Positioning
Deterministic SEC filing analytics without an LLM
Reproducible disclosure risk scores from filing HTML — same pipeline across CLI, Python SDK, HTTP API, and MCP. No token inference for scoring.
Deterministic pipeline vs LLM summarization
Disclosure Alpha scores SEC filing HTML with a fixed pipeline: section extraction (section_extractor_v1), text metrics (text_metrics_v4), boolean risk flags, year-over-year diffs, and weighted aggregation (deterministic_scoring_v2). Every response includes version strings for reproduction. LLM summarization can help humans read filings faster, but it is not a substitute for reproducible numeric scores. Model outputs vary across runs, lack stable version strings, and may paraphrase or omit filing text.
When LLMs are still useful
LLMs remain valuable for open-ended analysis, drafting, and agent orchestration. Disclosure Alpha's MCP Builder exposes deterministic tools (extract_sections_tool, score_filing_html_tool_wrapper) that agents can call — including 8-K raw HTML workflows that ticker routes do not support. Use deterministic scoring when you need comparable numbers across firms, time, and integration surfaces. Use LLMs when you need narrative synthesis on top of those tools.
Comparison
| Dimension | Deterministic pipeline | LLM summarization |
|---|---|---|
| Reproducibility | Identical input + version strings → identical scores every run | Non-deterministic; outputs vary across runs and model versions |
| Auditability | Artifact version strings in every response; inspectable pipeline | Opaque token generation; hard to audit or replay exactly |
| SEC filing fidelity | Scores derived from extracted section text and diffs — no paraphrase | Summarization may omit, merge, or hallucinate filing content |
| Cost at scale | Local CPU; no per-token API fees for scoring | Per-token inference costs scale with corpus size |
| Latency | Predictable; no network round-trip to an LLM provider | Depends on model size, context length, and API latency |
Not investment advice. Validation used Item 1A text for corpus scoring. See methodology and FAQ for supported claims.