Every idea this engine runs goes through the same pipeline: shadow trial → validation → a vote, or a grave with a cause of death. Nothing here is decoration — each entry is a real signal or assumption we ran, promoted, demoted or killed.
† A timestamp proves recency, not sanity.
Poisoned feeds served 2009-era prices under today's dates — three separate incidents (a $17 Microsoft, a $25 Signet, a $91 Meta). Buried under independent live-quote cross-checks on every source, and a chaos-test suite that replays each incident forever.
† A single keyless source cannot prove its own price scale.
Truncated histories from the last-resort feed poisoned two runs. Now: no independent verification, no price — the engine abstains instead of guessing.
† A composite score masked the absence of anything concrete.
One blended 'catalyst score' sounded rigorous and said nothing. Replaced by four specific, checkable catalysts: buybacks, insider clusters, 13D filings, debt traps.
† Deposits masqueraded as returns.
The account grew mostly from top-ups; charting it as performance would have been a lie. The public line is unrealized P&L vs cost — deposit-blind by construction.
† Free cash flow is meaningless for a bank.
Running owner-earnings DCF on banks, REITs and pipeline biotech fabricated numbers. A sector router now sends each to a method that fits — or to an honest abstention.
† A month of noise cannot judge a decade of cash flows.
Demoted, not deleted: 20-day scoring survives as an execution diagnostic. Doctrine now gets judged on 60-day checkpoints first, then 252-day — an external audit forced this one.
† Two contours priced two different markets.
The valuation model once priced a stock off a poisoned $91 while the screen saw $669 — and its false +30% margin was the number the money filter preferred. An invariant now rejects any name whose contours disagree by more than 20%.
An external audit argued our strict model double-punishes elite compounders. The re-derived model runs in shadow next to it; a pre-registered head-to-head resolves in October. Neither side gets to grade its own homework.
Wikipedia readership over 14/90/365-day windows as a demand proxy — spikes, builds and rollovers. Recorded next to every decision; earns a voice only if it predicts anything.
NT filings and auditor departures as early smoke. Shadow-journaled on finalists; the non-reliance flavor already graduated (see the active list).
Year-over-year rewrites of a 10-K's risk section — a documented value-trap predictor. Surfaced as narrative, never a veto, until the data speaks.
Whether curated value managers hold the name. Confirmation cue only — 45-day-stale public data never moves a score.
Blocks new buys after a 15% account drawdown. Under indictment: the audit argues it silences a value buyer exactly when entries get cheaper. Counterfactual cohort will be logged at the first real halt.
Sizing currently shrinks with the bear-bull spread of our own model. The audit calls it self-agreement, not risk. Pre-registered check runs against realized volatility.
Earnings-manipulation forensics. Started shadow, earned promotion: a raised flag now blocks a buy outright.
'Do not rely on our past financials' is the loudest sentence a company can file. Hard veto, no appeals.
Fundamental-health tiers gate conviction. Promoted from shadow with the forensics batch.
If a fair value needs implausible growth to exist, the fair value loses. Caps the model's own enthusiasm.
Born from the split-brain grave one scroll up. Every valuation must price the same market the screen sees, or it never touches money.
The one rule everything else serves: unverifiable data never becomes a fair value, and 'unknown' never masquerades as 'cheap'. The external audit named it the untouchable part — we agree.
Scoreboards: the Street vs the council · almost · the NO-ledger · how we price
The burial protocol comes from the owner's own research vault: every kill needs a named cause, so the same zombie can't be reborn under a new name. Research & education only — not investment advice.