spb/firc0001_uqo
Public
-
Thu, Nov 5, 2026 16
-
N8: rule-based baseline (snapshot rules.py) vs LLM on a stratified seeded sample of 60,000 ledger sections; rules-on-full vs rules-on-focused recall-loss bound
…
Items: FIRC0001-AI_C-19, THEO-09. 30,000 productive + 30,000 zero-mention sections (inverse-probability weights; DuckDB USING SAMPLE applied in subqueries), 200 s, 4 workers. P(rules >= 1 | LLM >= 1) = 85.7% (10-K 97.4, 10-Q 94.4, 8-K 69.2); weighted P(LLM >= 1 | rules >= 1) = 32.5%; type overlap on co-detected sections 84.6% (M&A 96.8%); rules on _focus(text) recover 26.2% of the rules-on-full-text mentions in over-budget sections (10-K MD&A 20.8%, 8-K 84.2%). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N13: export_stats.py (L1 revision) re-run read-only on M3U96b; 26/28 baseline CSV byte-identical, 2 identical up to tie order; stats.json 41 keys + 4 new exports
…
Items: FIRC0001-AI_C-11 DATA-01 DATA-08 ECON-09 ECON-13 EDIT-14 EDIT-17 THEO-13. 44 s; cross-checks vs the audits agree (gate 703,737; backfill 154,964 / 1,070; outage 5.25 h; 0.600 count 3,278; MD&A median 9,905; annual_panel 2,499). Flagged for the owner of export_stats.py: amount_raw_deviations 0 (NULL-blind <>, should be 584), october_wall_clock_hours 22.05 / 16.0 per s (measured from the backfill timestamp; 13.68 h / 31.9 rows/s), sp500_run.sections_logged 4,271 (should be 153,894); non-deterministic row order in two dumps. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N5: focus-window coverage of every ledger section (shown fraction, last offset, gate hits shown, plateau, short MD&As)
…
Items: FIRC0001-EDIT-08 THEO-09 AI_C-06 ECON-05 AI_C-20 DATA-08 THEO-13. _focus replicated span-exactly on 1,123,098 sections (0 mismatches; 23.4 GB streamed in 585 s); Python gate = DuckDB gate = 703,737. 10-K MD&A: median 9,905 words / 64,194 chars reproduced; median shown 9.3 %, last shown offset at 16 % of the text, 90.5 % of sections shown < 25 %; 62.7 % of all mentions come from sections shown < 25 %; 0 messages cut at 14,000 chars; mentions/section plateau ~1.5 beyond 20k chars. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N12: environment lock of the master node venv (uv pip freeze, 88 packages), .python-version 3.12.15, environment.md
…
Items: FIRC0001-DATA-10 (environment half). macOS 26.4, Apple M3 Ultra 28c/96 GB, CPython 3.12.15 (uv 0.13.4), duckdb 2.0.0, pyarrow 26.0.0, matplotlib 3.11.2. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N7: external consistency of 8-K mentions with 8-K item codes extracted from the body text (proxy, not precision/recall)
…
Items: FIRC0001-EDIT-01 (external benchmark), AI_C-19. Item codes found in 99.7% of the 400,601 8-K sections; P(ma_integration | Item 2.01) = 0.91 [0.90, 0.92] vs 0.03 for Item 5.02 (control); P(Item 2.01 or 1.01 | ma_integration) = 0.47; P(cost_reduction | Item 2.05) = 0.71 (lift 92x); by year, weight tercile and run. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N10: universe artefacts shipped with SHA-256; holdings CSV / company_tickers.json not archived; 2,573 -> 2,555 reconciled; missing filings; annual_panel = 2,499 CIKs
…
Items: FIRC0001-ADVE-15 AI_C-12 DATA-01 ECON-10 EDIT-16 LITE-09 THEO-14 (THEO-07, ECON-18). data/metadata/{universe,russell3000_constituents,sp500_constituents}.json + SHA256SUMS; 18 dual-class CIKs explain 2,573 lines -> 2,555 CIKs; 7 unmatched tickers absent from ticker_to_cik (10,430 entries); 2,587 and the 14/21 drop counts not reconstructible; 2 filings missing on disk (XEL 10-K 2014, LNT 10-Q 2012) + 14 parse errors; annual_panel/quarterly_panel carry 2,499 CIKs (paper: 2,501). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> -
N9: fixed-2010 cohort series and windowed-key (k = 3, 5) first-seen series for Table 14 / Figure 5
…
Items: FIRC0001-ADVE-13 ADVE-14 ECON-14 THEO-15. 2011-2025 first-seen range 6,240-8,216 (all issuers; 2010 = 7,835 left-censored, 2026 partial); fixed cohort of 1,389 issuers: new projects per issuer 4.7 (2011) -> 2.8 (2025) while mentions per issuer stay 9.6 -> 8.8; k = 3 / 5 episodes 131,823 / 126,318 vs 121,099. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N6: financing-instrument / buyback / dividend false-positive classes by item x type; Part I/II Item 2 collision test
…
Items: FIRC0001-EDIT-12 AI_C-13 THEO-08 (ADVE-01 ii). 22,412 mentions (7.3%) match the class: 54.5% of 10-Q unregistered_sales, 19.2% of 8-K, 52.6% of capex_program mentions; Table 5 restated side by side (capex_program 10,720 -> 4,725 projects excluding matched representatives); 1,001 unregistered_sales sections > 3,000 words (1.3%). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1 (follow-up): figure PNGs written by make_tables.py (dpi documented input) — orphan PNGs of FIRC0001-DATA-15 now generated
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N2: field-level grounding audit of all 308,065 mentions (years, locations, partners, amounts), fallback bounds, amount tail
…
Items: FIRC0001-EDIT-01 ADVE-01 AI_C-05 THEO-03 ECON-01 EDIT-02 EDIT-09 ADVE-03 ADVE-04 AI_C-04 AI_C-09 DATA-05 ECON-15. Section text streamed once (6.0 GB, 217 s, 4 workers), 1,000-replicate bootstrap clustered by section. Grounded: years 99.8%, partners 99.4%, amount number 93.7% (80.5% with currency token within 40 chars), locations 78.3% (8-K 67.8%), technology tags 63.6%; all structured fields 88.2% (8-K 76.6%). Upper bounds on regex backfill: years 98,438 / 252,469, technologies 16,949 / 36,740; max amount 1e12 (1 row), 161 >= 50 B, 24 < 0, 584 = 0 (amount_raw NULL). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1: de-literalise make_tables.py/export_stats.py; generated notes, macros (numbers.tex), prompt/regex from the snapshot, Table 12 longtable, figures (integer ticks, y from 0, per-issuer panel); stats.json never rewritten
…
Items: FIRC0001-AI_C-11, FIRC0001-ECON-09, FIRC0001-EDIT-17, FIRC0001-DATA-15, FIRC0001-ECON-13, FIRC0001-EDIT-14, FIRC0001-ADVE-13, FIRC0001-DATA-10, FIRC0001-EDIT-23, FIRC0001-ADVE-16, FIRC0001-AI_C-17, FIRC0001-ECON-18, FIRC0001-LITE-12, FIRC0001-DATA-01, FIRC0001-THEO-13; L2/L3 fragments: FIRC0001-ADVE-10, FIRC0001-AI_C-10, FIRC0001-DATA-06, FIRC0001-ECON-07, FIRC0001-EDIT-13, FIRC0001-LITE-08, FIRC0001-THEO-11, FIRC0001-THEO-12, FIRC0001-AI_C-15, FIRC0001-DATA-07, FIRC0001-EDIT-21, FIRC0001-AI_C-16, FIRC0001-EDIT-22, FIRC0001-EDIT-24, FIRC0001-LITE-07, FIRC0001-ADVE-06. export_stats.py edited only (run by N13 on M3U96b): run parameters from logs/ledger, new keys and dumps, read_only + threads=4. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N4: two runs, two 10-K segmenters, cohort regression, fiscal_year audit, 10-K funnel, git/mtime evidence
…
Items: FIRC0001-AI_C-07 EDIT-04 EDIT-19 EDIT-20 ADVE-02 ADVE-11 AI_C-03 THEO-06 DATA-13 ECON-03 ECON-08. No June git revision exists (first commit 0304133 on 2026-10-27); extractor files mtime 2026-06-17; df7905d..2324b58 = store.py +3/-1. Legacy MD&A carries 141 vs 272 '$' tokens (92% of pairs fewer); 86 full_document 10-Ks (35 invisible, 51 via legacy); 10-K funnel 341,207 -> 116,540 + 27,164 = 143,704 exact; cohort dummy +8.8 (all runs) vs -23.0 (October only). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N3: resolver composition, over-merge indicators, specification curve, ties, June->October reconciliation
…
Items: FIRC0001-ADVE-01 EDIT-07 THEO-01 AI_C-02 ECON-02 DATA-11 ECON-16 THEO-17 EDIT-06. _keyed rebuilt from resolve.py:17-32 (121,099 projects, identical id set); 9 keys x 5 thresholds; project count 21,027 (cik|type) .. 96,749 (type dropped) .. 121,099 (current) .. 198,488 (strict) .. 308,065 (mention); token 'acquisition' absorbs 40,407 mentions; 33,820 projects tie at max confidence; June block reproduces 39,930 / 19,227. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N1: ledger reconciliation (gate pass / true call count, backfilled block, outage, collisions, logs, throughput)
…
Items: FIRC0001-EDIT-02 EDIT-03 EDIT-18 ADVE-03 ADVE-05 ADVE-07 AI_C-01 AI_C-07 DATA-02 DATA-03 DATA-14 ECON-11 ECON-12 THEO-02 THEO-05 THEO-16 EDIT-05 ADVE-06 AI_C-08 DATA-04 ECON-06 LITE-07 THEO-10 DATA-08 ADVE-12 EDIT-15. Read-only DuckDB on M3U96b (12 s). 703,725 of 1,123,098 ledger rows pass the gate (true calls); backfilled block 154,964 = 153,894 + 1,070 rows < 30 words; 0 ledger rows inside the 5.25 h outage gap; 304 sections / 412 collided mentions; 14 silent JSON-decode failures outside the outage plus one anomalous post-resume batch (2,686 calls, 23 productive vs 536 expected); 149,352 reconciled from the log. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N0: audit infrastructure (read-only helpers, hash-verified 2324b58 snapshot, spid import shim, schema probe)
…
Stage D node batch N0 for FIRC-0001. sha256 of the nine snapshot files equals the node copies (extract/llm.py, rules.py, base.py, taxonomy.py, resolve.py mtime 2026-06-17); both DuckDB files open read_only=True; TEMP tables allowed; DuckDB temp_directory redirected to ~/firc0001_review/tmp. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-