FAULTLINE

What it is looking for

A company that is about to have a bad year rarely says so. What it does instead is become incoherent. The cash flow stops matching the profit. The inventory stops matching the sales. The KPI that headlined four consecutive quarters is not in the fifth deck. The word “record” is doing more work in the press release than it was last year. Each of these has an innocent explanation, and most of the time the innocent explanation is correct.

FAULTLINE’s whole premise is that the interesting signal is not any one of those things but the pressure of several of them at once, measured consistently, over time, against what management themselves said would happen.

Short-side equity research is the first experiment because a negative thesis is the cleanest thing to falsify: either the restatement, the auditor change, the going concern paragraph, the delisting or the drawdown arrives, or it does not. Being wrong is unambiguous and arrives on a schedule. That is a rare property in research and worth having early.

Contradiction intelligence for public companies is not the category, though. It is the test bed. Every mechanism here — a ledger of what was claimed, evidence bound to its source, a compulsory countercase, frozen scores — is the same machinery Strategy as Software points at private companies and their own board packs. Public filings just have the enormous advantage of being free, standardised, and thirty years deep.

The first real result, which was a null

The first full run went out on 29 August 2026 against fifty US retail and apparel companies, six of them since delisted — 5,701 filings, 82,173 point-in-time observations, 131 labelled negative events. The rules layer produced 852 contradictions across 588 snapshots.

Top-decile contradiction pressure had 0.94× the base event rate at 365 days. Deciles were not monotonic. M-score and accruals showed no lift either. In other words: on this sample, the rules predict precisely nothing.

That is written up in full, with the leaderboard and the bugs the real data exposed, because a result you only publish when it flatters you is not a result. It is also roughly what fifty companies and a drawdown-dominated event set were predicted to produce, which is why the honest gate is the wide universe rather than this one. The rules stay as the screen. The bet moves to the narrative layer, measured forward, scored within 24 hours of publication and frozen.

Two things the run taught, both slightly embarrassing and both useful. The rules are growth-quality rules — eight of ten need revenue growth, so a shrinking company barely registers, and Bed Bath & Beyond went bankrupt while sitting near the bottom of the leaderboard. And acquisition quarters swamp everything, because a rule that sees goodwill jump cannot tell a deal from a disaster without reading Note 2. Both are calibration items. Neither was visible from the design.

Principles it will not trade away

A contradiction means “investigate this”. It never means “this proves fraud”, and the language in every report is written to make that difficult to misread. Every conclusion keeps its provenance down to the accession number. Scores are frozen at publication and never revised. No model touches the backtest scoring path.

And nothing here is investment advice — it is a research instrument, run privately, on public filings, mostly to find out whether the idea survives contact with its own evidence.

Status

Internal. Runs daily and unattended, with a local web app for pipeline health and company drill-down, and an MCP surface so the reasoning can be done conversationally. Phase 2 — the wide backtest across every filer, with within-sector and within-momentum controls — is the gate that decides whether this becomes anything at all. It is designed so the idea can die cheaply, and before any money of consequence has been spent, which is a feature I have learned to build in on purpose.