How Tollbridge learns

The chain

1. Questions are asked — each paid call mints one decision with its frozen probability, hashes and lineage, before the outcome exists.
2. Outcomes resolve — an independent resolution engine evaluates each question against its data source; a decision resolves once, append-only, and a question that cannot be evaluated is recorded unresolvable.
3. Calibration accrues — resolved (probability, outcome) pairs land in published bands; each band publishes its empirical frequency and Brier score only at a 10-sample floor, and each question publishes at a 30-sample floor.
4. Release candidates are forged — recalibration, prior recounts and fusion refits are scored by offline walk-forward evaluation with availability-aware folds and a 30-row confirmatory floor.
5. Release gates decide — evidence floors (100 total / 30 per family), practical improvement (ΔBrier ≤ −0.005 with its block interval below zero), no-regression per family, and a fresh-process bit-exact rebuild. A candidate that fails is sealed rejected — and stays visible.
6. Activation is measured — an activated release changes the decision path; activation events, rollbacks, and the active release are public feed records, and every decision carries its release_version.
7. The index moves — the Decision Intelligence Index recomputes from the whole record: 100·(0.6·Brier skill + 0.25·calibration + 0.15·resolution rate), suppressed below 100 outcomes with the note "insufficient sample — no index claim".

Where players fit

Your calls are never inputs to the prediction you compete against. They aggregate — at a 10-graded-call floor — into a crowd benchmark published BESIDE the house record, not inside it. The separation is enforced by a repo-wide honesty gate: nothing from player calls is importable by the probability path. If the crowd aggregate ever feeds a model, that is a versioned intelligence release with its own methodology, never a silent default.

What your calls DO change: your rank, your quests, your badges, the crowd benchmark's sample, and the depth of the human-vs-agent comparison — measured evidence that grows every week you play.

The current state, publicly

The record is young. The live index publishes with its evidence counts; the learning curve is fitted from anchored walk-forward points; the release feed exists and is empty — no release has crossed its gates yet, and the house's skill against the base-rate reference is published with its sign even when negative. The ladder's next turn is blocked by BOTH the outcome milestone (public shortfall) and the release gate — which is the system working as designed.

The record · Calibration · The Arena · The ladder