## Rules every lane gets **HARD RULE (Ricardo, 11-04): the goal must not turn into process, and success is never activity.** Rounds launched, jobs run, designs scored, tests passed, checks added, ideas logged and dollars spent are not results. The only result is the thing that was asked for, shipped. A lane that measures itself by how much it ran has already lost the ask. **Restate the ask.** The first line of each round, job or campaign note is `You asked: . I will: .` A run whose difference from the previous one cannot be named is a repeat and must not start. **Stop rule.** Three runs that fail the same way end that approach; there is no fourth. Record in STATUS.md or STATE.md what failed, by number, and what changes. "Same way" means the same thing stopping the work, whatever the parameters were. On the first real failure after one of your own changes, revert it before investigating. **Log every run, win or not.** One line per run to this lane's `RESULTS.jsonl`: `log-result --ledger /RESULTS.jsonl --new yes|no --kind --what "" --link --referee HOLDS|NONE` A passing verdict (HOLDS, ACCEPT, PASS) on a `--new yes` row also needs `--verdict ""` answering three questions, each with the file, command, count or line you checked: `CONTROLS:` did an assumption override what the controls show? `CONTRADICTING:` was data that contradicts the conclusion dropped or ignored? `LABELS:` was a label or annotation trusted without checking it against the raw data? An unsupported "no" is refused, because it is the cheapest way to satisfy a rubric. Rows with `--referee NONE` or a failing verdict are unaffected. (BioBasicsBench failure modes, 11-06: these are what goes wrong, not arithmetic, and our own referee passed 98% of 528 claims.) `--new yes` only when something actually moved: a printed number beaten under the same definition and data, a stated open problem settled, a prediction tested on new data, or a shipped artefact that is better than the one before it. Reproductions, restricted variants you defined yourself and sensitivity checks are not new. Say "no" plainly when it is no. **Email in two languages (Ricardo, 11-05).** When a recipient's contact record carries a language other than English, that language is the **default** for the whole email, and it **opens** and **closes** in it. **English is required only where the content is technical** — numbers, variables, equations, methods, measurements — because English is the science language. Attachments are always English. The rule holds even when the recipient writes their own science in English. `rh outreach lint` fails a technical paragraph in their language, an opening or closing that is not in it, and an attachment that is not English; ordinary prose in English anywhere in the email is a warning, because it belongs in their language too. **No ops in friend emails (Ricardo, 11-06).** An email to a friend carries only what the friend needs. Never our infrastructure or operations: providers (RunPod, OpenRouter), pods, security incidents, key rotations, harness or watchdog details, or why something was re-run. Say "re-checked on my own machine; the numbers match." If the friend asks about one of these, answer in one plain line. The rule came from one sentence in an email of 11-05 — "The first two ran on a RunPod server that has since had a security incident" — which cost a round trip and told a friend about a host compromise he had no way to act on. `rh outreach lint` fails any non-public message carrying those terms (`OPS_DETAIL`); when the friend raised the subject themselves, name the term with `rh outreach draft --asked-about TERM` and it drops to a warning, so a reply can answer them. It takes a term from the check's own list, never the friend's words. **A result is not sendable until someone else rebuilt it (Ricardo, 11-05).** A referee pass is not evidence: 528 claims went through a referee that passed 98% of them. A result may go out when a **separate model rebuilt it from the public inputs with its own code and got the same answer**, and that re-check is recorded. Two of thirteen claims did not survive that on 11-05. Evidence **produced or modified on a pod** stops being evidence when the host is compromised, and becomes evidence again only once it has been **re-run after the incident** somewhere trusted, with the re-run recorded. `rh outreach lint` fails an email whose claim has no recorded independent re-check, whose re-run did not reproduce it, whose re-check dropped it, or whose pod-touched evidence was last re-checked before the configured incident: `RECHECK_MISSING`, `RECHECK_FAILED`, `RECHECK_DROPPED`, `POD_EVIDENCE_STALE`. Name the claims a draft reports with `rh outreach draft --claim DIR`. **How this is enforced.** `lane-yield` reads that ledger on a schedule. BURNING fails a lane that spends job-hours with nothing new; DRY_ROUNDS fails a lane that logs run after run with nothing new; RUBBER_STAMP fails a referee that passes everything. A lane that stops logging reads as a lane that stopped reporting, and is treated as such. Process is not the task: do not add gates, receipts, audits or review stages unless the spec asks for them; write the suggestion in STATUS.md and keep doing the task.