System health · NFAI engine
One screen to answer, in a few seconds: is everything live, did today's data arrive, is the engine learning, and is any pipeline stale. As of 11:30 AM ET · market Open.
Engine status
Data source
LIVE
Market
Open
Data as of
July 31, 2026 (3d)
Morning signal
CALL 56%
Orchestrators
steps 7 ran · 1 warning · next Tue Aug 4, 13:37 UTC
steps 8 ran · 0 warnings · next Tue Aug 4, 3:45 UTC
Pipelines
expand a row to diagnoseNot instrumented
Shown as — rather than a reassuring colour. The console reports only what it actually measures; a green dot on a metric it can't see would be the one real lie on this page.
Engine learning · self-measurement
0 calibrated · 3 measuringThe engine is measuring 3 models against reality on 23 observations — all still gathering evidence, no parameters moved.
The stop model looks optimistic by ~10.79 bps (loss runs ~0.67% past plan), but this is provisional — 7 of 20 observations. No parameter has moved.
Mean adverse stop slippage
10.79 bpsoptimistic
Evidence
7 / 20 obs · 35%Current: 0 bps — loss assumed capped at the stop level. No change proposed while the sample is provisional.
Investment committee · earned trust
0 graded · 3 seatedThe committee votes equally: of 3 seated members, none has yet cleared its sample bar to earn trust beyond the equal baseline. 8 subsystems produce views that aren't graded yet.
2 graded, below the sample bar — holds the equal baseline until it earns more
14 graded, below the sample bar — holds the equal baseline until it earns more
7 graded, below the sample bar — holds the equal baseline until it earns more
Trust withheld · no evidence stream yet
Validation (descriptive, sample 20)
CALL accuracy
50% (10)
PUT accuracy
50% (8)
No-trade rate
90%
Day types: gap and go 7 · expansion 7 · range 3 · gap and fade 2 · trend 1
Acceptance: accepted down 5 · rejected up 5 · rejected down 5 · accepted up 4 · neutral 1
Descriptive bookkeeping, not predictive probabilities. Accuracy here is retrospective until the sample is large enough for L4.
Recorder activity
Learning progress · 20 sessions collected
Freshness uses weekend-tolerant thresholds on each pipeline's last run. Per-call API latency and failed-job history are not instrumented yet and are deliberately not shown rather than faked. Not investment advice.
Recalibration history
None yet. A parameter moves only when the sample clears 20 observations — acting on a thinner sample would be manufacturing certainty, which the engine refuses.
Source: 7 realized stop-outs (broker exports). As of 2026-07-31.
The morning call has been right 9 of 14 decisive sessions (64.29%) — provisional at 14 of 20 observations. No parameter has moved.
Morning-call hit rate
64.29 %unbiased
Evidence
14 / 20 obs · 70%Current: 1.0 — calls taken at their stated confidence. No change proposed while the sample is provisional.
Recalibration history
None yet. A parameter moves only when the sample clears 20 observations — acting on a thinner sample would be manufacturing certainty, which the engine refuses.
Source: 14 decisive directional sessions (session records). As of 2026-07-31.
The engine's calls have been right 2 of 2 decisive graded decisions (100%), and following and overriding are so far a wash — provisional at 2 of 20. No parameter has moved.
Engine-call hit rate
100 %unbiased
Evidence
2 / 20 obs · 10%Current: discretionary — every engine call may be overridden. No change proposed while the sample is provisional.
Recalibration history
None yet. A parameter moves only when the sample clears 20 observations — acting on a thinner sample would be manufacturing certainty, which the engine refuses.
Source: 2 decisive graded decisions (decision ledger, forward-marked). As of 2026-07-31.
Every completed prediction becomes an observation; measured error updates the mathematics, never a black box. Provisional models move no parameters. This section is the seed of NFAI's learning record — new measurable models (probability calibration, expected-vs-realized return, volatility error) join it as they come online.
These subsystems produce a view but nothing yet grades whether that view is right, so the committee assigns them no trust — a held-open seat, not a hidden vote. Each becomes a graded member once its predictions are measured against reality.
No member holds permanent authority. Trust is re-earned every close from measured performance; until a model clears its sample bar it votes at the equal baseline (1.00×). Bounded to [0.25×, 1.75×] so nothing dominates or vanishes on a short run. The engine listens more to what has been right — and it changes nothing until the evidence says so.
