Ideagen Extraction Spec — What to Build

01 The play, in one line

Ideagen's entire cost base is analysts reading the sections our pipeline returns and hand-coding fields into datasets. If the item-3 rebuild emits those fields directly, that's a finished product we sell them — the highest-margin thing on the table, and it deepens the account instead of just serving search.

02 Where they actually look

They tell us exactly which form + section each dataset lives in, because they scope every search to it. This is the build target.

Form · section	Vol	Decodes to	Feeds dataset
10-K · 3-12	762	Part III Item 12 — Security Ownership	Insider pledging
20-F · 1-6 / 1-7	724/716	Item 6 Directors/Mgmt + Item 7 Major Shareholders	Pledging (foreign)
8-K · 2.06	614	Item 2.06 — Material Impairments	Impairments
10-K · 2-8 / 2-9a	131/98	Item 8 Fin. Stmts + Item 9A Controls	Out-of-period · ICFR
10-Q · 1-1 / 1-2	~60	Item 1 Fin. Stmts + Item 2 MD&A	Out-of-period · estimates
8-K · 4.02	46	Item 4.02 — Non-Reliance	Restatements
8-K · 1.05	ramping	Item 1.05 — Cybersecurity Incident	Cyber (new line)

Two structural must-haves: they always pair each form with its /A amendment (amendments are first-class), and their forms run wide — foreign (20-F/40-F) and funds (N-CSR/NSAR) are in scope, so extraction can't be US-10-K-only.

03 The field spec — WHAT TO CODE PER DATASET

Restatements / Non-Reliance

Priority 1

8-K Item 4.02 · 10-K/A · 10-Q/A

type: Big-R | little-rperiods_restated[]error_category[]accounts_affected[]magnitude_$ + dirauditor_involvedaudit_cmte_concludeddisclosure_date

Material Impairments

Priority 1

8-K Item 2.06

asset_type: goodwill|intangible|PP&Eamount_estimatecash_vs_noncashreasondate_concluded

Insider Pledging

Priority 2

10-K Item 12 · 20-F Items 6/7 · DEF 14A

personroleshares_pledgedpct_of_holdingslender/brokermargin_loan_flagas_of_date

ICFR / Material Weakness

Priority 2

10-K Item 9A

material_weakness_flagicfr_effectiveweakness_categoryremediation_statusauditor_attestation

Auditor Changes

Priority 2

8-K Item 4.01

former → new auditordismissal_vs_resignationdisagreements_flagreportable_eventseffective_date

Cyber Incidents

Strategic

8-K Item 1.05 · SEC 2023 rules

material_flagdiscovery_datenature: ransomware|access|…systems/data_affectedfinancial_impact_estthreat_actor

Out-of-Period Adjustments

Later — narrative

10-K / 10-Q · Item 8 notes + MD&A

immaterial_flagdirection: over|underamountaccountprior_periods[]reason

Changes in Estimate

Later — narrative

10-K / 10-Q · notes

estimate_typeeffect_directioneffect_$prospective_flag

04 Output contract — what makes it saleable

One ask that turns extraction into a product

Emit each dataset as a row keyed by accession + section, carrying the coded fields plus a source pointer — the paragraph/span each value came from. Ideagen's product is accuracy: they need every field traceable to the filing text, provenance is how they QA it and how we grade it, and booleans must be tri-state (yes / no / not-disclosed), never defaulted.

05 Build order & the pilot

Start with the clean 8-K items — 4.02 restatements, 2.06 impairments, 1.05 cyber. Single-topic, structured, easiest to code accurately; they prove the pipeline.
Then the well-bounded ownership/controls sections — pledging (10-K 12 / 20-F 6-7), ICFR (10-K 9A), auditor changes (8-K 4.01).
Do the narrative datasets last — out-of-period and estimate changes are highest-value but hardest to code; tackle them once the structured items are landing accurately.

Suggested pilot

Pick one structured 8-K item (4.02 or 1.05), run extraction over the matched sections, and grade precision/recall against a hand-checked sample — that proves the "section → coded row" machine end-to-end before wiring the harder datasets. Names will supply the exact candidate query + a labeled sample for whichever section sections picks.