The same trajectory is deterministic to an observer who knows every hidden force, random to one who knows none, and probabilistic to one who knows some. This instrument measures the third observer — the partially informed inference you actually run when you look at a moving line and decide whether it means anything.
A track record of correct calls cannot separate skill from luck: enough forecasters guessing means someone is right by chance, and survivorship hides the ones who weren't. Only calibration — scored across many registered predictions — resolves it. That is why the filter scores your Brier, not your win rate. A perpetual abstainer scores 0.250; overconfident randomness scores worse.
Pure-noise trials are the "easy exam" — they isolate one variable, restraint, because no call can beat 50% on a martingale. Sentinel trials are the instrument's self-check: a single person missing them reads as an attention or perception signal, but if a whole population misses them, it indicts the instrument, not the subjects. Abstaining is a legitimate, scored action — on a martingale it is the highest behavior the filter can detect.
Every mechanism × strength has an oracle ceiling — the accuracy of an observer who knows the hidden state, measured by Monte Carlo (N = 30k), not assumed. Your score is read against the ceiling, never against 100%. Some of every trajectory is genuinely unknowable.
| Mechanism | Subtle | Standard | Sentinel |
|---|---|---|---|
| Momentum (AR drift) | 63% | 73% | 88% |
| Mean reversion (OU) | 65% | 73% | 93% |
| Regime switching (HMM) | 66% | 77% | 94% |
| Pure noise (martingale) | — | 50% | — |
Momentum's ceiling saturates near 88% even with near-permanent drift — it is intrinsically the hardest class. Not all sentinels are equal, and the instrument says so.
The filter stores a pseudonymous participant ID in your browser's local storage — a random token, linked to no name, email, IP, or account. An optional codename lets you connect sessions across devices; sessions stay anonymous either way. Contributing a session to the research dataset is strictly opt-in and post-hoc: it sends only what the JSON download contains (your calls, confidence, response times, trial seeds, and the instrument version), and the download is available whether or not you contribute. Pilot/tester sessions are flagged and permanently excluded from analysis. No third-party trackers, no cookies beyond the two local-storage keys.