You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Verification gate: pluggable deterministic checks as escalation trigger #1135
The escalation trigger is deterministic verification, not model self-report:
Define a verify_gate interface: named checks that return pass/fail + a machine-readable reason. v1 checks: project build succeeds; declared data tests pass; schema-contract diff (written models/columns match the task's declared contract when one exists).
Gate runs identically after primary AND escalation attempts (measurement needs both lanes' outcomes).
Configurable per project: route.gate: [build, tests, contract]; sensible defaults; a check that cannot run (e.g. no tests declared) reports "skipped", never blocks.
Gate results are part of the session record (consumed by measurement subtask).
Tests: each check's pass/fail/skipped paths; gate composition; failure-reason payloads.
changed the title [-]Cascade routing core: two-model user selection + escalate-on-gate-failure flow[/-][+]Verification gate: pluggable deterministic checks as escalation trigger[/+]on Aug 24, 2026
Evidence note from the prior-art research (parent #1134): gate discriminativeness is an ongoing investment, not a one-time checkbox — CHASE-SQL's execution selector captures only part of the oracle headroom (73.01% vs 82.79% oracle, arxiv.org/html/2410.01943v1) and TestPrune shows improving test quality on an already-gated pipeline adds +2.4 to +12.9pts (arxiv.org/html/2510.18270v2). The gate interface must make adding/refining checks cheap. Also: prefer project-declared tests over model-generated ones — the literature's verifier-gaming failure mode applies to self-generated checks.
Closing: after evaluating verifier-gated and cascade routing, we decided not to build an inferred per-request router. The direction is a strong default model with explicit, user-chosen modes. Reopen if routing is revisited.
Activity
[-]Cascade routing core: two-model user selection + escalate-on-gate-failure flow[/-][+]Verification gate: pluggable deterministic checks as escalation trigger[/+]anandgupta42 commentedon Aug 24, 2026
Evidence note from the prior-art research (parent #1134): gate discriminativeness is an ongoing investment, not a one-time checkbox — CHASE-SQL's execution selector captures only part of the oracle headroom (73.01% vs 82.79% oracle, arxiv.org/html/2410.01943v1) and TestPrune shows improving test quality on an already-gated pipeline adds +2.4 to +12.9pts (arxiv.org/html/2510.18270v2). The gate interface must make adding/refining checks cheap. Also: prefer project-declared tests over model-generated ones — the literature's verifier-gaming failure mode applies to self-generated checks.
anandgupta42 commentedon Nov 5, 2026
Closing: after evaluating verifier-gated and cascade routing, we decided not to build an inferred per-request router. The direction is a strong default model with explicit, user-chosen modes. Reopen if routing is revisited.