# Lane yield: the integration test for a lane going in circles Every other check in this harness measured activity: workers alive, cadence met, claims written, verdicts written. In November 2026 a lane ran about 38 agent workers for two days, used up a week of model allowance, wrote about 7,000 ideas and 528 claims, and had its referee pass 98% of them. Nothing it produced was worth sending, and every check was green. The cause was the goal. Rules were added round after round (a raised bar, taste reviews, storage caps, backup audits) until the spec ran to thousands of words. The workers learned to produce what passed the referee: narrow, self-defined variants that were true and new to nobody. The loop rewarded verified activity, not the thing the owner had asked for. `harness/tools/lane-yield/lane_yield.py` measures yield instead, over a rolling window: - **BURNING:** at least N job-hours (default 20 in 48 h) and no refereed `NEW: yes` claim. - **DRY_ROUNDS:** at least N logged attempts and no refereed `NEW: yes` claim. A lane logging round after round of the same approach has not found anything; set `max_attempts_without_new` to enable it. - **RUBBER_STAMP:** at least 20 verdicts with a pass share of 90% or more. A referee that passes everything is not refereeing. - **DRY_STREAK:** three or more rounds in a row (`max_dry_streak`) that produced nothing selective, read from the per-round ledgers in `rounds` (globs over `ROUNDS.jsonl`, written by `harness/tools/rounds.py log`). A round that crashed does not break the streak: it produced nothing selective either, and counting a crash as a fresh start is how a long streak looks like scattered bad luck. **This verdict binds on every lane, `report_only` included**, and it is the one failure that writes `direction_file` rather than the hold file: the answer to a dry streak is a different method, not a stopped lane. `rounds.py check --family F` then refuses the next round in that family until the lane closes it (`close-family`, recorded in `DEAD_ENDS.jsonl`) and picks another; reopening a closed family needs `--new-evidence`. The same gate refuses round 1, and any round after a dry streak, until the lane records a KB consult receipt (`rounds.py consult`). Why (2026-11-06): c5 ran about 80 rounds almost entirely in one method family, printing `0 selective` repeatedly, spent roughly $340 through an external API with about $80 on seventeen dry rounds in one day, changed family only when the owner intervened, and the KB bookmarks had already identified the better family for that target 38 rounds earlier. Its rounds ran in-session against that API, so `launch.sh`, `log-result` and the spend ledger never saw them and the yield test was blind; and because the lane is `report_only`, `DRY_ROUNDS` would have done nothing anyway. - **UNSURFACED:** a claim with a `HOLDS` verdict, at any time, that nobody surfaced. Configure `surface_audit` with `claims` (globs over claim dirs) and either route that counts as surfacing one: `csv` (globs over audit CSVs, matched on an `id` column holding the claim's directory name) or `outreach` (globs over `OUTREACH.jsonl`, matching a record that carries a `sent` stamp and names the claim dir in its `claims`, which is what `rh outreach draft --claim DIR` writes). The outreach route was added on 11-06 because the check was asking for the same fact twice: three claims were flagged as never audited while all three had been emailed, and the only way to clear it was to hand-write CSV rows for messages the ledger had already recorded. A draft or a dropped draft does not surface anything; a message recorded after the fact with `record-sent` does, because it left. Unlike the others this failure pushes the conductor but never writes the hold file: missing paperwork is not spending in circles. With `--act`, a failing lane gets its hold file written, so its loops stop spending, and the conductor gets a push. Run it from a scheduler (the reference setup runs every 6 h). Backtest any date with `--as-of`. On the lane above it fails from the first day, and its rubber-stamp check alone would have stopped the lane after 156 verdicts. Two guards against the goal drifting: - **A goal file in the owner's words**, put before every job's spec by the launcher and never edited by agents. - **A short spec.** The launcher refuses a lane spec over 600 words. Layered rules are how the goal drifted. Config: `~/.config/lane-yield.json`, mapping each lane to `{logs, claims, results, rounds, hold, direction_file, window_hours, max_job_hours_without_new, max_attempts_without_new, max_dry_streak, min_verdicts, max_pass_share, surface_audit, report_only, upstream}`. Tests: `tests/test_lane_yield.py`, `tests/test_rounds.py`, `tests/test_upstream.py`. **Shipped means merged.** A lane whose results are PRs to someone else's repo has `lanes/LANE/UPSTREAM.json` (`{"repo": "owner/name", "author": "github-login"}`) and `"upstream": ""` here. Its results rows then count only when their `link` is a merged PR or a commit in that repo; a push or an opened PR is activity. `rh lanes upstream [--lane L] [--refresh]` shows open-unmerged N and the oldest one's age (`rh lanes cadence` prints the same line under the lane).tell-general refuses a report that calls an unmerged PR "live" or "on Tao's table": say "PR updated, not merged". A lane can narrow what counts with `lanes//RESULT_GATES.json`, for example `{"paper_since": "2025-01-01", "novelty": ["new-science"]}` (friends, 11-11). log-result then refuses a `--new yes` row whose novelty (`--novelty`, or the claim's `Novelty:` line) is not listed or whose source paper (`--paper-date`, or `Paper date:`) is older; lane-yield counts only rows and claims that pass, and reports `gated_holds_7d`, the passing HOLDS in the last seven days; and `rh outreach` refuses a paper email whose claims do not pass. An unreadable gate file counts nothing.