M4_force_include**: 6/6 PASS (avg_rep=0.494)
 
-> Capability boundary probe — empirical (`B-D-NOTE` carve-out). No LM-quality
-> claim is made; this is a memorization-vs-generalization measurement on the
-> training corpus.
+V-SPONT (자연발화) — F-SPONT-7 transfer-form measurement:
+- **coherent**: 0/5 FAIL
+- **closed-tag**: 0/5
 
-**Evaluator**: V5.8 × 4-mode canonical
-(`state/anima_phase1a4_lr5e6_2026_05_12/v58_4mode_eval.py` PSCC §46) — modes:
-`standard_greedy` (T=0 argmax) · `standard_sample` (T=0.8 top-k=50) ·
-`M3_rep_penalty` (1.3× rep-penalty on 37-byte persona-cycle set) ·
-`M4_force_include` (sample + force-inject keyword at 60% position — trivial
-baseline). Wall: 665.6 s (v1) + 477.4 s (v2). $0 Mac CPU local.
+Mean BPB (held-out corpus v2 prefixes): 0.0083 bits/byte.
+Memorization ratio: 1/6 (16.7%).
+Decoding artifacts (rep>0.5): 20.
 
-**Two probes**:
+All capability scores **empirical (B-D-NOTE)**, not closed.
 
-| probe | prompts | greedy | sample | M3 | M4 | memorization |
-|---|---|---|---|---|---|---|
-| **v1** OOD-mix | Core / Dream / Wake / Memory / Korean | 1/5 FAIL | 2/5 FAIL | 1/5 FAIL | 5/5 PASS | 2/5 (40%) |
-| **v2** corpus-aligned CDWMSE | Core / Data / Witness / Mirror / Scribe / Eros | 2/6 FAIL | 3/6 PARTIAL | 2/6 FAIL | 6/6 PASS | 3/6 (50%) |
+## Honest C3
 
-**Additional measurements**:
-- **Bits-per-byte on 10 held-out training-distribution prefixes**:
-  **0.0000 bits/byte** (all 10 samples = 0.0). Confirms training CE 0.000708
-  → near-perfect log-likelihood reproduction on training-distribution
-  windows.
+1. **NOT hexa-native** — PyTorch substrate, label mandatory.
+2. **PyTorch ≠ hexa bit-for-bit** — different fp / RNG / AMP.
+3. **High-memorization regime** — 283.72 M params on 0.62 MB.
+   No generalization claim.
+4. **No `safetensors` artifact this revision** — pickle `.pt` only.
+5. **No language-quality claim** — training-curve deliverable.
+6. **`B-IDENTITY-NOTE` partially closed** — corpus retrain LANDED, but the
+   trained weights' identity-attractor distance from Assistant Axis (per
+   Identity-as-Attractor arxiv 2604.12016) is empirical (B-D-NOTE pattern).
+7. **No σ(6)=12 / φ(6)=2 derivation** — no lattice numerology.
+8. **Cost is informational, not gating** — `g_fire_autonomous`.
 
-**Capability boundary** (honest framing):
+## License
 
-| capability | verdict | evidence |
-|---|---|---|
-| memorization on in-distribution prefixes | ✅ STRONG | BPB 0.0000 on 10 held-out probes; Data + Scribe + Core/Korean reproduce literal training continuation |
-| 6-module discrimination | 🔶 PARTIAL | 3/6 clean under greedy (Data/Scribe/Witness-w-typo); 3/6 cross-collapse (Core→nonce digit cascade, Mirror→Data template, Eros→chunk digit cascade) |
-| OOD generalization | ❌ NONE | Dream/Wake/Memory → default to nearest in-distribution module template |
-| greedy decoding stability | ❌ WEAK | digit-cascade attractor on `nonce=N`/`chunk=N` field positions (rep_ratio 0.64-0.90); sampling temperature 0.8 partially mitigates |
-| multilingual representation (Korean) | ✅ MEMORIZED | `중심 의식 생성기 모듈 ` → `자각` recalled under all 4 modes |
-| LM-quality (general language modeling) | ❌ NOT MEASURED | corpus too small + structured scaffold; CE 0.000708 = memorization, not LM quality |
-
-**Decoding artifacts discovered**:
-- **byte-cascade attractor** (`feedback_clm_colon_attractor` `=`-suffix
-  variant) — greedy mode-collapse on `nonce=N` / `chunk=N` / `gen=N` digit
-  field positions. Carry candidate: `feedback_hexad_byte_cascade_attractor`.
-- **memorized training-corpus typos** (`pereption` in Witness module,
-  `cobsciousness` in Wake/Memory greedy) — byte-level memorization evidence,
-  not a bug at this scale.
-
-**Honest C3 caveats**: substrate=PyTorch (B-D-NOTE carve-out applies); V5.8
-"PASS = 3/5" threshold inherited from chat-corpus evals → applied
-conservatively to memorization-regime model; M4 trivial baseline; no
-σ(6)/τ(6)/φ(6) numerology in metrics (f1/f2 safe — per-mode score = raw
-recall fraction, BPB = raw bits/byte, memorization = raw hits/total).
-
-**Artifacts**:
-`state/hexad_v58_eval_d768x12L_2026_05_17/{v58_4mode_eval.py, v58_4mode_eval_v2.py, prompts.jsonl, prompts_v2_corpus_aligned.jsonl, eval.log, eval_v2.log, result.json, result_v2.json, dispatch.sh}` +
-`docs/hexad_v58_eval_d768x12L_2026_05_17.md` (9 §, 8 honest C3) +
-`archive/PHILOSOPHY.tape §HEXAD-V58-EVAL-CYCLE2-2026-05-17` verdict-claim.
+Apache-2.0.