Autopsies · public losses
Autopsies — Every loss graded, published, fed back.
Every losing EMI trade gets an autopsy. Dual-horizon postgame. Four loss buckets. An execution-vs-thesis split that distinguishes "right idea, bad entry" from "signal was wrong". Published in public, fed back into Phase 5 reliability weights — because that's how the model actually learns.
Dual-horizon truth
Right idea, wrong window?
Lots of losses aren't actually signal failures — they're execution failures. EMI's dual-horizon postgame stores both strict-horizon and grace-window truth, so a trade that hit target one day after the horizon doesn't get filed as "signal wrong". The autopsy keeps thesis validation and execution quality on separate axes.
The four buckets
Why this loss happened.
Every autopsy assigns a dual-horizon loss bucket. The bucket determines what Phase 5 learns next cycle.
| Bucket | What it means | Reliability weight |
|---|---|---|
| signal_wrong | Thesis didn't validate — strict or grace. | Negative delta |
| entry_wrong | Thesis validated late. We bought too early. | Mild delta |
| structure_wrong | Validated, but stop hit before structure confirmed. Bad geometry, not bad idea. | Mild delta |
| management_wrong | Validated, hit target after our exit. Trail too tight or partial too eager. | Process delta |
Only signal_wrong losses fully decay the source's reliability weight. The other three buckets train geometry and management, not signal trust. This is how the system distinguishes "the model is wrong" from "the trader is wrong".
Bounded learning
Per-update bounds, floors and ceilings.
The reliability-weight updater is bounded by ±0.05 per update and clipped by per-source weight_floor and weight_ceiling. A death-spiral guard freezes further negative deltas once a source sits at its floor — recovering signals always have a path back up.
Autopsies refresh automatically when shadow truth lands. A trade resolved on day 1 isn't frozen — if the grace-window outcome arrives on day 5, the autopsy re-files itself.