A system that rewrites itself toward a score
q08 is not a fixed rule set. It has a target, the Brier score of its forecasts against the market price, and a loop that changes its own parameters, its own configuration and its own code to move that score, then checks whether it moved. This page lists every change the loop has made, read live from the tables the loop writes.
Honest status: the calibration tuner needs 50 settled calls before it may refit itself, and 47 more must settle. Until then it holds the neutral setting and only the product manager's knobs and experiments below are moving.
The loop
- 1 · Measure
Every forecast is hash-chained before its event and scored against the market price when it settles (Brier score). That score is the target the system optimises.
- 2 · Propose
Once enough calls have settled, the tuner fits a recalibration and a model-versus-market blend; the daily product manager proposes a change to a knob; the research lane turns a published method into a code task.
- 3 · Trial
A change runs as one experiment at a time with its own sample. Calibration changes start a new parameter version, so old and new are never mixed.
- 4 · Keep or revert
The result decides: the winner is written to the config, the loser is reverted, a domain where the model is significantly worse than the market is paused, and every step keeps its reason and its revert statement.
- 5 · Rewrite
A method that needs new code becomes a code task with a test; the code loop builds it, the deploy probes it, and the next measurement says whether it helped.
What it is running now
| Parameter version | 1 |
|---|---|
| Calibration (slope, shift) | 1 , 0 |
| Weight on the model versus the price | 0.5 |
| Domains paused for losing to the market | none |
Every change it has made
| When (UTC) | Kind | Change, reason and undo |
|---|---|---|
| 2026-11-12 23:35 | parameter | pause new signal entries until 2026-11-13T23:35:26.434Z drawdown 19.7% reached the 15% breaker undo: DELETE FROM pt_control WHERE key IN ('paused_until','paused_reason') |
| 2026-11-12 15:15 | knob | max_updates: 4 → 5 coverage 1, 810 tokens a forecast and 12 open questions: one more update an hour undo: DELETE FROM fe_config WHERE key = 'max_updates' |
| 2026-11-12 15:15 | knob | undefined: undefined → undefined |
| 2026-11-12 15:15 | experiment running | broadcast_format: arm registered-first sample still collecting |
| 2026-11-11 23:35 | parameter | pause new signal entries until 2026-11-12T23:35:24.373Z drawdown 15.6% reached the 15% breaker undo: DELETE FROM pt_control WHERE key IN ('paused_until','paused_reason') |
Changes to its own code
The research lane has read 0 published items and filed no code task yet. It files one when a method matches a named gap and scores above its threshold; the code loop then builds it with a test.
What it cannot change
Its own scoring rule, the hash chain, the paper-money limit and the rule that every forecast is written before its event. The target is fixed so the system cannot improve by moving the goal. Everything else is open to revision.