orb(gate): next guardrail-loosening tranche — backtest-cleared paths only

[JSONbored]

Problem

Reducing the manual-review surface toward the 97% automation target must not outpace the evidence. The auto-loosening machinery (visible + held-out split clearance, hard floors, flag gates) exists; this issue is the next planned tranche run through it.

Scope

• Enumerate the current manual-review trigger paths with volumes (from the ledger) — which categories contribute most manual load
• For the top candidates, run the standard backtest clearance; loosen only what clears both splits; record corpus checksums + SHAs per run
• Explicitly out: visual-change PRs remain held for human review

Acceptance

Each loosened path has a linked backtest record; automation-rate series (sibling issue) reflects the change; no loosening without clearance.

[linear-code]

JSO-6788

[JSONbored]

Scope item 1 (enumerate the manual-review trigger paths with volumes from the ledger) — measured on edge-nl-01, all-time agent.action.hold:

holds	category
10,042	protected author (maintainer exemption)
5,222	no auto-action planned
1,467	CI still pending
983	unstable mergeable_state
552	CI failing
141	merge withheld, mergeable_state unstable
94	merge withheld, mergeable_state blocked
55	base conflict
46	close-audit adjudication
15	gate blocker ai_consensus_defect
9	precision circuit breaker

The maintainer-exemption bucket dominates at 2× the next category, and it is still accruing (675 holds on 07-29, 122 on 07-30). Broken down by what the gate actually wanted to do:

holds	blocker behind the withheld close
4,741	(no named blocker)
4,074	missing_linked_issue
214	ai_consensus_defect
200	manifest_missing_tests
42	secret_leak
18	linked_issue_scope_mismatch

The top candidate, with two independent signals

missing_linked_issue on protected authors is 4,074 holds that can never produce a close — the author is exempt by design, so the evaluation runs, the hold is recorded, the label goes on, and the outcome was never in doubt. That is not enforcement, it is pure volume, and it is the largest single named contributor to the manual-review surface.

Separately, the deployment’s own anomaly detector already flagged the same rule on JSONbored/awesome-claude:

gate false-positive spike: missing_linked_issue blocked 46 PR(s), 16 merged anyway (35% false-positive, 0 overridden) — the gate is holding mergeable PRs. Keep it advisory / loosen it.

Across the whole ledger the ratio is worse: of 237 PRs this rule blocked, 155 merged anyway — 65.4%.

Caveat, stated because it bounds the claim: "blocked then merged" is not automatically a false positive — some of those merged after the author added a linked issue, which is the gate working as intended. That ambiguity does not apply to the maintainer subset, where the close is structurally impossible, so the 4,074 figure is the unambiguous one and the 65.4% is the upper bound on the contributor side.

Suggested first tranche

Exempt protected authors from missing_linked_issue evaluation entirely rather than evaluating-then-withholding. It removes ~4,074 holds and the manual-review labels they carry, changes no contributor-facing outcome, and needs no backtest clearance because it loosens nothing that could ever have fired — this is removing a no-op, not relaxing a gate.

The contributor-side loosening (the 65.4%) is the one that genuinely needs the standard split clearance, and should be a separate tranche once the false-positive share is separated from the "author fixed it" share. Visual-change PRs stay held per the issue’s own out-of-scope note.

[JSONbored]

Scope bullet 1: the manual-review surface, enumerated from the Orb's ledger

Measured against decision_records on edge-nl-01 (the deployment that actually performs reviews). 2,429 verdicts total.

Action	Verdicts	Share
hold	1,679	69%
close	380	16%
merge	370	15%

So the manual-review surface is 69% of all verdicts. Broken down by the clause that decided it:

reason_code	Holds	Distinct PRs	Read as
success	518	146	Gate concluded success, held anyway
missing_linked_issue	482	96	Policy hold
guardrail_hold	317	96	Guarded-path hold
ai_consensus_defect	164	23	Finding-driven
ai_review_inconclusive	152	39	Finding-driven
secret_leak	17	5	Correct, keep
linked_issue_scope_mismatch	14	4	Finding-driven
content_lane_deliverable_missing	9	2	Policy hold
neutral	6	1	—

By repo, for the top three buckets (window 2026-07-26 → 07-31):

Bucket	loopover	metagraphed	awesome-claude
success	377	52	89
missing_linked_issue	365	62	55
guardrail_hold	86	231	—

The headline: success is the largest bucket

Per DecisionRecord.reasonCode's own contract, the code is "a blocker class, policy_close:<kind>, or the gate conclusion". So a hold carrying success means the gate reached a success conclusion and the PR was held anyway — 518 holds over 146 distinct pull requests, 31% of all manual load. Nothing in the finding set is holding these; something downstream of the gate is.

That makes it the top loosening candidate by volume and the one where loosening is least likely to change any outcome, since the gate already concluded success on every one of them.

Before it clears, the cause has to be identified rather than assumed. success is a bucket, not a mechanism, and at least three distinct paths can land in it — guarded-path holds, the maintainer lane, and the machine-paced review-cadence heuristic. Worth noting that LOOPOVER_REVIEW_REPUTATION is currently off on this box precisely because that heuristic was holding maintainer PRs, so some share of these 518 may be an artifact of a known incident rather than a standing policy. Splitting success by its actual holding path is the next step, and it is a prerequisite for backtesting it — a clearance run against a bucket that conflates three mechanisms would not mean anything.

Not done, deliberately

No guardrail has been loosened. The acceptance bar is per-path backtest clearance on both splits with recorded corpus checksums, and this comment is only the first scope bullet — the volumes needed to choose candidates. secret_leak (17) and the visual-change path stay out of scope regardless.

Ranked candidates for the clearance run

1. success holds, once split by holding path — largest volume, gate already says success.
2. guardrail_hold — 317, heavily concentrated in metagraphed (231), which suggests one or two guarded paths dominate and could be enumerated narrowly.
3. missing_linked_issue — 482, but this is a deliberate policy hold rather than a calibration threshold, so loosening it is a policy decision, not a backtest one. Listing it for completeness, not proposing it.