Methodology

Model and policy versions

model version

arena-model@0.1.0

envelope version

cutoff-envelope@0.1.0

ruleset version

anniversary-s2@0.1.0

advisory policy version

advisory-policy@0.3.0

slot-pressure version

slot-pressure@0.1.0

snapshot manifest

m-us-3v3-2026070818

Slot gate graduation status

authorizing

CORPUS-GRADUATED — the dual rating-plus-slot gate governs the authorized state. Every predeclared objective passed on the synthetic corpus (docs/BACKTEST_REPORT.md §7). This is a corpus-graduation: it carries a standing requirement to re-confirm at the live-capture gate (G3).

Predeclared graduation objective (docs/SLOT_SAFETY.md §graduation)

These bounds were fixed before any combined result existed, and PR8's audit may not adjust a single number here — it only scores against them.

G-FS — False-safe non-increase, with real correction

The combined gate must not increase false-safe verdicts on any held-out fold versus the incumbent rating-only gate, and must additionally distinguish the two controlled dense-ladder cases (identical rating evidence, thin slot reserve) that the rating-only gate cannot tell apart.

G-REACH — Reachability floor

On held-out freeze points that are genuinely safe, the combined gate must still render SAFE TO SIT for at least six of every ten of them — a gate that is permanently abstinent fails regardless of its false-safe improvement.

G-ABS — Abstention bound

The INSUFFICIENT DATA rate under the combined gate may exceed the incumbent rate by at most 5 points on the held-out corpus, measured on a 0-100 point scale.

G-FLIP — Advisory churn bound

The advisory flip count under the combined gate may exceed the incumbent by at most one extra flip for every ten the incumbent already produced — a relative bound, not an absolute one.

G-LINEAGE — Not one lineage

The aggregate false-safe reduction must not be attributable to a single lineage: removing any one lineage from the aggregate must still show the combined gate at or below the incumbent.

Designed refusals, named

Every refusal in this app is a designed state with its own reason — never a blank card, a spinner, or a guess softened into "safe".

INSUFFICIENT DATA

the rating gate itself refuses — a designed refusal, never a guess.

ABSTAIN

an envelope/threat surface has too little history to fit a rung — renders as UNASSESSED, never as safe.

WATCHED_CHARACTER_ABSENT / WATCHED_CHARACTER_INELIGIBLE

the watched character has no position in the eligible set this snapshot — no fabricated position is produced.

UNRESOLVED_TIE / TIE_BREAK_INDETERMINATE

a load-bearing tie at the boundary cannot be resolved by the cited ruleset — refuses rather than breaking the tie with a stable-id fiction.

INCOMPLETE_PAGE_COVERAGE / MIXED_SCOPE / UNSAFE_SNAPSHOT_MANIFEST

upstream capture integrity failed — a position is never computed against a population the engine cannot trust to be whole.

Data quality

Backtest summary

Calibration history is the synthetic backtest corpus (docs/BACKTEST_REPORT.md), not live capture — every envelope verdict inherits that report's standing requirement to re-confirm on live capture once one exists. The only sealed real lineage today is a single six-snapshot capture window; the graduation objective above is scored on held-out synthetic seasons under leave-one-season-out, never on the live product.