Corroborate by independent reruns

SkillDev tools

Use when a candidate patch or answer needs independent corroboration before it is trusted. Run N isolated sandboxed attempts and return the strict-majority byte-identical patch with its agreement count, or report no consensus. Not for multi-stance investigation, use council.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Corroborate by independent reruns skill

What this skill tells your AI

The instructions your AI receives, as published by outlinedriven/outline-driven-development in .devin/skills/corroborate-by-independent-reruns/SKILL.md and read by ahel’s review.

Contract

FieldBound contract
TriggerA candidate patch or answer needs independent corroboration before it is trusted
AuthorityReversible local: writes only the named corroboration report; rollback is discarding it if rejected. No remote mutation. Each run mutates only its own isolated sandbox, discarded after capture.
Side effectRuns N isolated sandboxed attempts and returns the majority patch plus its agreement count; byte identity is a corroboration signal, not a correctness oracle; no-change runs never carry consensus
DoneEither a strict majority of launched runs produced byte-identical output and it is returned with its count, or no consensus is reported

Inputs

  • The candidate patch or answer to corroborate.
  • N, the number of isolated runs to launch (N >= 1).
  • The sandbox execution command or task prompt each run performs independently from a clean baseline.
  • Optional: a baseline output to classify no-change runs against.

Procedure

  1. Receive the candidate, N, and the sandbox execution command. Bound scope to N isolated runs; do not widen to additional runs or mutate the shared working tree. Done when: scope is bounded to exactly N isolated runs with no shared-tree mutation.
  2. Launch N isolated sandboxed attempts, each starting from the same clean baseline and running the same task without access to any other run's output. Done when: N isolated runs are launched, each from a clean baseline with no cross-run access.
  3. Capture each run's full output bytes. Done when: every launched run's full output bytes are captured.
  4. Classify each output: a no-change run whose output is byte-identical to the baseline or empty is excluded from consensus and never carries consensus. Done when: every output is classified as candidate or no-change, with no-change runs excluded from consensus.
  5. Group the remaining outputs by byte identity. If a single byte-identical group is a strict majority of the N launched runs (count > N/2), return that output with its agreement count. Done when: outputs are grouped by byte identity and either a strict-majority group is identified or no majority exists.
  6. If no byte-identical group reaches a strict majority, report no consensus. Done when: no-consensus is reported with the per-run output groups, or the strict-majority output is returned with its count.
  7. Treat byte identity as a corroboration signal only, never as a correctness oracle; do not assert the majority output is correct. Done when: the result is presented as a corroboration signal with no correctness claim.

Failure and recovery

  • Sandbox failure: a run that errors or cannot capture output counts as a launched run that produced no candidate; it contributes to no byte-identical group.
  • Partial results: never return a partial majority; consensus requires a strict majority of all N launched runs, not a plurality.
  • Non-mutation: each run operates in an isolated sandbox discarded after capture; the only named local artifact is the returned corroboration report, which the caller discards if rejected.
  • Blocked or non-converged: if no byte-identical group is a strict majority, the terminal result is no consensus with the per-run output groups. Do not fabricate agreement, promote a plurality, or claim the majority output is correct.

Output

Either the strict-majority byte-identical patch or answer with its agreement count, or a no-consensus classification listing the per-run output groups. Never a correctness claim.

Signals

GitHub stars
54
Forks
10
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
corroborate-by-independent-reruns
Source
github.com/outlinedriven/outline-driven-development