Corrections are part
of the benchmark.

Mini Ledger V5 uses compatible R2 through R5 source revisions. R2 established fixed turns, the explicit wire contract, and source-only verification. R3 updated DotAgents and corrected cache/usage accounting. R4 hardened stream completion, redaction, and retry reliability. R5 adds Droid 0.186.0 through the sealed CLIProxy route as 15 fresh terminal runs.

The campaign does not rewrite those results as if they came from one implementation. Every accepted run names its source revision, challenge and schedule hashes, result key, duration, usage, attempts, and trace. The task, scoring, models, requested high reasoning, and 30-minute per-turn limit remain fixed.

Operator authorized one fourth recovery attempt on 2026-07-31 after two disconnected upstream streams and one sealed turn timeout; task, adapter, model, and timeout remain unchanged. The fourth attempt keeps the same logical job, task, adapter, model, reasoning level, and timeout. No fifth attempt is authorized.

V5-R1 separated agent and verifier containers, but verifier stages still shared candidate state, root-owned fixtures could be unreadable to the candidate, and one concurrency oracle contradicted the public idempotency contract. DotAgents also received the wrong workspace path. These defects could move scores in either direction, so no R1 trial is an official result.

R2 gives every stage and holdout case a source-only workspace, applies candidate ownership to fixtures, states the wire protocol in every turn, distinguishes infrastructure faults from agent failures, and archives each retry as a separate immutable attempt.

The native runner isolated the candidate workspace and harness home, but the agent process could still traverse the parent Git checkout. That checkout contained current and predecessor verifier source. Because the exposure was common to all 60 runs, none qualifies as a sealed result.

We then reviewed executable command and file-path fields across all published semantic traces. Eight show direct access to non-provided verifier source. Six reached holdout source. One opened the current V4 verifier; seven opened only the closely related V3 predecessor. The other 52 are labeled “no verifier access observed”—not “clean.”

Removing the table would obscure both the scale of the incident and its possible effect. The historical ordering remains available for diagnosis, with contaminated generations marked individually. It is not carried into replacement Elo and should not be interpreted as a clean harness comparison.