What This Does and Does Not Claim

Canonical scope statement for public docs. Other pages link here instead of repeating full disclaimers.

Supported product claims

Disclosure Alpha does:

  • Parse 10-K and 10-Q HTML and extract named sections (Item 1A, MD&A, controls, etc.); 8-K via local HTML or MCP Builder only (see surface matrix below)

  • Compute deterministic text metrics, boolean flags, and section diffs — no LLM required

  • Produce reproducible 0–100 disclosure risk scores with versioned artifact strings in every response

  • Expose the same pipeline via CLI, Python SDK, HTTP API, and MCP

Form-type support by surface

Surface

10-K / 10-Q

8-K

CLI --html

Yes

Yes

CLI --ticker / EDGAR

Yes

No

HTTP ticker routes

Yes

No

MCP Analyst

Yes

No

MCP Builder

Yes

Yes (local HTML)

Scores summarize language and change signals in filings. They are research and integration tools, not trading signals.

What’s proven

Headline result: on 478 S&P 500 FY2025 Item 1A sections (deterministic_scoring_v2), company-specificity correlates ρ ≈ 0.87 with an independent NER-based measure (Spearman).

Full cohort counts, boilerplate construct validity (ρ ≈ 0.74), post-filing volatility association (Q5/Q1 ≈ 1.15), and limitations: Evidence and Validation.

Unsupported claims

Disclosure Alpha does not:

  • Provide buy/sell signals or return prediction

  • Replace reading the underlying SEC filing

  • Guarantee full S&P 500 index coverage in any empirical cohort

  • Claim earnings-surprise or other outcome prediction

  • Offer investment signals — scores are not validated as alpha

Language signal vs risk score vs investment signal

Term

Meaning

Language signal

Raw metrics (word ratios, flags, diffs) from filing text

Risk score

Weighted 0–100 components and headline overall_disclosure_risk_score

Investment signal

Not provided — scores are not validated as alpha

Deterministic and “no LLM required”

Given the same filing HTML and the same artifact versions (parser_version, metrics_engine_version, scoring_model_version, dictionary version), output is reproducible. No external model API is called in the scoring pipeline.

Version pinning: Versioning and Reproducibility.