ADR-0131 — Minimum useful rigor: no ceremony without leverage

Context

A refinement report ("Ceremony Does Not Improve Agent Output by Itself") argued Suspec should default to less process and more proof. Most of its direction is already shipped doctrine — the who-should-not- use case (ADR-0057, docs/01), disposable run-scoped task briefs, manual/cost-tiered heavy review (ADR-0122, ADR-0124), and skills paired with their counterweight ([[SWESKILLS]] with [[SKILLSBENCH26]]). What was missing is the doctrine stated once, plainly, with the internal measurement that motivates it recorded honestly rather than as a slogan.

The measurement, in its own words (suspec-works DP-9, the first controlled A/B of the whole loop — local, unpublished, pre-registered). A 15-task battery run through four arms (full loop / plan-then- implement / gate-only / skills-only): every arm hit 100% oracle-pass with zero regressions (A 15/15, B 15/15, D 6/6, E 3/3), at a ceremony cost of ~2.6× for the full loop. The finding's own conclusion: "On tractable, clearly-specified tasks the loop is correctness-neutral overhead — now with data, matching the docs' own guidance. The loop's value for hard, ambiguous, or long-horizon work where implementers do ship defects remains plausible and unproven." The direct gate test could not fire: a capable model on unambiguous briefs shipped no defects, so the gate had nothing to catch — "the framework's own 'who should not use this' guidance (small, clear, tractable work) confirmed with data, not a refutation of the loop for hard work." Cost is model-confounded (two models mid-run); only the correctness parity is robust.

Three guardrails on the way in, so the doctrine records the evidence and not more than the evidence:

1. No "ceremony doesn't improve output" claim. DP-9 showed parity on tractable work and could not test the hard-work regime; the slogan overstates it, and canon already forbids it — ADR-0117 (no count-bearing prose in public/current copy) and ADR-0063 (no data-shaped claim without a level and, for enforcement, a shipped tool).
2. A non-ceremony counterweight exists. Under the same discipline, implementers refused to fabricate evidence when asked (suspec-works #90) — a benefit of the honesty framing that is not process bulk. Cutting rigor to zero is not the lesson; cutting unleveraged rigor is.
3. Targeted structure still measurably helps. A lean plan the coder then follows buys a real gain ([[SELFPLAN]] +25.4% Pass@1; [[PLANCOMPLY]] plan + reminders improve compliance), and reducing detail can also help by not misleading retrieval ([[UNDERSPEC-HELPS]]). The variable is leverage and task- fit, not artifact count — the same split [[SWESKILLS]]/[[SKILLSBENCH26]] found for skills.

Decision

Default to the least rigor that still leaves enough proof; add structure only where it changes execution quality or reviewability. Two operational tools carry the doctrine, both introduced by this ADR:

1. The rigor ladder (docs/reference/rigor-escalation.md): named L0–L5 rungs (prompt-only → lean spec → spec + task → + independent review → Revolver panel → orchestrated waves), default L1 or L3, the top rung is never the default, keyed to the existing risk axis — size / diffusion / churn / change-type (ADR-0094). Level: convention.
2. The artifact leverage test (ADR-0109, extended this pass): every artifact, section, and template earns its place by improving at least one of {clarity, scope, execution-context, verification, reviewability, durable-memory}; if it has no consumer, cut it. Level: checklist (a spec-check / review item, not a shipped enforcer).

One convention line, reinforcing existing skip-clauses (not a new field). Escalating to a heavy mode — a Revolver panel, an orchestrated fan-out, a durable task brief on otherwise 1:1 work — SHOULD state its reason (the risk that warrants the cost). This is prose, resolved by review; it adds no machine- readable complexity_justification field. Level: convention.

What stays fixed. The independent-review invariant (ADR-0119) and the evidence gate (ADR-0121) are not "ceremony" to trim — they are the minimum, not the surplus. Minimum useful rigor lowers the default rung, never the floor that keeps a completion claim honest.

Consequences

• A single stated doctrine the ladder and the leverage test both point at, consistent with the vocab tiers, the output-economy floor, and the risk-weighted-review rule already in canon.
• Not enforced. No hook, no checks.yaml rule, no contract change (ADR-0063). The ladder is a how-to reference; the leverage test is a review checklist item; the escalation-reason line is a convention a reconcile pass MAY note but does not gate.
• The DP-9 record stays private and descriptive (single operator, N=15, model-confounded cost); this ADR cites it as recorded rationale, not as a published result. The burden of proof for the loop's hard-work value moves to a harder task battery, not more easy-task replications — a future suspec-experiments run, out of scope here.

Affected obligations / constraints

• Refines: ADR-0057 (practical-first positioning) and ADR-0094 (risk-weighted scrutiny — adds the named ladder). Extends: ADR-0109 (the leverage test generalizes output-economy from agent output to all artifacts).
• Honors: ADR-0117, ADR-0063 (no slogan, no unlevelled claim), ADR-0119, ADR-0121 (the floor is not trimmed).
• Grounded by: [[SELFPLAN]], [[SWESKILLS]], [[SKILLSBENCH26]], [[PLANCOMPLY]], [[UNDERSPEC-HELPS]], [[BUILDAGENTS]]; the [[INSTRROT]] boundary (stale artifacts are prevalent but their output-quality effect is unmeasured — the leverage test cuts by consumer, not by an unproven decay claim).
• Does NOT change: the artifact formats, the verdict model, or the checks contract; introduces no enforcement and no provider-specific mechanism.

Status

Accepted (2026-07-04).