A language model writes the prose. A deterministic layer deletes what it cannot support.

Where the model is, and where it is not

The forensic screens, the distress and earnings-quality math, the reverse-DCF and the grade are computed in Python from the filings. The model writes narrative on top of them: the thesis, the bull and bear read, flag wording, the plain-English explanations.

One exception, stated because a page about honesty that hid it would be worthless: the six business-quality bars are the model's own read. Four of them — Revenue Growth, Balance Sheet, Earnings Quality and Risk Profile — are capped down to the deterministic pillar covering the same ground, so a bar can never read stronger than the math behind it. Profitability and Outlook have no deterministic counterpart and are the model's read alone. Where the filings are too thin to compute a grade at all, the headline quality number falls back to that read as well.

The filter, running

Each example below is a real call: the left-hand text goes into the named function in explanation_guard.py, and the right-hand text is what that function returns. Struck-through sentences are the ones it removed. A test in the repository feeds these same strings to the same functions and fails if the output stops matching what this page displays, so the page cannot drift away from the behaviour it is describing.

Seven examples is a sample, not the layer. It is currently 50 checks across about 5.1k lines — counts measured from the module by that same test, so they understate rather than flatter.

The same sentence, two different filers

The checks above can mostly be read off the wording. This one cannot, and it is the clearest illustration of what "deterministic" buys you: Eight consecutive years is a checkable assertion about a specific filer. The guard compares it against how many fiscal years that filer has actually filed, so the verdict is a property of the company's history rather than of the sentence.

What the layer will not do

• It does not rewrite. Nothing is paraphrased into something more defensible — a sentence is kept as written or removed.
• It does not empty a field to win. Most checks refuse to delete the last surviving sentence, because a blank panel is a worse answer than a weak one and this codebase has watched guard chains erode notes into bare number lists before.
• It does not catch everything. It is a filter on shapes of unsupported writing, not a fact-checker over the whole filing. A wrong claim phrased with a real mechanism and a real figure passes.
• It runs per sentence, not per clause, for most checks — so a claim smuggled into a compound clause can survive one whose sibling clause supplies the grounding. Several checks carry clause-level variants for exactly this reason; not all of them do.

Where the data comes from

Everything the screens read is primary source: company filings from SEC EDGAR, and fund holdings from N-PORT, the report funds file with the SEC. There is no redistributed vendor dataset in the analysis path and no per-seat data licence sitting underneath your subscription — which is why fund look-through covers what it covers rather than what a data contract permits.

The calibration, with its limits

The grade is tested by replaying it at points in the past and measuring what happened next. Across 2,847 point-in-time observations spanning cutoff years 2016–2024, a name graded A or B went on to lose more than half its value 2% of the time (95% confidence interval 1.18–3.43%); a name graded D or F did so 12% of the time (9.74–14.88%).

Read that as a comparison between the grades and nothing more. The test universe deliberately over-weights companies already known to have blown up, so that there are enough disasters in it to measure a split at all. That makes the two rates comparable with each other and not with a market-wide base rate — neither figure is a claim about what share of ordinary stocks collapse.

The rest of what qualifies these numbers is published with them, not buried. The API returns the limits below alongside the figures above, and this page prints whatever it returns rather than a chosen subset:

• 1-year single-stock returns are wide and lopsided — 8 in 10 land between -43% and +80%. This is a quality/risk read, not a price forecast.
• Small-cap (under ~$2B) returns are noise over 1 year; the grade's edge is in large- and mid-caps.
• Cutoff years span 2016–2024 — most observations sit in the 2021–2024 windows, point-in-time approximated from as-filed XBRL, so later restatements aren't un-restated.
• The test universe deliberately over-weights known blow-ups and bankruptcies so the disaster-avoidance split is measurable — crash rates compare the grades against each other, not against market-wide base rates.
• 240 of these 2847 observations carry an AAER- or SAIC-based grade cap — too few yet to break out separately with statistical confidence, but the cap is included in every stat above, same as any other grade input.

Panel last recomputed 2026-07-30 · full method and track record

How to check any of this

• Run a company you already know well and read the narrative against its filings. The claim is not that it is always right — it is that it does not say things it cannot support.
• Compare the numbers on any analysis to the filer's own EDGAR documents. Every figure in the deterministic tables traces to a filing.
• Ask for a sentence that violates one of the rules above. If the layer is working, the failure mode you find will be a thin answer, not a confident wrong one.

Stockonomy is an educational research tool. Nothing here is investment advice, a recommendation, or a solicitation to buy or sell any security. Forensic signals flag probability, not certainty. Data is sourced from public SEC EDGAR filings.