fredgar-ai

Leadership & stakeholder analytics — methodology and responsible use

This Fredgar AI subsystem extracts company leadership from SEC filings and computes a transparent “stakeholder orientation” (people-vs-profits) signal. Read this before relying on any output.

What it is — and is not

It is a way to (a) list officers/directors/owners disclosed in SEC filings and (b) summarize, with fully disclosed inputs, whether a company’s capital allocation over a period leans toward reinvestment (capex, R&D) or shareholder payout (buybacks, dividends).
It is not an approval rating, a popularity score, a personality/competence judgment of any named individual, an endorsement, or investment/HR/financial advice. Signals describe company behavior over a period, not a person’s character.

Every score decomposes into its source XBRL concepts and periods. Always verify material conclusions against the original filings (linked throughout).

Data sources

Source	Used for	Status
SEC Forms 3/4/5 (ownership)	Officer/director/owner roster, titles, role flags, tenure bounds, insider buy/sell	Implemented — legally public, parsed from EDGAR
SEC XBRL facts (10-K/10-Q)	Capex, R&D, buybacks, dividends, revenue → orientation signals	Implemented
Filing text (bios, human-capital, MD&A)	Quotes, initiatives, narrative	Foundation present (filing corpus + chunks); narrative extraction is the AI path below
Earnings call transcripts	Quotes, forward direction	Not scraped. Not on EDGAR and usually copyrighted/paywalled — ingest only user-supplied/licensed transcripts
LinkedIn / third-party profiles	Background enrichment	Not scraped (ToS/privacy). A pluggable `external` enrichment field on `Person` accepts licensed/manual data only

The orientation index (method v1.0)

orientation_index ∈ roughly [−1, +1], a weighted average of the available signals (missing signals are excluded and disclosed). Latest full-year (~annual) facts are used.

Signal	Definition	Weight
`allocation_balance`	`(capex + R&D − buybacks − dividends) / (capex + R&D + buybacks + dividends)`	0.40
`capex_intensity`	`capex / revenue`, scored vs a 5% reference — a local/physical-investment proxy	0.25
`rnd_intensity`	`R&D / revenue` — capability/people investment	0.20
`insider_alignment`	Net Form 4 acquired(+)/disposed(−) across leadership — directional “skin in the game”	0.15

Label: index ≥ 0.33 → “Reinvestment / stakeholder-tilted”; ≤ −0.33 → “Payout / shareholder-tilted”; otherwise “Balanced”.

Known limitations (important)

Outsourcing/offshoring is not directly observable from XBRL. capex_intensity is only a proxy for building physical/local capacity. True outsourcing signals require geographic-segment disclosures or the text/AI path — not this index.
insider_alignment counts all Form 4 share changes, including option exercises, grants, and tax withholding — not just open-market conviction buys. A refinement would weight by transaction code (P/S vs A/F). The raw value is always disclosed.
Employee-count / wage signals are omitted because there is no reliable, universal us-gaap XBRL tag for them; add them when a company’s human-capital data is structured.
Tenure (first_seen/last_seen) is a filing-date range, not an official appointment record.

AI analysis (Claude by default, grounded)

Narrative extraction — leadership initiatives, verbatim quotes, and stated forward direction — runs through a pluggable analyzer with Claude as the default provider (ENABLE_AI_ANALYSIS=true). It stays purely additive: it degrades to a no-op unless the optional anthropic SDK and a credential are present, so the structured, source-cited signal above is always the default output and a credential-less checkout makes no API calls.

How it stays grounded. When enabled, the analyzer (AnthropicAnalyzer, default model claude-opus-5-5) is sent only excerpts from this company’s already-ingested SEC filing text, each labeled [S1], [S2], …, and is constrained by a strict system prompt to:

Ground every item in the excerpts and cite the source label.
Reproduce quotes verbatim — never invent, paraphrase, or reconstruct a quote; if no verbatim quote exists, return none.
Make no personal, character, competence, popularity, or “approval” judgment about any named individual — only disclosed company/leadership actions and plans.
Use no outside knowledge and no speculation.

Output is validated against a JSON schema (structured outputs), persisted as a LeadershipAnalysis row with the used_sources it cited, and degrades gracefully: if the analyzer is off (or no credential/SDK is configured), no filing text is available, or the model call fails, a clearly-marked empty/error result is returned and stored — never a fabricated one.

Auth & setup. AI analysis is on by default; you only need the SDK and a credential to make it live. Authentication is token-first — a Claude Code OAuth token is preferred and sent as a Bearer token (with the oauth-2025-04-20 beta header); an API key is the fallback. Set only one — the API returns 401 if both are present.

pip install -r requirements-ai.txt                    # installs the optional `anthropic` SDK
export CLAUDE_CODE_OAUTH_TOKEN=$(claude setup-token)  # preferred: Claude Code OAuth token
# or, instead of the token:
# export ANTHROPIC_API_KEY=sk-...                     # fallback: metered API key
# ENABLE_AI_ANALYSIS defaults to true; AI_ANALYSIS_MODEL defaults to claude-opus-5-5

This is narrative extraction with citations, not an opinion about people, and not investment/HR advice. Always verify quotes and initiatives against the cited filings.

Commands & API

python manage.py sync_leadership --ticker TSLA --limit 25          # extract officers/directors
python manage.py compute_stakeholder_assessment --ticker TSLA     # compute the index
python manage.py analyze_leadership --ticker TSLA                  # LLM narrative (Claude; see above)

GET /api/v1/companies/{id}/leadership/ — roster with titles, roles, tenure, net insider shares
GET /api/v1/companies/{id}/stakeholder-assessment/ — index + decomposed signals + caveats
GET /api/v1/companies/leadership-compare/?cik=…&cik=… — compare leadership + orientation across companies
POST /api/v1/companies/{id}/analyze-leadership/ — run the LLM narrative analysis (admin only)
GET /api/v1/companies/{id}/leadership-analysis/ — latest stored narrative analysis (with caveats)

The static site renders a Leadership table and a Stakeholder orientation section (index, signal breakdown, and these caveats) on each company page, with CSV/JSON export. When a grounded AI analysis exists, it also renders a Leadership analysis (AI) section (summary, initiatives, verbatim quotes with source labels, and the AI caveat).

This site is open source. Improve this page.