Every call is registered with its direction and horizon before the outcome exists, then judged against realized price when the horizon elapses. You get a structured verdict over REST or MCP — and the sample size, base rate, precision, recall and misses that go with it.
Not another feature vector. A scored one. Anyone can compute a signal; almost nobody publishes what happened next, at the stratum where the number was earned. Fahali's output arrives with the denominator attached — how many times this class of call has been made, how often the outcome was already going that way without it, and how much of the qualifying population it actually caught. That is the part you cannot backfill and the part you would otherwise have to build and wait a year to earn.
One structured verdict per instrument, over REST or MCP, shaped so a model or an agent can act on it without parsing prose.
Direction, the horizon it is claimed over, and the condition that would invalidate it — all fixed at registration time, before the outcome is known.
What was measured from a real feed, what is a proxy, and what is missing — as separate lists. A proxy is never returned as a measurement.
A confidence value is meaningless without the scale it was drawn on, so the scale ships with it. Where a figure could not be computed, the field carries the reason instead of a plausible substitute.
Precision and recall, the signal-conditioned base rate, and the effective sample — for the engine, horizon tier and instrument class the call came from. Never one platform-wide figure.
Not all of it — and this is the section most vendors do not write. Several of Fahali's own detectors have been measured, over large samples, as worse than their own base rate on a forward question. Microstructure imbalance, liquidity-wall and liquidity-vacuum reads, the microstructure-absorption proxy, and volatility expansion and compression are all in that group.
They were not deleted and they were not quietly left in the brochure. They keep running and feed the committee read as sensors, but they are gated out of making standalone published claims, and the judged record covers only the lanes that make judged claims. If you were planning to consume one of those as a forecast, the honest answer is that it has not earned that, and you would rather learn it here than from your own P&L.
Pull it yourself. The judged record is a public endpoint with no key: app.fahaliai.com/api/track-record/lead-time returns one row per stratum, and a stratum that fails the publication floor is withheld with a stated reason rather than rounded up. The methodology describes the judge and the bar it scores against.
Read-only by design. Fahali does not route orders, manage capital, or provide buy/sell recommendations. It surfaces structured intelligence so you and your systems can decide. Past outcomes are published including the misses; future outcomes are uncertain.
A structured verdict per instrument: direction, the horizon it is claimed over, a confidence with its scale, the evidence behind it separated into measured and proxy, what is missing, and the condition that would invalidate it. Available over REST and MCP. Fahali is read-only by design — it routes no orders and has no path to capital.
Every signal is registered with its direction and horizon before the outcome exists, then judged against realized prices when the horizon elapses. The judged record is published per stratum — engine, horizon tier and instrument class — with sample size, signal-conditioned base rate, precision, recall and the misses, at app.fahaliai.com/api/track-record/lead-time. Strata that fail the publication floor are withheld with a stated reason rather than rounded up.
Not all of them. Several detectors have been measured over large samples as worse than their own base rate on a forward question, so they are held back from making standalone published claims. They keep running and feed the committee read as sensors, but the published record covers only the lanes that make judged claims. The distinction is on the record rather than in the marketing: a detector that has not earned a forward claim is not sold as one.
Yes. A free agent key needs no card and is enough to evaluate the response shape and the judged record before any paid plan. Read access to the public market read and the judged record requires no key at all.
Nothing here is gated. Pull today's market read, or mint a free key and evaluate the response shape yourself.