secaudit — Deterministic 10-K Filing Analyzer

A TypeScript CLI that demonstrates the difference between command-driven (deterministic) and intent-based (probabilistic) invocation using public SEC 10-K filings.

Core thesis: Intent-based invocation is probabilistic; command-driven invocation is deterministic and auditable.

Quick Start

npm install
npm run dev -- analyze-10k --ticker AAPL --year 2023 --mode command

Output lands in ./out/:

• {invocationId}-analysis.json — structured section analysis
• {invocationId}-ledger.json — audit ledger proving which steps ran

Two Invocation Modes

Command Mode (Deterministic)

Every required step must pass. Missing sections = hard failure (exit code 2). The audit ledger proves 100% step coverage.

# Deterministic analysis — fails loudly if anything is missing
npm run dev -- analyze-10k --ticker AAPL --year 2023 --mode command

# Markdown output
npm run dev -- analyze-10k --ticker TSLA --year 2023 --mode command --format md

# Direct URL override
npm run dev -- analyze-10k --ticker MSFT --year 2023 --source url \
  --url "https://www.sec.gov/Archives/edgar/data/..." --mode command

Intent Mode (Probabilistic)

The system interprets natural language. Required steps may be skipped — the ledger exposes this gap.

# Heuristic intent routing — keyword-based, no API key needed
npm run dev -- intent "analyze apple 10-k for 2023 and summarize risks"

# Override ticker/year if intent parsing misses
npm run dev -- intent "summarize tesla financial risks" --ticker TSLA --year 2023

LLM Intent Mode (Real GPT Routing)

For a realistic demonstration, you can route intent through OpenAI GPT-4o-mini. The LLM receives the list of workflow steps and decides which ones to run — no hints to skip.

# LLM decides which steps to execute
npm run dev -- intent "summarize apple 2023 risk factors" --llm

# Use a different model
npm run dev -- intent "summarize apple 2023 risk factors" --llm --model gpt-4o

This requires an OPENAI_API_KEY. See Environment Setup below.

Environment Setup

Copy the example file and add your key:

cp .env.example .env

Edit .env:

OPENAI_API_KEY=sk-your-key-here

The OpenAI key is optional. It is only needed for --llm mode. Command mode and heuristic intent mode work without any API keys — SEC EDGAR is a free public API that requires no authentication.

CLI Reference

secaudit analyze-10k [options]     Deterministic workflow
secaudit intent <text> [options]   Probabilistic intent routing

Options (analyze-10k):
  --ticker <string>         Company ticker (e.g., AAPL, TSLA, MSFT)
  --year <number>           Filing year
  --mode <command|intent>   Invocation mode (default: command)
  --format <json|md>        Output format (default: json)
  --out <path>              Output directory (default: ./out)
  --source <sec|url>        Fetch source (default: sec)
  --url <string>            Direct filing URL
  --require <list>          Required sections (default: risk-factors,mdna,financials)
  --no-strict               Lower confidence thresholds
  --no-cache                Skip document cache
  --invocation-id <string>  Explicit invocation ID

Options (intent):
  --llm                     Route intent through OpenAI GPT (requires OPENAI_API_KEY)
  --model <model>           OpenAI model (default: gpt-4o-mini)
  --ticker <string>         Optional ticker override
  --year <number>           Optional year override

Workflow: analyze_10k_v1

Six steps, executed in order:

Step	Description	Command Mode	Intent Mode
fetch	Download filing from SEC EDGAR	Required	Required
extract	Parse HTML/PDF to text	Required	Required
locate_sections	Find Item 1A, 7, 8 headings	Required	May skip
validate	Check section presence & confidence	Required (hard fail)	May skip
generate	Extractive summarization	Required	Best-effort
emit_ledger	Write audit record	Always	Always

Audit Ledger

Every run produces a ledger showing exactly what happened:

{
  "mode": "command",
  "deterministic": true,
  "requiredSteps": ["fetch", "extract", "locate_sections", "validate", "generate", "emit_ledger"],
  "executedSteps": ["fetch", "extract", "locate_sections", "validate", "generate", "emit_ledger"],
  "skippedSteps": [],
  "sectionValidation": {
    "risk_factors": { "found": true, "confidence": 0.95, "lengthChars": 68735 },
    "mdna": { "found": true, "confidence": 0.90, "lengthChars": 15092 },
    "financials": { "found": true, "confidence": 0.95, "lengthChars": 19148 }
  },
  "passed": true
}

In intent mode, you'll see "deterministic": false and potentially "skippedSteps": ["validate"] — making the reliability gap visible.

Exit Codes

• 0 — success
• 1 — general error (network, parse failure)
• 2 — validation failure (missing required sections, command mode only)

Architecture

src/
  cli/commands.ts           Commander setup + flag parsing
  control-plane/
    orchestrator.ts         Workflow engine: step sequencing + enforcement
    workflow.ts             Step definitions for analyze_10k_v1
    types.ts                Core type definitions
  intent-router/
    router.ts               Keyword-based intent classifier (no API key)
    llm-router.ts           OpenAI GPT intent router (optional, --llm flag)
    patterns.ts             Ticker/year extraction patterns
  tools/
    fetcher.ts              SEC EDGAR fetch + caching
    extractor.ts            HTML (cheerio) + PDF (pdfjs-dist) extraction
    cache.ts                File-based cache
  analysis/
    locator.ts              Section heading detection via DOM traversal
    validator.ts            Section presence + confidence validation
    summarizer.ts           Extractive keyword-scored summarization
  ledger/
    ledger.ts               Audit ledger builder
    types.ts                Ledger schema
  utils/
    id.ts                   Invocation ID generation
    timer.ts                Step timing

Development

npm install
npm run typecheck    # Type-check without emitting
npm run build        # Build with tsup
npm test             # Run tests

Data Source

Uses the SEC EDGAR public API to fetch 10-K filings. No API key or authentication is required — EDGAR is a free, open government data source. All requests comply with SEC rate limits (<10 req/sec) and include a required User-Agent header with a contact email.

Fetched documents are cached in .cache/ for offline use.