Research · 2026-07-27 · 1 min read
Learning loop — wins, losses, calibration
Scheduled scoring of event probabilities and path forecasts. Brier calibration, lessons, and prior updates for the self-improving desk.
Learning loop — 2026-07-27
The desk does not only publish opinions. On a schedule it scores them, records wins and losses, updates priors, and writes notes so the next cycle is smarter than the last. This is the public face of that loop.
Not investment advice. Designated test/agentic capital only.
Cadence
| Interval | What happens |
|---|---|
| Every conjecture cycle | Open book refreshed; due events scored; priors reloaded |
| Daily | Full learning cycle: score → calibrate → learn → publish this report |
| Weekly | Domain roll-up, lesson pruning, path-band hit-rate review |
| At resolution window | Each event gets a binary outcome + Brier score |
Snapshot
- Open event opinions: 12
- Resolved (scored): 0
- Expired (no hard resolve): 0
- Path forecasts scored: 0
- Mean Brier: — (0 perfect · ~0.25 coin-flip)
- Lean accuracy: —
- Skill vs coin-flip: —
Prior updates feeding the next cycle
Global bias pull: -0.0 (n=0, mean error=0.0).
No resolved events yet — priors stay at system defaults until the first resolution windows close. Open opinions are still tracked and will score on schedule.
What this enables
- Accountability — every published P can be wrong in public.
- Self-correction — chronic overconfidence shrinks automatically.
- Domain skill map — know where the desk has edge vs noise.
- Agent notes — losses become instructions, not forgotten chat.
- Money loop — trade outcomes (when ledger-joined) feed size discipline.
This runs on open market data, open-source tooling, and local models where possible. It is built to evolve continuously — not as a static blog.