sec_edgar_ex96_pdfcommodityOfficial surveyAuthoritative

SEC S-K 1300 TRS - LLM extraction lane

SEC (extracted via OpenAI gpt-5.5)

DepositResource + DepositEconomic across the FULL EX-96 universe (HTML tables, Workiva div-rendered HTML, and PDF TRS), FY2021+

Source

How it's published

Feed (what to expect)

Ad-hoc OpenAI batch over the TRS universe; vendored JSONL, re-run per vintage.

Cadence: ad-hoc batch (re-run the OpenAI batch per discovery vintage)
Access: operator-vendored records.jsonl from `strata ex96-pdf submit|collect` (no network at store-build)
Format: JSONL (pre-parsed DepositResource + DepositEconomic records)
Licence: public domain (17 U.S.C. 105)
Vintage: 2026-06-16
Upstream: data/raw/sec_edgar_ex96_pdf/<vintage>/records.jsonl

Quirks

LLM-extracted via gpt-5.5 (OpenAI Batch API) - semantically structured, not deterministic table parsing
TEXT-first: pdfplumber/text-strip + page/chunk selection sends only the relevant pages, never whole-PDF vision
Supersedes the deterministic sec_edgar_ex96 HTML lane (rank 37 vs 38) - that lane's column-mapping mis-handles polymetallic tables and is kept QA-only
commodity_price economic metric carries the priced metal in commodity_raw

Bridge CacheOnlyFetcher: reconstructs the 1,167-record corpus persisted by the `ex96-pdf` lane (gpt-5.5 over all 456 SEC EX-96 exhibits). Emits DepositResource (reserves/grade) + DepositEconomic (NPV/IRR/opex/capex/mine-life/commodity_price). Re-collect with `strata ex96-pdf submit --force` then poll `collect --write`.

Reconciliation

Role in the canonical graph

Authority tier: Official survey
Provenance: Authoritative
Canonicality: tier 1
Primary for: —
Contributes IDs: —
Name-block matching: trusted
Reconcile profile: default (tier 1)

Schema

What it feeds & emits

Feeds (axes)

Supply

Record kinds

DepositResourceDepositEconomic

Identifier schemes

—