Decision IntelWhat it found before anyone knew the answer
Three public decisions that resolved badly. In each one the packet was read with retrieval off, entity names redacted and the outcome withheld, and the structural read was recorded before it was compared to anything. Below each is what actually happened. The gap between the two columns is the whole product.
Zillow (Zillow Offers)
No deceleration trigger: a leveraged, algorithm-driven bet on continued home-price appreciation with no kill-switch and no downside case if prices flattened.
Fermi America
Anchor-tenant / timeline mismatch — substantially all contracted capacity on a single tenant with no definitive leases for the remainder; a ~$20B valuation detached from pre-revenue fundamentals; capital committed to long-lead build before tenants signed.
WeWork (The We Company)
Founder-control governance concentration + a path-to-profitability unsupported by unit economics; "community-adjusted EBITDA" as an unfalsifiable accretion metric.
Why we do not call this a prediction
These are famous decisions, so a language model has read about their outcomes somewhere in training. Turning retrieval off proves the run had no route to the internet; it does not prove the model has no memory. What is defensible is narrower and still worth your time: the finding cites the document’s own language, the structural detectors that produced it are pure functions of the extracted text and the record’s structural spine recomputes identically on every export, and you can check every quote against the filing yourself.
The version of this that survives any scrutiny is forward, not backward. We audit a filing whose outcome does not exist yet, hash the record, and wait. There is no hindsight available to a timestamp. That clock is running and the entries are unresolved by definition, which is why they are not on this page yet.
What it did not find
One caution before any of these numbers. Our case library is weighted toward failures by construction, so a reader that flagged everything would look excellent on it. The informative rows are the disagreements, which is why the misses are here and no rate is computed anywhere on this page.
For every library case whose subject is a real SEC filer, we pulled that company’s own annual report from before the case year and ran the deterministic layer over it. One selection rule, no per-case judgment, no model in the loop. Fifty-six cases resolved to a filing; ninety-four were dropped by a rule that refused to guess.
Four pre-crisis banks read as recoverable
Lehman Brothers, Bear Stearns, Washington Mutual and AIG, each read from its own last annual report before 2008. Those filings carry leverage and funding structure in the language of routine disclosure, and the structural register does not treat a matched-book funding profile as a condition worth naming.
Recorded as one class rather than four accidents. Writing a rule to catch the four failures we already know about would be hindsight wearing a lab coat, and it would tell you nothing about the fifth.
Eighteen of forty-six failure-side cases produced no structural signal
Including the four above. The rest are named individually in the validation report rather than aggregated into a percentage. Two have a document-side explanation that is separated out instead of being counted against the engine.
A structural reader measures the snowpack, not the trigger. A company can carry a real concentration and compound anyway, and two exceptional successes in the same read came back fragile.
Two failure classes are absent from the register entirely
Clinical-trial outcomes and brand or demand collapse. Neither is expressible in the structural vocabulary the engine reads today.
Left as a stated gap. A detector written to close a gap it cannot actually see is how a rules library turns into a hindsight machine.
The protections column is shorter than the risks column
Across the merger register the engine credited a full protection 79 times against 1,282 unanswered flaws. Only four of twenty-three risk classes have ever produced a fully-credited protection.
That is a limit of our catalog rather than a finding about deals. It is here because a buyer will eventually ask, and the answer is better volunteered than discovered.
The comparison set a single firm cannot assemble
A partner sees their own deal flow. Every structural read is scored against populations of decisions that have already resolved, matched on structure rather than on sector or story. This is the part that makes "this shape has failed before, and here is the mechanism" a checkable sentence rather than an opinion.
Form S-4, 2012 to 2017
An outcome-blind merger census. Every registration in the window, selected so the filer is the acquirer and the outcome joins deterministically.
Schedule 13E-3 census
Hand-adjudicated Item 1013 and 1014 fairness reasoning. The closest thing the public record holds to an investment committee memo.
89% challenge rate
SEC staff review letters, and the amendments filed in response. What an unhedged assertion turns into once an independent reader challenges it.
What a claim becomes when someone pushes back
A comment letter is an allegation. An amendment is an admission. Select a challenge type to see what filers conceded once a reader with no stake in the deal asked.
The same test, on a deal you already know
Everything above is a public filing you can check. The version that matters to you is your own: send one closed transaction, we read the packet as of the decision date with the outcome withheld, and you compare what it named against what happened. You hold the ground truth, so you are grading us rather than trusting us.
Request a blind audit