The metric the operator named is legitimate research output per unit of scarce human judgment. That needs a denominator, and the denominator is only useful if it is not gamed. This file records every point where the run believed it needed the operator, whether the need was real, and what was substituted.
The run was instructed not to optimise for zero interventions. Two were genuinely deferred; the rest were substituted, and the substitutions are named so they can be scored later.
I-1 — Which research programs to fund (SUBSTITUTED)
Decision required: the operator explicitly declined to assign a direction. Why the system could not resolve it from the repository: ROADMAP.md states a next build, but the mandate was that the current agenda might be wrong. Substitution attempted: five competing programs were constructed with theses, objections, cheap falsification tests and a scoring rule fixed before any result returned (PROGRAMS.md), so the allocation is auditable after the fact instead of being a matter of taste. Outcome: the scoring rule fired as designed — the candidate-generation programme (P4) failed its own pre-registered test (fewer than three survivors of a five-minute attack) and did not get renewed; the audit programme (P1) returned six defects and was extended. Capability that would remove this intervention: none needed. This is the kind of judgment the system can make provided it commits to the criteria first.
I-2 — Whether to spend the run's wall clock installing a Lean toolchain (SUBSTITUTED)
Decision required: several gigabytes and an unknown fraction of the run, for a capability the environment did not have. Why it looked like a human call: the operator warned specifically against building infrastructure because infrastructure is satisfying to build. Substitution: it was made reversible and cheap to abandon — the install ran in the background against a fixed falsification test ("does the repository's own zero-sorry claim reproduce?"), so its cost was wall clock the run was spending anyway and its value was a replication rather than a capability. Outcome: replicated (8715 jobs, zero sorrys). Justified.
I-3 — Whether to modify zeta/ at all (SUBSTITUTED, with a rule)
Decision required: hunts/README.md forbids a hunt from modifying zeta/ without explicit permission, and the operator's mandate explicitly allowed modifying the lab. Substitution: a rule fixed in advance — a change to zeta/ is allowed only where a defect was reproduced by the director independently of the agent that reported it, and only with a regression test or an explicitly corrected contract. Six changes met it. Nothing was changed because it looked better.
I-4 — Whether hunts/local_positivity/ could be repaired by this run (SUBSTITUTED)
Its MISSION.md scopes that hunt to itself. The run's justification for editing it is that provenance_report() was publishing a false statement about its own module's theorem in machine-readable output; leaving a known-false field in returned data because of a scope rule would be the rule defeating its purpose. Recorded here so the exception is visible rather than silent.
I-5 — Adjudicating the shift result (SUBSTITUTED — the load-bearing one)
Decision required: an explorer produced a five-part refutation of a repository claim, some of which was right and some of which was not. Adjudicating that is exactly the "severity calibration" and "skepticism routing" that meta/operator-functions.md identifies as the operator's residual role. Substitution: two further agents, each blind to the other, one paid to destroy the result and one paid to find prior art. They converged: the numbers survived, the inferences did not, the mathematics turned out to be known, and the skeptic found a defect the original had missed. What this measures: the substitution worked because the parties were independent and had opposed incentives, which is the same condition meta/ledger.py records as the guard on self-adjudication. It is not evidence that a single agent could have done it. Residual human need: the decision to accept the skeptic over the explorer was made by the director, which is a party to the dispute. That is the one structural weakness in this run's adjudication, and it is recorded rather than resolved.
Genuinely deferred to the operator
D-1 — Nothing here should be published outside this repository. The run produced no new mathematics (the knownness gate says so explicitly, at ~0.9 confidence). No external review is warranted and none is requested.
D-2 — The rung-3 re-plan is a build decision, not a measurement. The run measured that a feasible configuration exists and priced it at ~130k terms against 79.5k. Whether to spend that compute — and whether to instead do the architectural fix (a polar or mean-value enclosure of m^{-s}, worth ~2× on the whole certificate) — is a resource call with no reversible experiment that substitutes for it. It is the one place this run stops and asks.
Honest accounting
| quantity | value |
|---|---|
| interventions believed necessary | 7 |
| substituted by an independent agent or a pre-committed rule | 5 |
| genuinely deferred | 2 |
| interventions the substitution demonstrably failed | 0 (no substitution was later found to have been wrong — but the run has no independent check on that, which is itself the limit of this table) |
The last row is the one to distrust. A ledger kept by the system whose autonomy it measures cannot certify its own substitutions, and this one does not claim to.