spb/wpai0002_uqo

1# Decision log23Each entry records the decision, its rationale and the evidence. Entries D1–D14 were reconstructed on42026-11-09 from the code and documents in the repository at commit 43899c8 plus the uncommitted work of that day.56| ID | Date | Decision | Rationale | Evidence |7|---|---|---|---|---|8| D1 | 2026-11-09 | The project lives in `Recherche-UQO/ia/wp-ai-0002`, with id AI-0002 in the "AI" series and origin `gitsrv:srv/git/wpai0002_uqo.git` | follows AI-0001; no id collision | `CLAUDE.md` |9| D2 | 2026-11-09 | PMC is acquired **only** through the AWS Open Data bucket `pmc-oa-opendata` (per-article JSON + JATS, MD5-verified) | PMC FTP packages and the OA Web Service were retired in August 2026; the bucket is the official bulk channel | `src/htr/ingest/pmc_s3.py`, `docs/gates/GATE_A.md` |10| D3 | 2026-11-09 | Pre-2018 status is verified per document from JATS `<pub-date>` (ppub → collection → epub …), the arXiv v1 date or book front matter; PMCID ranges are only a sampling prior; deposit and upload dates are never used | mandate Phase 1: dates must be verified, not inferred | `parse/jats.py::_parse_pub_dates`, `ingest/arxiv_oai.py`, `ingest/books.py` |11| D4 | 2026-11-09 | Conservative licence policy: ND, NC-ND, TDM and ARXIV-NONEXCLUSIVE are tier C (no training by default); OTHER-OA and UNKNOWN are metadata only; the release set is CC0, PUBLIC-DOMAIN, CC-BY; the document's own licence statement overrides the provider field | open access ≠ open licence; rewriting is a derivative work | `licenses/policy.py`, `ingest/corpus.py` |12| D5 | 2026-11-09 | Splits are made **before** rewriting. The unit is the near-duplicate cluster ∪ author group (venue optional); disciplines and periods can be held out whole; assignment is deterministic by hash | prevents leakage through duplicates and authors | `data/splits.py`, `data/splits/smoke/splits_summary.json` (group_leaks = []) |13| D6 | 2026-11-09 | Placeholder protocol: citations, math and URLs are replaced by `⟦C/M/U n⟧` before any LLM call; rewrite and target are stored in placeholder form and restored at inference | removes the most fragile spans from the rewriting task and makes preservation checkable | `parse/protect.py`, `rewrite/engine.py`, `inference/pipeline.py` |14| D7 | 2026-11-09 | Rewriters: `claude-opus-5-5` is primary (effort low, Message Batches); `claude-sonnet-5-5` and `claude-haiku-5-5` are reserved as held-out models for RQ4; held-out prompts H1/H2 are never used in training; prompt version `rw-1.0` | pinned ids and prices; generalisation is a primary RQ | `configs/rewriting/pilot.json`, `rewrite/prompts.py` |15| D8 | 2026-11-09 | Label smoothing ε = 0.1 for L_token (0 in the overfit test) | standard regulariser for seq2seq | `models/losses.py` |16| D9 | 2026-11-09 | Numbers are kept as text, **not** placeholders. They are checked strictly (no number missing, ≤ 1 extra) and up-weighted through L_factual | numbers carry style (formatting, units) and must be learned; masking would hide factual errors | `validate/checks.py`, `training/data.py` |17| D10 | 2026-11-09 | From-scratch SentencePiece unigram tokenizer, trained on train-split text only, with atomic control and placeholder tokens | avoids test leakage and third-party tokenizer licences; makes the tokenizer part of the release | `models/tokenizer.py`, `scripts/train_tokenizer.py` |18| D11 | 2026-11-09 | Pointer/copy gate over source positions, on by default (`use_copy`); `--no-copy` ablation | most of the target is copyable from the rewrite; the mandate requires constrained preservation | `models/transformer.py`, `models/config.py` |19| D12 | 2026-11-09 | Pilot rewrite ceiling is **30 USD** (hard stop), batch mode; anything larger needs explicit approval | mandate: spend is opt-in | `configs/rewriting/pilot.json`, `docs/reproducibility/COMPUTE_BUDGET.md` |20| D13 | 2026-11-09 | The novelty claim is narrowed after the literature review: we do not claim to be first to invert LLM paraphrases (Rivera Soto et al. 2024, 2026); the claim is fidelity, rights-verified pre-2018 scientific originals, factual integrity, non-identifiability and memorisation controls, and discipline/era/section analysis | verified literature | `docs/RESEARCH_GAP.md` §5 |21| D14 | 2026-11-09 | Corpus scope widened: Medium stage (≥ 10 000 docs and ≥ 200 000 passages), economics and finance (arXiv q-fin/econ, Crossref CC BY journals), pre-2018 English reference books (Gutenberg, NCBI Bookshelf, OpenStax only if dated); a usable model release is a first-class deliverable; torch may be installed on other nodes | author additions | `docs/MANDATE.md` header, `CLAUDE.md` |22| D15 | 2026-11-09 | Mock and dry-run providers are the default; real calls need `--live`; request ids are deterministic and results append-only | resumability, no accidental spend | `rewrite/engine.py`, `scripts/rewrite.py` |23| D16 | 2026-11-09 | Model release format: SafeTensors + `config.json` + SentencePiece model + release metadata; usable API `HTRReconstructor`, deterministic decoding and preservation warnings | mandate Phase 7 | `training/trainer.py::save_checkpoint`, `inference/pipeline.py` |2425**Pending decisions**, to be logged when they are taken:2627- final SESOIs and the PREREGISTRATION v1.0 freeze hash;28- the size labels vs the measured parameter counts (RESEARCH_PLAN §5);29- the public release of weights trained on Claude rewrites (Anthropic terms, `LICENSE_MATRIX.md` §4);30- the hosting namespace;31- the B2 base model and its licence.3233## D-FREEZE — 2026-11-09 17:59 UTC3435PREREGISTRATION v1.0 frozen at commit `e74b71266d5c0e8c20f181a50df139f52c771d0f` (pre-specified in-session; SESOIs = v0.1 proposals, author confirmation pending). No confirmatory split had been scored with a model before this commit; only B0/B1 sanity checks on the smoke corpus.3637## D17 — 2026-11-09 (after the exploratory validation of the M1 pilot)3839**Decision.** Add a denoising pretraining stage on unlabelled, rights-cleared original passages (train split only; BART-style span masking, word dropout, local shuffling, rule-based paraphrase) before fine-tuning on (rewrite → original) pairs, and block repeated 4-grams at decoding. **Why.** The from-scratch M1-tiny scored chrF 32.9 vs 71.2 for identity on the validation split with degenerate repetitions: 3 900 pairs (≈ 1.7 M target tokens) cannot train a 27 M sequence model from random weights. Pretraining uses only data we already hold (no API spend) and is standard practice; it is reported as part of the model recipe, and token-loss-only fine-tuning without pretraining remains the documented control. Exploratory, before any confirmatory scoring.4041## D18 — 2026-11-09 (pilot diagnostics)4243**Finding.** Every M3 variant that included the factual term L_fact (weight 0.5, normalised by the number of protected/numeric tokens) failed to generalise (val CE ≈ 5.4 flat while train CE fell), and so did M1 + L_fact alone (`pilot_m1_fact_tiny`), whereas M3 token-only reached val CE 3.80 ≈ M1 (3.69). The culprit is therefore the normalisation of L_fact, not the dual-latent architecture nor the style/adversarial/contrastive terms. **Decision.** L_fact is now normalised by all target tokens (an extra weight λ_fact on protected tokens, bounded); presets unchanged. Seed variance on the pilot is large (M1 val CE 3.69 / 4.09 / 4.31 over seeds 1–3; M2 seed 2 = 5.03), so pilot-scale comparisons between architectures are not conclusive — the Medium stage is required.4445## D19 — 2026-11-09 (single-term diagnostics)4647**Finding.** L_contrastive alone (`pilot_m3_cononly_tiny`, val CE stuck ≈ 5.3) also prevented generalisation, L_semantic alone did not (best 4.08). Cause: six rewrites of the same original co-occur in length-bucketed batches and InfoNCE pushed their content vectors apart. **Decision.** Same-passage items are masked out of the InfoNCE negatives (`passage_ids` carried in batches). Presets unchanged; M3 and ablations are re-run as v3 after this fix.4849**D19 follow-up (v3 runs).** With same-passage masking the M3 results are unchanged (full 5.31, no-style 4.82, no-adv 5.30, no-con 5.29, no-fact 5.30): only 1.3 % of off-diagonal pairs in length-bucketed batches share a passage, so the mask is correct but was not the cause. At pilot scale (≈ 3 900 pairs) every auxiliary term except L_semantic degrades generalisation, the style/adversarial branch most (val CE ≈ 5.3 vs 3.6–3.8 token-only). The dual-latent hypothesis therefore remains **untested in a regime where it could work**; it is carried to the Medium stage with the denoising-pretrained backbone and the loss weights re-tuned on validation (λ_style, λ_adv ≤ 0.05 as starting points), and reported as a negative pilot finding.5051## D20 — 2026-11-10 (author decision)5253**Decision.** The Medium-stage rewriter is **claude-haiku-5-5** (author chose the Haiku option among Opus ≈ 1 070 / Sonnet ≈ 535 / Haiku ≈ 27 USD). Job `medium_v1_haiku`: 20 000 passages of `medium_v1` (train 14 000 · val 2 000 · test 3 000 · held-out domain 500 · held-out period 500) × 6 generations = 120 000 requests, Message Batches, prompts rw-1.0, placeholder protocol 1.1. **Consequence for the design.** The training rewriter changes from Opus 5.5 (pilot) to Haiku 5.5; the Opus pilot pairs (974 passages, 5 383 validated) and the Sonnet/Haiku pilot held-out sets become the cross-model generalisation material (RQ4), and the paper reports rewriter strength as a factor. Validation thresholds (`val-1.0`) are unchanged; acceptance rates are reported per model.5455## D21 — 2026-11-10 (author request)5657**Decision.** Two additions: (i) **RQ7 — detector sensitivity**: originals, LLM rewrites and HTR reconstructions are scored by the commercial AI-text detector Pangram (API key in `~/Desktop/Cluster/secrets/pangram/`, never in the repo; every call ledgered and cached). The measurement answers "how does a deployed detector react to inversion outputs"; a lower AI-likelihood is reported as *detector behaviour*, never as evidence of human authorship, and the dual-use implication (inversion as a paraphrase attack, cf. Krishna et al. 2023, Rivera Soto et al. 2024) is discussed in the ethics section. (ii) **HTR-Edit**, a non-generative, non-autoregressive text-editing model (LaserTagger/Felix family: per-token KEEP/DELETE tags + insertions from a phrase vocabulary learned from aligned rewrite→original pairs, encoder initialised from the Medium denoiser), which cannot hallucinate beyond its phrase vocabulary and leaves untouched spans byte-identical. Both are exploratory additions after the pre-registration freeze and are labelled as such.5859## D22 — 2026-11-10 (author request: cloud GPU, "more instances", "a logic that scores higher")6061**Decision.** (i) Full-scale training moves to rented NVIDIA H100 SXM instances (Lambda Cloud, us-southeast-1; author62provided the account and key; billing 4.29 USD/h per 1×H100 and 8.38 USD/h per 2×H100; `scripts/lambda_*.sh`): instance 163(1×H100) HTR-Base (335 M) denoising pretraining 30 000 steps → pair fine-tuning → predictions; instance 2 (2×H100) HTR-Large64(1.07 B, 40 000 + 30 000 steps, 8 192 / 6 144-token batches) and HTR-Base seed 2; instance 3 (2×H100) the "HTR 2.0" tracks65below. The cluster (MPS) keeps the Tiny/Small/seed/HTR-Edit runs. (ii) **HTR 2.0 — four pre-declared, *exploratory* tracks66to raise the lexical score above the identity ceiling**, each reported against the same B0/B1/M\* numbers, all tuned on67validation only and applied once to test: **T1 decoding** (nucleus sampling + MBR with chrF++ utility + source-anchored68reranking; `scripts/decode_search.py`); **T2 minimum-risk training** (Shen et al. 2016: expected 1 − chrF++ + fact penalty69over {gold, greedy, 8 samples}, CE anchor; `htr.training.mrt`, `scripts/train_mrt.py`); **T3 pretrained backbone** (STRAP-style70inverse paraphrase, Krishna et al. 2020: `google/flan-t5-large` 780 M fine-tuned on the same pairs/splits, system71`B2_flan-t5-large`; pre-2018 contamination of the backbone's pretraining data is a validity threat to be stated); **T472discriminator** (`FacebookAI/roberta-large` original-vs-rewrite on the train split, `scripts/train_discriminator.py`) as a73reranking signal and as a reward for a KL-regularised policy-optimisation variant (HUMPA, ICLR 2025 / AuthorMist 202574recipes with a source-fidelity term). The confirmatory results (Gate E) are **not** re-run; HTR 2.0 systems enter the paper75as exploratory additions with their own validation-tuned choices disclosed. **Why.** The frozen M\* beats identity by only76+0.44 chrF++ (below SESOI); the author asked for a mechanism that scores materially higher. The literature review77(STRAP; HUMPA; AuthorMist; Adversarial Paraphrasing, UMD 2025; MBR/MRT) points to pretrained priors and direct metric78optimisation, not to architecture. **Ceiling stated up front:** the best-of-9 oracle over Tiny samples is only +1.7 over79identity on validation, while the perfect-edit oracle (HTR-Edit labels) is +13: the headroom is in *knowing* the edit, not in80picking among samples. A "100 %" reconstruction is impossible by construction (non-identifiability, RESEARCH_PLAN §2).8182## D23 — 2026-11-11 (author decision: public release of the weights)8384**Decision.** The author asked to publish the six served models on the Hugging Face Hub under `spboucher04/` (`htr-tiny`, `htr-base`,85`flan-t5-large-htr`, `flan-t5-xl-htr`, `qwen2.5-1.5b-htr`, `qwen2.5-3b-htr`; `scripts/export_hf_cards.py`, `scripts/hf_publish.py`;86token in `~/Desktop/Cluster/secrets/huggingface/.env`, created through the author's logged-in Safari session). Model cards carry the87synthetic-output notice, the no-detection/no-integrity-use clause, the training-data provenance (rights-verified pre-2018 originals;88Claude Haiku 5.5 rewrites; pairs not released), the pretraining-contamination caveat for the backbones, the Pangram block with its89caveat, and the metrics of `docs/STATUS.md`. **Licence chosen for the fine-tuned weights: CC BY-NC 4.0 (non-commercial research use)**,90because the training pairs include CC BY-NC sources and LLM outputs; `qwen2.5-3b-htr` additionally inherits the Qwen Research Licence;91the `htr` inference code shipped inside `htr-tiny`/`htr-base` stays Apache-2.0. The open question of `LICENSE_MATRIX.md` §4 (provider92terms on training with Claude outputs) is thereby decided by the author for the weights; the pairs themselves remain unreleased.93