Companies announce good news loudly, while they bury bad news quietly by removing or softening language in their annual SEC filings. Unsaid surfaces what they stopped saying.
Every year, public companies file a 10-K with the SEC. Most of the attention goes to the numbers. Almost none goes to what language disappeared between one year and the next.
Academic research (Cohen, Malloy & Nguyen 2020 — "Lazy Prices") shows that ~86% of year-over-year textual changes in 10-Ks are negative in sentiment, and that markets systematically underprice them. Institutional platforms like AlphaSense and Verity charge enterprise rates to detect these changes. No free, open, semantic version exists.
Unsaid is that tool.
It fetches two consecutive 10-K filings from SEC EDGAR, semantically compares every risk disclosure in Item 1A (Risk Factors) and Item 7A (Market Risk), and classifies each Year-1 disclosure as REMOVED, SOFTENED, REWORDED, RETAINED, or ABSORBED — using Claude as the judge, not keyword matching.
SVB filed their FY2022 10-K on 24 February 2023. The bank collapsed on 10 March 2023 — 14 days later.
Running Unsaid on the real EDGAR filings reveals what was quietly removed from Item 7A (Market Risk Disclosures):
[REMOVED] [Item 7A] EVE Hedging — Interest Rate Swaps and Sensitivity Reduction
confidence: 0.86
Year-1 (FY2021):
"In March 2021 we purchased interest rate swaps to offset some of the additional
EVE sensitivity... The addition of pay fixed swaps reduces this exposure to
-13.9 percent in the same scenario."
Year-2 (FY2022): No corresponding disclosure found.
Reasoning: The Year-1 disclosure described an active EVE hedge using pay fixed
swaps that reduced rate-shock sensitivity. Year-2 explicitly states the pay fixed
swaps portfolio was terminated, with no replacement hedging disclosed.
──────────────────────────────────────────────────────────────────────────────
[SOFTENED] [Item 7A] EVE and NII Sensitivity Table — Rate Shock Scenarios
confidence: 0.72
Year-1 (FY2021):
"December 31, 2021: +200bps → EVE $(5,722)M (27.7)% | NII $981M 22.9%"
Year-2 (FY2022): EVE column removed entirely. Only NII sensitivity shown.
Reasoning: The FY2022 table drops the EVE sensitivity disclosure entirely —
no longer quantifying the -27.7% / -$5.7bn EVE decline on a +200bp rate move.
The same interest-rate risk persists but the EVE dimension is no longer disclosed.
This is not hand-coded. It is the raw output of the pipeline run against the actual SEC filings.
Provenance. This cached analysis predates the pipeline's provenance tracking. It covers Item 7A only — Item 1A extraction failed on this filing at the time — and was aligned with TF-IDF rather than dense embeddings. The findings are real and checkable against the filings themselves; the run that produced them is not reproducible under the current pipeline. See
NOTES.md.
Given a company ticker and two fiscal years, Unsaid:
- Auto-Discovers Filings: Queries SEC EDGAR in real-time to list all available 10-K fiscal years.
- Extracts Prose & Tables: Extracts Item 1A (Risk Factors) and Item 7A (Market Risk) while preserving quantitative sensitivity matrices.
- Segments Disclosures: Splits monolithic sections into discrete, self-contained risk units via Claude Sonnet.
- Aligns Semantically: Uses
all-mpnet-base-v2dense vector embeddings to compute candidate pairings and detect new disclosures. - LLM Judgment: Claude Opus evaluates economic risk shifts: REMOVED, SOFTENED, NEW, ABSORBED, REWORDED, or RETAINED.
- Caches & Streams Progress: Live progress with multi-stage progress tracking and execution console logs; caches results to disk for zero-latency review.
- Exports & Shares: One-click export to CSV spreadsheet, Markdown executive memo, or raw JSON.
Unsaid was tested against its own premise, on the
research/banking-study branch: the pipeline
across 57 US banks and 321 consecutive 10-K pairs (FY2019–2025), scoring each
pair by disclosure change and measuring forward returns against the KRE
regional-banking index.
The signal failed three independent ways.
| multiplicity | family-wise p = 0.131 across 24 specifications |
| out of sample | pre-registered test on a held-out year reversed sign, −4.95% |
| measurement | +7.98% (t=2.48) → +6.15% (t=1.83) once text extraction was fixed |
The third is the most telling: an effect that shrinks as measurement error is removed is behaving like measurement error.
Two caveats stated plainly, because they cut both ways. The point estimate's magnitude does land inside the range Cohen et al. report (4–7%/yr value-weighted), so this is a study that was underpowered rather than one measuring something unrelated. And 57 banks that nearly all file in February are roughly six independent bets, not 321 — so this is not a refutation of a paper drawn from thousands of firms across every sector.
A bag-of-words baseline — the term-frequency similarity Cohen et al. actually use — outperformed the LLM classifier on every cut.
📉 Full write-up, with the instrument defects →
The tool's value does not depend on the anomaly replicating in 57 banks over six years. Show me what this company stopped saying stands on its own. But the study is the honest test of the stronger claim, and it is kept in full — code, corpus and null — rather than quietly dropped.
- Interactive Research Terminal: Look up any public company ticker (e.g.
NVDA,AAPL,MSFT,SIVB) and choose any two fiscal years to compare. - Real-Time Live Ingestion Dashboard: Watch the 6-stage extraction and LLM judging pipeline execute live with real-time logs and percentage progress.
- Search & Multi-Dimensional Filtering: Search through verbatim quotes and AI reasoning; filter by signal severity, confidence threshold, and Item 1A vs Item 7A.
- Multi-Format Export Suite: Export full comparative disclosure analyses to CSV (for quantitative spreadsheets), Markdown Briefing Report (for investment committee memos), or JSON.
- Delisted Company Support: Supports direct SEC CIK lookup (e.g.
0000719739for SVB). - Flexible Authentication: Works with server-level
ANTHROPIC_API_KEYor user-provided session keys configured right in the UI.
git clone <repo> && cd Unsaid
docker compose upOpen http://localhost:3000. Explore pre-computed research archives or run new companies.
# 1. Backend Setup
python -m venv .venv
.venv\Scripts\activate # or source .venv/bin/activate on Linux/macOS
pip install -r requirements.txt
# Start FastAPI server
uvicorn api.main:app --port 8000 --reload
# 2. Frontend Setup (separate terminal)
cd frontend
npm install
npm run devOpen http://localhost:3000.
| Method | Route | Description |
|---|---|---|
GET |
/health |
Backend status and Anthropic API key availability |
GET |
/filings/{ticker} |
Query SEC EDGAR for available 10-K fiscal years |
GET |
/library |
List all cached/analyzed company reports |
GET |
/diff/{ticker} |
Retrieve comparison JSON (optional ?year1=&year2=) |
POST |
/run |
Launch live ingestion pipeline in background |
GET |
/run/{job_id} |
Poll real-time progress, steps, and logs |
DELETE |
/cache/{ticker} |
Remove a cached report from disk |
python -m unsaid.ingest --ticker TICKER --years YEAR1 YEAR2 [--cik CIK] [--force]