Quality Gates

Where this fits

Quality Gates are the after-draft surface in the conversation quality lifecycle. BMO generates a candidate answer (or plan, recipe, or proposal) first; an optional Quality Judge evaluates that draft against evidence and a rubric before the run is finalized.

Changes live output?

Advisory by default. Armed send_back can retry at most once. Quality Gates are not the prompt enhancer: the enhancer adds context before generation; Quality Gates evaluate output after a draft exists.

Maturity: Advanced operator and maintainer surface. Use when you need a bounded post-draft or proposal-readiness judgment, not as a replacement for ordinary review or tests.

Reach for it when

Failures are visible only after a draft exists: unsupported claims, weak plan criteria, missing prompt/recipe hardening, or proposal evidence that is not ready for human review.

Inspect

Job	Route	What it shows
Posture (CLI)	`bmo config show-quality-gates` [`--format=json`]	Armed surfaces, judge model status, metadata-only recent decision ring
Posture (TUI)	`/quality-gates` (alias `/quality_gates`)	Same shared renderer as CLI
Agent / MCP	`quality_gates_posture`, `bmo_quality_gates_posture`	Bounded JSON for automation
Decision trail	`/debugger` → `quality_judge_*` events	Verdict, action, reason, rubric failures, evidence refs
Run ledger	`GET /v1/agent-runs/{run_id}/events`, `list_agent_runs`, `get_agent_run_events`	Same events outside the TUI
Manual rubric	`bmo quality-gate check --path …` [`--surface plan\|prompt_recipe`]	Plan or recipe hardening without enabling runtime gates

Trust boundary: posture and recent-decision readouts carry hashed session/run/proposal/finding prefixes only — never prompts, answers, diffs, plan bodies, recipe bodies, or provider secrets.

Live BMO TUI: the /quality-gates posture readout reports the foundation state, the resolved judge, the per-surface arm state (answer, plan, prompt_recipe, patch_proposal), and the bounded metadata-only decision ring — the surface to inspect before trusting or debugging a verdict.

After a real turn with Quality Gates armed, the Agent Debugger shows `quality_judge_*` events — verdict, action, and rubric metadata alongside the ordinary run timeline.

Conversation quality — lifecycle map; use step 3 of the infra regression walkthrough
Prompt Stack — jump here when the surprise is input assembly, not draft quality
Shadow Evals — jump here for bmo eval compare + eval posture after prompt changes; complements prompt_recipe manual checks
Quality orchestration — separate server fanout path; does not replace post-draft gates
Independent verification — a Quality Gate pass, warn, send_back, or escalate does not ingest mutation aftercare, accept a Patch Proposal, or record a parent outcome. Aftercare and parent evidence stay on that page and on Change synthesis.

Deep dive

What it does

When enabled, the answer gate can return one of four verdicts:

Verdict	Meaning
`pass`	The draft meets the rubric.
`warn`	The draft has an issue, but can still be shown.
`send_back`	The draft should be revised before final output.
`escalate`	The issue needs operator attention.

The default posture is conservative. The feature is off unless configured, and advisory decisions do not block chat output. Armed send-back retries at most once.

Unavailable judge paths are shown as unavailable, not pass. Runtime policy may fail open for answer and patch-proposal infrastructure failures, but the posture and recent-decision ring keep that state visible.

Manual document checks

bmo quality-gate check --path docs/plans/example.md
bmo quality-gate check --surface prompt_recipe --path .bmo/recipes/example.yaml

The plan rubric checks acceptance criteria, source grounding, scope boundaries, system-wide impact, validation commands, and open questions.

The prompt/recipe rubric checks output contract, source ownership, tool availability assumptions, artifact destinations, instruction hardening, and a validation or review gate. This command is manual and opt-in; it does not install hooks or block ordinary edits.

prompt_recipe checks complement — but do not replace — deterministic bmo eval run scenarios documented on Shadow Evals.

Runtime status

The TUI stays quiet for normal pass/advisory outcomes. Runtime activity is shown only for judging in progress, armed send-back, or escalation.

Agent Debugger — inspect Quality Gate events.
Patch Proposals — code-change proposals can use Quality Gate evidence as a review-readiness signal.
TUI - phases vs runtime activity — how the compact status surface works.
Configuration — generated config reference.

Implementation details: quality-gates.md.