spb/wpai0002_uqo

1# STATUS — WP-AI-0002 (weights published on Hugging Face, D23; updated 2026-11-11, 03:30 UTC)23## Gates4| Gate | State | Evidence |5|---|---|---|6| A — repository, governance, 100-paper smoke, quality audit | **GO** | `docs/gates/GATE_A.md` |7| B — 500–1 000 validated Claude pairs, spend, human review | **GO** (974 passages, 5 383 validated pairs, 29.68 USD total spend) | `docs/gates/GATE_B.md` |8| C — leakage-safe splits, reproducible baselines, benchmark + statistical plan | **GO** (splits frozen; B0/B1 run on val; pre-registration v1.0 frozen at `e74b7126` + A1) | `docs/gates/GATE_C.md` |9| D — M1/M2/M3 training, one-batch overfit, pilot results | in progress — exploratory val (502 pairs): B0 chrF 71.2 · B1 68.4 · **B2 T5-small 68.8** (W1 to originals 0.112 < 0.170, i.e. closer style distribution but lower lexical fidelity) · **M1-tiny from scratch 32.9** (degenerate). Best val CE over seeds 1–3: M1 3.69/4.09/4.31 · M2 3.72/5.03/3.71 · M3 token-only 3.80/3.74/3.58 · M3 with auxiliary losses ≈ 5.4 → diagnosed (D18): the L_fact normalisation, not the architecture; fixed and re-running. **Denoising pretraining (D17)** works: pilot denoiser (12 000 steps on 17 k train passages) val CE 1.58; fine-tuned on the pilot pairs → **val chrF 63.6** [62.2, 64.9] (from scratch 32.9), number recall 0.93, placeholders 0.88, W1 0.150 (< identity 0.170); still −7.6 chrF below identity on this small pair set (B2 T5-small −2.4). Denoiser alone (no pair fine-tuning) chrF 58.1. Medium denoisers tiny (M3U96b) and small 130 M (M4M64a) running, fine-tuning queued after them. | `artifacts/metrics/pilot_val_exploratory/`, `artifacts/runs/`, `docs/DECISION_LOG.md` D17–D18 |10| E — larger training | **GO** (Medium pairs, confirmatory pass run once; H1/H2 vs B0 detectable but below SESOI, H1/H2 vs B1 and H4 supported, H3 not established, H5 not supported) | `docs/gates/GATE_E.md` |11| F — release + paper | **GO** (2026-11-11): paper **v1.0** compiled (79 pages, all `\\pending{}` markers resolved, new §10 "Beyond the vanilla Transformer" with the HTR 2.0 tracks, figures/tables generated from artefacts by `scripts/build_figures.py --tag HtrTwoTest` and `scripts/build_htr2_figures.py`, val/domain/period deltas by `scripts/htr2_split_deltas.py`); release candidates `artifacts/releases/HTR-Tiny-0.3.0`, `HTR-Base-0.1.0`, `flan-t5-xl-htr` load independently; six models served on htr.spboucher.ai with model cards. **Author to do:** confirm SESOIs; decide weights release vs Anthropic terms; decide the platform's default model; independent rater study = future work. | `docs/gates/GATE_F.md`, `paper/main.pdf` |1213## HTR 2.0 — exploratory tracks after Gate E (D22, 2026-11-10; all numbers are direct paired chrF++ differences vs B0 identity, ± 1.96 SE, no bootstrap yet; the Gate E confirmatory numbers are unchanged)1415| System (track) | val (3 000) | test (15 649) | held-out domain (2 493) | held-out period (2 574) | numbers / placeholders (test) |16|---|---|---|---|---|---|17| M\* HTR-Tiny (Gate E, reference) | +0.42 ± 0.14 | +0.44 [+0.30, +0.58] (confirmatory) | −0.10 | −0.53 | 0.991 / 0.992 |18| **HTR-Base 335 M, H100** (`medium_m1_base_lambda`; denoiser 30 k steps val CE 1.225; FT best at step 1 000, CE 1.266 — lr 1e-4 overfits afterwards) | +0.83 ± 0.13 | **+0.69 ± 0.05** | +0.29 ± 0.13 | −0.12 ± 0.10 | 0.992 / 0.993 |19| **T3: flan-t5-large 780 M, STRAP-style inverse paraphrase** (`medium_b2_flan_t5_large`, system `B2_flan-t5-large`; best val CE 0.753 at 5 k of 11 k steps) | **+2.04 ± 0.15** | **+1.79 ± 0.07** | **+1.27 ± 0.18** | **+1.09 ± 0.14** | 0.986 / 0.992 |20| **T3b: Qwen2.5-1.5B SFT, decoder-only LLM backbone** (`medium_llm_qwen25_15b`, system `LLM_qwen2.5-1.5b`; best val CE 0.733 at 4.5 k of 6 k steps, lr 1e-5) | **+2.19 ± 0.12** | **+1.97 ± 0.06** | **+1.78 ± 0.13** | **+1.30 ± 0.14** | **0.997 / 0.998** |21| HTR-Base seed 2 (`medium_m1_base_lambda-s2`, denoiser CE 1.231, FT best at step 1 000 CE 1.267) — replicates Base | +0.67 ± 0.14 | +0.63 ± 0.06 | +0.32 ± 0.14 | −0.07 ± 0.10 | 0.991 / 0.992 |22| HTR-Large 1.07 B from the *half-trained* stage-2 denoiser (CE 1.628; stages 1–2 diverged) — under-trained, **negative**; stage 3 also diverged (INCIDENTS I4), no further Large runs | −0.64 ± 0.12 | −0.71 ± 0.05 | −1.15 ± 0.13 | −1.62 ± 0.12 | 0.987 / 0.987 |23| HTR-Base FT v2, lr 2e-5 (`medium_m1_base_lambda_lr2e5`, from the seed-2 denoiser; best CE 1.259 then rising — Base over-fits the pairs at any lr) | +0.68 ± 0.13 | +0.55 ± 0.06 | | | |24| HTR-Base on MPS (M4M64c, `medium_m1_base_mediumdenoiseinit`; denoiser 24 k small-batch steps, CE 1.959 vs 1.225 on the H100) — under-trained, superseded by the H100 Base | −0.21 ± 0.13 | −0.26 ± 0.06 | | | |25| Qwen2.5-1.5B SFT, **seed 2** (`medium_llm_qwen25_15b_s2`, best val CE 0.7332) — replicates seed 1 | +2.17 ± 0.12 | **+1.97 ± 0.06** | +1.75 ± 0.13 | +1.28 ± 0.14 | |26| **T3: flan-t5-xl 2.85 B** (`medium_b2_flan_t5_xl`, system `B2_flan-t5-xl`; best val CE 0.688 at 5 k of 9 k steps, lr 5e-5, grad-ckpt) | **+2.56 ± 0.16** | **+2.39 ± 0.07** (CI above the +2.0 SESOI) | +1.78 ± 0.18 | **+2.12 ± 0.15** | 0.985 / 0.992 |27| **T3b: Qwen2.5-3B SFT** (`medium_llm_qwen25_3b`, system `LLM_qwen2.5-3b`; best val CE 0.701 at 6 k steps, lr 8e-6) — best fidelity of all systems | **+2.42 ± 0.13** | **+2.27 ± 0.06** (CI above the +2.0 SESOI) | **+2.05 ± 0.13** | +1.91 ± 0.14 | **0.998 / 0.999** |28| HTR-Large, final attempt: FT lr 1.5e-5 from the stage-3 denoiser (`medium_m1_large_lambda_lr15e6`, best CE 1.522, stable) — still **negative**; Large track closed (I4) | −0.60 ± 0.10 | −0.79 ± 0.05 | −1.05 ± 0.11 | −1.10 ± 0.10 | 0.988 / 0.989 |29| HTR-Small v2 (M3U96b; denoiser 22.5 k steps CE 1.849, FT CE 1.520) | −0.13 ± 0.14 | — | — | — | 0.987 / 0.993 |30| HTR-Edit (non-generative tagger, M4M64a) | −2.27 [−2.54, −2.03] | not scored | — | — | 1.000 / 1.000 |31| T1 decoding (Tiny M\*, 1 000 val pairs, K = 8 nucleus samples): greedy +0.54 → MBR +0.66 → MBR + source anchor +0.70; oracle best-of-9 +1.74 | | | | | `artifacts/metrics/decode_search_tiny_val/` |32| T2 MRT on Tiny M\* (α 0.5 and 0.005, lr 1e-5/2e-5, 750–800 steps): val chrF 68.96 → 68.80–68.86 | **negative** (abandoned) | | | | `mrt_tiny_mstar_a005`, `mrt_tiny_mstar_a05_abandoned` |33| T4 discriminator RoBERTa-large original-vs-rewrite (`disc_roberta_large`) | val AUROC 0.994, acc 0.976 | | | | reward/reranking signal; not yet used in a reported system |3435Reading: the pretrained English prior is the lever (the T5 backbone gains 4× M\* and, unlike every HTR model, its gain **transfers to the held-out36disciplines and to the pre-1900 books**); scale of our own Transformer helps modestly (Base +0.69 vs Tiny +0.44 on test). Running / queued:37flan-t5-xl 3 B (instance 1), Qwen2.5-1.5B SFT (instance 3), HTR-Base FT v2 lr 2e-5 (instance 2), HTR-Large stage 3 with the fp32 copy mixture38(instance 2; stages 1–2 diverged at lr 2e-4 / 8e-5, see INCIDENTS I3). Pangram (same 54-pair test sample, `artifacts/metrics/pangram_test_htr2/`, 7.52 USD): flan-t5-large 54/54 "Human" (mean ai_likelihood 0.000), HTR-Base 53/54, **Qwen2.5-1.5B 48/54 (6 "Mixed", mean 0.032; `pangram_test_qwen`, 3.64 USD)** — the lexically closest system is *not* the most detector-clean; reported as detector behaviour only (D21). RQ6 memorisation audits on validation (`artifacts/metrics/audit_qwen15b_val`, `audit_flant5large_val`): max Jaccard to HTR training originals 0.000, no same-passage leaks, for both backbones — these audits cover the HTR pairs only, not the backbones' own pretraining corpora.3940**Exploratory test report with all systems** (`artifacts/metrics/medium_test_htr2_exploratory/`, 15 649 pairs, 2 000 document-clustered resamples):41chrF++ B0 68.91 [68.35, 69.48] · Tiny M\* 69.35 · Base 69.60 · flan-t5-large 70.71 [70.19, 71.24] · **Qwen2.5-1.5B 70.88 [70.33, 71.42]**.42Stylometric distance W1 to the originals (rewrite 0.154): **flan-t5-large 0.083 (−46 %)** · Tiny 0.104 (−32 %) · Base 0.108 · Qwen 0.123 (−20 %). **Final report (2026-11-11 01:31 UTC, 27 systems, `report.json`)**: flan-t5-xl **+2.39 [+2.20, +2.57]**, W1 0.081 · Qwen2.5-3B **+2.27 [+2.13, +2.40]**, W1 0.100 · Qwen 1.5B s2 +1.97 [+1.83, +2.11] · Base MPS −0.26 · Large lr 1.5e-5 −0.79 · Large stage-3 FT −1.25 · Small v2 −0.09 · HTR-Edit −2.20 [−2.36, −2.05].43Reading: the two backbones split the objectives — Qwen is the closest lexically and keeps the most numbers/placeholders, flan-t5-large moves the44text furthest towards the pre-2018 stylometric distribution and is the only backbone with 0 % Pangram flags; the paper reports both, no single "best".4546## Medium-stage pairs (Gate E data, 2026-11-10; `data/pairs/medium_v1_haiku/validated/validation_summary.json`)47| Item | Value |48|---|---|49| Rewriter | claude-haiku-5-5 (author decision D20), prompts rw-1.0, protocol 1.1, Message Batches |50| Requests | 120 000 on 20 000 passages (train 14 000 · val 2 000 · test 3 000 · held-out domain 500 · held-out period 500) |51| Outcome | 119 724 ok · 88 refused · 188 truncated · **13.52 USD** (project total 43.20 USD) |52| Validation (`val-1.0`) | **103 934 passed (86.6 %)**, 16 066 quarantined (numbers missing 10 700 · entity recall 2 682 · formatting 784 · extra numbers 454 · placeholders 451 · negation 434 · near-verbatim 426) |53| By variant | A 17 558 · B 16 815 · C 17 881 · D 18 043 · E 33 637 (of 20 000 / 40 000) |54| Pairs by source | PMC 63 726 · Crossref econ/finance 20 028 · Gutenberg 12 862 · Bookshelf 4 948 · arXiv 1 398 · OpenStax 972 |55| Pairs by discipline (top) | biomedicine 12 043 · **economics 11 818** · life sciences 10 789 · **finance 10 737** · social sciences 10 716 · engineering 9 379 · CS 8 098 · chemistry 6 047 |56**Medium validation (exploratory, 3 000 val pairs, 2 000 bootstrap; `artifacts/metrics/medium_val_exploratory/`):**5758| System | chrF++ | edit sim | number recall | placeholders | W1 to originals ↓ | paired chrF vs B0 |59|---|---|---|---|---|---|---|60| B0 identity | 68.42 [67.60, 69.23] | 0.678 | 1.000 | 1.000 | 0.164 | — |61| B1 generic paraphrase | 65.67 | 0.566 | 1.000 | 1.000 | 0.226 | −2.75 |62| **HTR-Tiny, Medium denoiser init** (`medium_m1_tiny_mediumdenoiseinit`, val CE 1.47) | **68.75 [67.94, 69.54]** | 0.678 | 0.991 | 0.992 | **0.100** | **+0.33 [+0.09, +0.58]**, p_Holm 0.015 |63| HTR-Tiny, pilot denoiser init (`medium_m1_tiny_pilotdenoiseinit`, 16 k tokenizer) | 67.87 | 0.674 | 0.985 | 0.988 | 0.126 | −0.55 |6465| HTR-Tiny M3 architecture, token-only (`medium_m3tokenonly_tiny_mediumdenoiseinit`, val CE 1.49) | **68.78** | 0.680 | | | 0.106 | **+0.35 [+0.13, +0.59]**, p 0.002 |66| HTR-Tiny M2 metadata-conditioned (val CE 1.49) | 68.57 | 0.677 | 0.988 | | 0.098 | +0.15 [−0.08, +0.39], n.s. |67| HTR-Small 130 M (`medium_m1_small_mediumdenoiseinit`, denoiser 15 k steps val CE 1.85 → FT val CE 1.50) | 68.62 | 0.675 | 0.990 | 0.996 | **0.095** | +0.20 [−0.04, +0.45], n.s. |68| HTR-Tiny M3 token-only, seed 2 (val CE 1.48) | 68.73 | | 0.990 | | 0.105 | +0.31 [+0.09, +0.56], p 0.020 |69| HTR-Tiny M3-light (style/adv/con/sem/orth/fact at small weights; val CE 1.49) | 68.48 | | 0.991 | | 0.098 | +0.06 [−0.18, +0.32], n.s. |70| HTR-Tiny M1 seed 2 (val CE 1.474) | 68.81 | | | | 0.097 | +0.38 [+0.14, +0.63], p 0.003 |71| HTR-Tiny M1 seed 3 (val CE **1.470**, lowest → **M\*** by the frozen rule) | **68.84** | | 0.991 | 0.995 | 0.100 | **+0.42 [+0.20, +0.66]**, p 0.003 |72| **B2** T5-small (60 M) fine-tuned on the Medium pairs, 3 000 steps | 68.38 [67.60, 69.14] | 0.680 | 0.989 | **0.713** | 0.117 | −0.04 [−0.40, +0.30], n.s. |7374Reading: every Medium-denoiser-initialised HTR model sits at or just above identity on chrF (detectable gains of +0.15 to +0.35, all75below the pre-registered +2.0 SESOI) while **halving the stylometric distance to the originals (W1 0.095–0.106 vs 0.164)** at a cost of76≈ 1 pp in number/placeholder preservation. Metadata conditioning (M2) and model size (Small vs Tiny, same step budget) do not change the77picture; the Small model is under-trained relative to its size. Running: M3-light and M1 seed 2 (M4M64a), M1 seed 3 (laptop), B2 T5-small78on the Medium pairs (M3U96b). M\* selection follows the frozen rule (lowest validation CE among same-budget candidates).7980## Incident (2026-11-09 ~23:30 UTC)81M3U96b and M4M64a froze (no ping) under Medium-scale denoising pretraining (MPS memory near the limit) — `docs/INCIDENTS.md` I1.82The Medium tiny denoiser was relaunched on M4M64c with `PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.6`; the HTR-Small pretraining is on83hold until a 64 GB+ node is back. **If the two nodes do not come back by themselves, they need a power cycle.**8485## API spend (ledgers, Message Batches = 50 % price)86| Job | Model | Requests | ok / refused | USD |87|---|---|---|---|---|88| `pilot_opus` | claude-opus-5-5 | 5 844 | 5 801 / 43 | 26.85 |89| `pilot_heldout_models` | claude-sonnet-5-5, claude-haiku-5-5 | 600 | 600 / 0 | 1.43 |90| `pilot_heldout_prompts` | claude-opus-5-5 (H1, H2) | 300 | 297 / 3 | 1.40 |91| **Total** | | **6 744** | | **29.68** |92Measured mean tokens per request on Opus: 691 in / 323 out → 0.0046 USD per request in batch mode.9394## Medium corpus v1 (final acquisition, built 2026-11-09 night; `data/parsed/medium_v1/build_summary.json`, splits `data/splits/medium_v1/` with v0 assignments preserved)95| Item | Value |96|---|---|97| Documents kept | **14 470** (14 900 stubs; 2 parse errors, 35 post-2017, 61 undated, 36 licence-excluded, 16 non-English, 234 exact + 46 near duplicates) |98| By source | PMC 10 911 · Crossref econ/finance **1 730** · Bookshelf 1 286 chapters · Gutenberg 448 · arXiv 90 · OpenStax 5 |99| By licence | CC-BY 9 702 · CC-BY-NC 1 849 · PUBLIC-DOMAIN 1 737 · CC-BY-NC-SA 1 070 · CC0 109 · CC-BY-SA 3 |100| Disciplines | biomedicine 8 828 · life sciences 2 222 · **economics 1 289 · finance 564** · social sciences 363 · engineering 252 · chemistry 174 · CS 168 · earth 147 · physics 102 · statistics 84 · general reference 58 · humanities 56 · mathematics 45 · other 118 |101| Paragraphs / **eligible passages** | 897 419 / **451 072** |102| Splits | train 11 035 · val 1 270 · test 1 261 · heldout_domain 253 · heldout_period 510 · quarantine 141 (fixed-split conflicts) · 0 leaks |103104## Medium corpus v0 (built 2026-11-09 evening; `data/parsed/medium_v0/build_summary.json`, splits `data/splits/medium_v0/`)105| Item | Value |106|---|---|107| Stubs parsed / documents kept | 13 773 → 13 372 (2 parse errors, 35 post-2017, 42 undated, 36 licence-excluded, 16 non-English, 228 exact + 42 near duplicates) |108| By source | PMC 10 911 · Bookshelf 1 184 chapters · Crossref econ/finance 895 · Gutenberg 287 · arXiv 90 · OpenStax 5 |109| By licence | CC-BY 8 867 · CC-BY-NC 1 849 · PUBLIC-DOMAIN 1 478 · CC-BY-NC-SA 1 070 · CC0 105 · CC-BY-SA 3 |110| Disciplines | biomedicine 8 706 · life sciences 2 208 · **economics 689 · finance 319** · social sciences 318 · engineering 237 · chemistry 173 · CS 168 · earth 128 · physics 100 · statistics 84 · mathematics 45 · humanities 42 · general reference 37 · other 118 |111| Paragraphs / **eligible passages** | 660 108 / **371 419** (chapter 85 657 · other 57 689 · results 55 116 · methods 53 319 · discussion 44 692 · introduction 27 984 · abstract 13 252 · background 8 851 · conclusion 8 074 · data 6 632 · theory 6 202 · related work 1 999 · limitations 1 304 · preface 648) |112| Splits (pilot assignments preserved; held-out physics + earth sciences; pub_year < 1900) | train 10 268 · val 1 190 · test 1 166 · heldout_domain 257 · heldout_period 472 · quarantine 19 (fixed-split conflicts); 0 leaks |113| Tokenizer | `artifacts/tokenizers/medium_v0` — SentencePiece unigram 32 000 on 258 264 train-split passages |114| Pretraining | `denoise_m1_tiny_medium` (M3U96b, 30 000 steps, no API cost) running |115Acquisition still running when v0 was built: Crossref (895/1 730), Gutenberg (287/500), Bookshelf (1 184 chapters); v1 will be rebuilt when they finish.116117## Corpus (rights-verified, pre-2018, English)118| Source | Status | Documents / eligible passages |119|---|---|---|120| PMC Open Access (AWS `pmc-oa-opendata`) | pilot built; Medium-stage ingestion running (seed 3, target 10 000 more docs; 5 100+ fetched so far) | pilot 1 180 / 23 658 |121| arXiv q-fin + econ (CC-licensed, v1 < 2018) | adapter written; live smoke 12 docs; pool count in progress | see `data/manifests/discover_arxiv_*` |122| Crossref CC-BY economics/finance journals (16 verified journals) | adapter written; **1 730 eligible articles** counted; 15 fetched → 566 eligible passages | `data/manifests/discover_crossref_count.json` |123| Books: Project Gutenberg (public domain, nonfiction LoCC Q/R/T/H/B/G/L/Z) | adapter written; pool **7 799** English nonfiction texts; 12 fetched → 2 037 eligible passages, 10/12 dated from front matter | `data/manifests/discover_books_gutenberg_*` |124| Books: NCBI Bookshelf (BITS XML via FTP, 3 941 pre-2018 books) | adapter written; 3 books → 17 chapter stubs, 299 passages | `configs/data/books_sources.json` |125| Books: OpenStax / CNX pre-2018 editions (Internet Archive, dates from the PDF itself) | adapter written; 5 items | idem |126127## Open decisions for the author1281. Publisher robots.txt: MDPI and Springer disallow named AI crawlers; our fetcher uses a research user agent and the129   Crossref text-mining links for CC BY content. Confirm this policy (documented in `docs/RISK_REGISTER.md`).1302. Anthropic terms: publishing weights trained on Claude rewrites — recorded as a risk to confirm, not a conclusion.1313. SESOI thresholds in `docs/PREREGISTRATION.md` v0.1 must be confirmed before the v1.0 freeze.1324. **DECIDED (author, 2026-11-10): Haiku 5.5.** Job `medium_v1_haiku` submitted: 60 batches, 120 000 requests, pre-flight133   estimate ≈ 13 USD (batch); collection + validation run automatically when the batches end (D20). The earlier plan for reference:134   20 000 passages of `medium_v1` (train 14 000 · val 2 000 · test 3 000 · held-out domain 500 · held-out period 500) × 6135   generations = 120 000 requests. Pilot-calibrated estimate (Medium passages are longer than the PMC pilot's):136   **≈ 1 070 USD with Opus 5.5 (batch)**, ≈ 535 USD with Sonnet 5.5, ≈ 27 USD with Haiku 5.5; a 10 000-passage plan halves these.137   Per-request lower bound from the pilot's actual spend: 551 / 276 / 14 USD. Nothing has been submitted.138139## Known limitations at this point140- Discipline labels for PMC are heuristic (journal title / subject categories); relabelled once (165/1 180 changed).141- Pre-1950 PMC material is OCR and yields almost no eligible passages; historical style will come from books.142- Validator heuristics count spelled-out numbers as missing; entity matching is a proxy.143144## Public platform (2026-11-10)145**https://htr.spboucher.ai** — research preview serving **HTR-Tiny 0.3.0** (M\*, 33.5 M params, CPU inference ≈ 0.1 s per paragraph; the pilot model `HTR-Tiny-pilot-rc0` was served until 2026-11-10 morning): single-page UI (examples, metadata, word-level diff, preservation chips, warnings), JSON API `POST /api/reconstruct`, `GET /api/info`, `/model-card`, `/health`; 30 requests/min per client, 8 000 characters max. Deployed with `mld` on **m4mm:8540** (PM2 `htr-web`, tunnel route on BHS64); source = `m4mm:~/apps/htr` (git `htr-platform.git` on spbgit; laptop staging copy `~/Desktop/Cluster/htr-platform`). To ship a better model: copy a new release dir to `m4mm:~/apps/htr/model/` and `pm2 restart htr-web`.146147## Cluster usage (2026-11-10, author: "exploite le cluster encore plus")148All training runs on the Apple GPUs (MPS) of the nodes — there is no CUDA hardware in MacLustr. In flight: **M3U96b** HTR-Small v2 (denoiser resumed 15 k → 45 k steps, then 30 k fine-tuning steps), **M4M64c** HTR-Base 335 M (denoiser 24 k steps + fine-tune), **M4M64a** HTR-Edit non-generative tagger (agent), **m4mn** (new node, torch installed) M3 token-only seed 3 + M2 seed 2, laptop: confirmatory scoring + Pangram evaluation (agent). Memory caps on every node (`PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.7`, `LOW=0.5`).149