Research log.
Every audit pass on every program, logged with date, version, and what changed. arXiv-style versioning across the six working papers. The current entry of each project lives in its body; the full history lives here.
The auto-research loop, instrumented.
Every pipeline run emits a RunManifest with code/data/config hashes alongside metric values, so candidate runs can be A/B'd against a baseline run id without re-deriving anything. Every proposed change to a prior — hand-shipped today, auto-proposed once the LLM loop ships — is logged in hypotheses.jsonlwith predicted delta, actual delta, and accept/reject decision. This panel is the loop's public state.
| Paper | Latest version | Runs | Hyp. | Accepted | Code hash |
|---|---|---|---|---|---|
| I. Researcher Restriction Index | v0.2 | 1 | 1 | 1 | a4887b4ef36e |
| II. Five-Year Plan Cascade | v0.2 | 28 | 28 | 0 | 2a2909b10b48 |
| III. Doctrinal Event-Response | v0.3 | 10 | 11 | 2 | 855b06d31e65 |
| V. AI Sentiment & Infrastructure | v0.2 | 10 | 10 | 1 | c90e28d6c72c |
Encode the BIS Oct-7-2022 regime-break by adding a derived feature `us_coauthorship_density_post_2022` = density × indicator(date >= 2022-10-07). The v0.1 logistic learns one coefficient across the full window; the v0.2 fix lets the model see the sign-flip.
brier: -0.006 · auc: +0.014
Sign-flip is empirically visible in the learned coefficients (pre-flip -0.439, post-flip add-on +0.361). Both metric deltas exceed the 0.005 acceptance threshold. The hypothesis was correct AND the effect size matched the prediction.
Replace unigram TF-IDF retriever with char-bigram TF-IDF (analyzer='char_wb', ngram_range=(2,5), max_features=20000). Fallback in place of the intended bilingual sentence-transformer, which stalled on CDN model download.
recall_at_3: -0.500 · mae_lag_months: +1.900 · subsidy_band_accuracy: -0.400
Honest negative result. Char-bigram retrieves cascades whose excerpts LOOK similar but whose realized sectors are unrelated. Confirms the v0.5 lift must come from semantic similarity (EN-paraphrase ↔ ZH-original), not richer surface features. Recorded as v0.2 baseline alongside the v0 keyword retriever; sentence-transformer code path stays wired for when ST weights cache.
Add TranslationVariant dataclass + translation_variants field to DoctrinalPassage. Populate from translation_notes.md across 5 anchors (6 variants, 2 doctrinally heaviest in Xi 2019 Taiwan).
framing_top1_accuracy: +0.000 · tier_mae: +0.000 · lever_f1: +0.000
Schema-only change. No metric movement expected and none observed. Accepted because it unblocks downstream work — the framing classifier can now feature-flag on variant_count or doctrinal_load. v0.3 should use translation_variants as a feature input.
Load 3 deferred anchors: Xi 2017 19th Congress, Xi 2021-08 common prosperity (CFEAC), Third Plenum 2024 communique. Widen corpus 12→15.
framing_top1_accuracy: +0.000
n=9 historical pairs is unchanged by adding anchors; backtest metrics are insensitive to corpus expansion until new pairs are added. Accepted as correctness improvement, not metric improvement. Need new historical event-response pairs in v0.3+ to measure lift.
Build a lexicon-based scorer (100 pos + 108 neg tokens, bilingual EN+CJK) and run it on the 60-article calibration set. First real model run for Sentiment paper.
mae: +0.227 · accuracy_3class: +0.717 · pearson_r: +0.757
Lexicon scorer beat the conservative predicted bar on every metric. Block-mean ordering reproduces the US-A +0.68 → US-C -0.33 → CN-C +0.42 arc — the load-bearing v0.1 claim survives a real model. Honest failure: CN-B/CN-C under-stated because bear-register tokens fire on Chinese-ascendant pieces that namecheck US slip. v0.2.1 LLM scorer benchmarks against MAE 0.227.
Researcher Restriction Index — 2022 sign-flip encoded as separate feature
- Open Question 03 resolved: US co-authorship density now has two coefficients — pre-Oct-2022 (protective, −0.439) and post-Oct-2022 (risk-loading add-on, +0.361). Effective post-flip slope collapses to −0.078.
- Sign-flip is empirically visible in the learned coefficients — the BIS Oct 7, 2022 regime-break is real in the data, not an editorial claim.
- Brier 0.088 → 0.082 (−7% error). AUC 0.908 → 0.922.
- observation_date added to LabeledProfile so positives use event date, negatives sample from event-date pool — temporal mix balanced across classes.
Five-Year Plan Cascade — Sentence-transformer code path (char-bigram fallback shipped)
- sentence-transformers installed via uv; bilingual ST code path wired in features.py + baselines.py.
- Model weights download stalled at CDN; fallback to char-bigram TF-IDF retriever shipped as the actual v0.2 baseline.
- Honest negative result: char-bigram recall@3 = 0.07 vs. keyword baseline 0.57 — confirms that the v0.5 lift requires semantic similarity (EN-paraphrase ↔ ZH-original), not richer surface features.
- When ST weights cache, the pipeline auto-promotes — no code change needed. retrieval_backend field added to backtest JSON so we can prove which retriever ran.
Doctrinal Event-Response — 3 deferred anchors loaded (corpus 12 → 15)
- §1.13 Xi 2017 19th Congress political report — verified on chinadaily.com.cn (new era + principal contradiction reformulation).
- §1.14 Xi 2021-08 common prosperity (CFEAC meeting) — verified on gov.cn (common prosperity + third distribution).
- §1.15 Third Plenum 2024 communique — verified on gov.cn (dating corrected from v0.1 placeholder 'third-plenum 2025' to actual Third Plenum 15-18 Jul 2024). Xinhua English 404'd; gov.cn Chinese is the verified primary.
- Corpus widened 12 → 15. PENDING_ANCHORS list now empty.
Comparative Tech Firms — Temporal premium series
- 6 pairs with ≥3 time points charted across end-2023 → mid-2026 (4 reference dates).
- Anthropic↔Zhipu detonates to −98% in mid-2026 once Zhipu's HKEX listing window blows up trailing P/S to ~1,066× — clean listing-venue rerating signature.
- TSMC↔SMIC narrows monotonically +145% → +141% → +60% → +40% — cohort's cleanest premium-narrowing series.
- xAI↔Minimax cleanest revenue-curve-maturation: +1,289% → +662% → +70% as both firms scale revenue past near-zero base.
AI Sentiment & Infrastructure — LLM scorer scaffold (pending API credential)
- Full LLM-scorer code path written: run_llm_calibration, llm_score, probe_llm_provider, --llm CLI flag. Provider-agnostic. Fallback chain: Anthropic Haiku 4.5 → OpenAI gpt-4o-mini → Gemini 2.5 Flash.
- Did NOT run: Gemini key returns FAILED_PRECONDITION (region block), Anthropic key commented out in ~/.zshrc, no OpenAI key configured.
- Honest pending stub written to data/sentiment/calibration_v021.json with status: pending_api_key. §V chart unchanged (lexicon-only); pending callout added below.
- To complete: export ANTHROPIC_API_KEY (or OPENAI_API_KEY) and run `cd projects/sentiment && .venv/bin/python -m forward_indicators_sentiment.calibration --llm`. Pipeline auto-populates without code change.
AI Sentiment & Infrastructure — Real reliability diagram (lexicon baseline)
- Lexicon-based scorer (100+108 tokens, bilingual EN+CJK) now scores the 60-article calibration set end-to-end.
- Headline: MAE 0.227 · RMSE 0.324 · Pearson r +0.757 · 3-class accuracy 71.7% (US 60% / CN 83%).
- §V scaffold replaced with a real 5-bin reliability diagram + block-level model-vs-hand comparison.
- Load-bearing claim survives: block-mean ordering reproduces the v0.1 US-normalizing-while-CN-ascendant arc. Lexicon scorer is now the benchmark for the v0.2 LLM scorer to beat.
Doctrinal Event-Response — translation_variants schema + heaviest surfaced
- TranslationVariant dataclass + translation_variants field added to DoctrinalPassage schema.
- 6 discrepancies populated across 5 anchors (Mao 1956, Jiang 1995, Xi 2012, Xi 2019 ×2, Xi 2024 NQPF).
- 2 doctrinally heaviest surfaced on §I of the page: Xi 2019 reunified clause (CSIS vs Xinhua) + Xi 2019/2022 use-of-force three-variant progressive hardening.
- Translation discrepancies are now queryable from Python — not just narrative in markdown.
Five-Year Plan Cascade — Calendar-month arithmetic fix
- _resolve_at switched from timedelta(days=30×m) → dateutil.relativedelta(months=N). Eliminates 1-day drift at 24 / 36 months.
- All 7 resolve_by dates now anchor cleanly on the NPC adoption day (12th): §2.1 / 2.6 / 2.7 → 2027-09-12; §2.3 / 2.4 → 2028-03-12; §2.2 → 2028-09-12; §2.5 → 2029-03-12.
- Page resolveBy strings updated to match Python. Page ↔ Python reconciled at the calendar-day level.
Comparative Tech Firms — Sector subgroup view
- New §VI: Aggregate stats now disaggregated by sector.
- Foundation models (n=4): median −75.75%, range −96.5% to +70%, Chinese-dominant 3 of 4 — largest dispersion (166.5 pp).
- Semiconductors (n=2): median −19.5%, mixed.
- Humanoid robotics +233% (n=1, Western), Payments +117% (n=1, Western — cleanest regulatory-overhang pricing pair).
Researcher Restriction Index — Python propagation sprint
- Page Table 1 brought into agreement with its own v0.1.1 revision log: §1.4.2 / §1.6.2 / §1.7.1 URLs now match what the log claimed was fixed.
- Python historical_events.py: Tencent inductive event actually removed (v0.1.1 log claimed removal but file was stale).
- Python URLs for §1.4.2 / §1.6.2 / §1.7.1 audited and confirmed already-pinned.
- Source-of-truth direction restored — Python ↔ page consistent.
Five-Year Plan Cascade — Python propagation + pipeline re-run
- §2.7 quantum propagated to Python (live_predictions.py).
- Pipeline re-run. predictions.jsonl regenerated with correct dates — every row was previously stale at 2027-11-21 (computed against the old 2027-03-15 anchor before v0.1.1).
- Registry now contains 12 predictions: 6 v0 15FYP drafts + 1 v0.1.1 quantum + 5 14FYP/13FYP backtest precedents.
- 1-day off-by-one between Python timedelta(30×m) and hand-curated page dates logged as v0.2 cleanup.
Doctrinal Event-Response — Python propagation sprint
- corpus.py: 11 of 12 verification_status fields flipped event_verified_quote_placeholder → verified.
- Primary URLs pinned in corpus.py for the 11 verified anchors (FLP mirrors, Stanford DigiChina, Xinhua, Qiushi, SCIO, etc.).
- historical_pairs.py §2.8 BIS Dec 2024: MOFCOM Notice No. 46 + CSIA/CAIA/CATAA/IAIA URLs added, window tightened 1 → 2 days.
- Translation discrepancies remain in markdown only — translation_variants schema field scheduled for v0.2.
Researcher Restriction Index — Primary-source pinning + cohort expansion
- §1.4.2 fabricated Bloomberg URL replaced with SCMP + Wire China + MIT Sloan triad.
- §1.6.2 originator corrected from Bloomberg to The Information (2025-03-14).
- §1.6.3 rescoped: Moonshot+Zhipu → Dec 2025 DeepSeek quiet rollout.
- §1.7.1 founding citation URL pinned (Bloomberg 2026-05-26).
- Internal cohort expanded 5 → 10 (5 real-named feature-vector validation profiles added under public-figure guardrail).
- Inductive Tencent pre-travel-review event flagged as likely over-extrapolation, removed from Python corpus.
- 11 of 14 unverified entity-count cells in §1 events table replaced with verified counts.
Five-Year Plan Cascade — Date integrity + quantum continuation
- Critical: 15FYP NPC adoption date corrected from 2027-03-15 → 2026-03-12 per NDRC PDF. Constant was off by 12 months; every downstream resolve_by shifted accordingly.
- §2.7 quantum information added as 7th live prediction (continuation cascade at state-lab magnitude band 1–5bn USD-eq).
- 5 cross-FYP verbatim excerpts pinned with Chinese + URLs.
- 36 ministerial follow-ons documented in cascade evidence pack.
- 23 monitor channels documented in 15FYP monitor pack.
Doctrinal Event-Response — Verbatim verification + translation discrepancies
- 11 of 12 anchors flipped event_verified_quote_placeholder → verified (primary URLs pinned to FLP, Xinhua, Qiushi, SCIO, Stanford DigiChina, CSIS Interpret, Creemers).
- PD Aug 2020 Dual Circulation remains placeholder pending People's Daily print archive retrieval.
- §2.8 BIS Dec 2024 enriched with MOFCOM Notice No. 46 + CSIA/CAIA/CATAA/IAIA industry statement.
- 6 translation discrepancies documented; doctrinally heaviest: Xi 2019 "reunified" clause + the three-variant 2019→2022 use-of-force hardening.
- "Section 1059" naming misnomer identified — FY2025 NDAA §1059 is unrelated to Dec 2024 BIS chip controls.
Comparative Tech Firms — Cohort expansion + valuation evidence
- Cohort expanded 9 → 12 documented pairs.
- Cohere ↔ 01.AI added (enterprise frontier model, revenue asymmetric).
- Mistral ↔ DeepSeek added (open-weights frontier, premium −82% to −89%).
- Stripe ↔ Ant Group added (payments scale, premium +117%) — cleanest regulatory-overhang pricing pair in the dossier.
- Valuation evidence pack pinned for 16 firms.
- Trajectory data captured for 24 of 36 cells.
AI Sentiment & Infrastructure — DC event reclassification + calibration scaffolding
- Meta 14 May 2025 row reclassified: previously logged as 0.30 GW scale-back; re-read of Susan Li's Q1 call shows reweight INTO AI, not DC retreat. v0.2 disposition: drop OR split scaled_back status.
- DC event evidence pack pinned (primary-source attempt across all 12 logged events).
- PRC capex evidence pack pinned (Alibaba FY25 Q3, ByteDance FT, East-Data-West-Compute II, provincials).
- 60-article hand-coded sentiment calibration set built — feeds v0.2 reliability diagram.
Lab — Initial release of all 5 working papers
- Working Paper I-01 — Researcher Restriction Index. 5 pseudonymized profiles, 17 historical §1 events, logistic regression vs. 3 baselines, Brier 0.088.
- Working Paper II-01 — Five-Year Plan Cascade Analyzer. 10 documented 14FYP cascades, 6 live 15FYP predictions, TF-IDF retrieval, leave-one-out backtest.
- Working Paper III-01 — Doctrinal Event-Response. 12 doctrinal anchors, 9 historical event-response pairs, live prediction for PRC talent-curb expansion.
- Working Paper IV-01 — Comparative Tech Firms. 9 matched West ↔ East pairs across foundation models, semis, payments, e-commerce, EVs, cloud.
- Working Paper V-01 — AI Sentiment & Infrastructure. 12 US AI DC cancellations, 15 Chinese AI capex entries, sentiment series, live 13-week-forward delta prediction.
arXiv-style. Each project page's body reflects its current reading. The full revision history lives in the appendix of each paper (§ Revision log) and is mirrored here as a cross-project feed. Audit passes are dated, versioned (v0.1.x), and named by sprint. The lab's practice is that prior framings are not silently overwritten — they sit in the appendix so any reader can see what we used to think and why we changed our mind. RSS at /research-log/feed.xml ships with v0.2.