Skip to content

0005: The post-v0 batch is counted, and the budget rises by each piece's measured size

Status: Accepted (owner-ratified 2026-10-18, capped at 3,785; see Ratification) · Oct 17, 2026 · amended by each PR of the batch below (last: #27, which closes the batch)

Context

ADR 0004 left the v0 budget at 2,800 counted lines with 17 to spare (2,783). The next unblocked issues all add new flightdeck code, not ported code:

  • 17: flightdeck task import-aidlc, which turns ai-dlc scenarios into task directories and specs. ai-dlc's

    contracts and suite checks go to ported/aidlc/ uncounted (ADR 0004 rule 1), but nothing in ai-dlc wrote Harbor tasks or flightdeck specs, so the converter is new code.

  • 13 (holdout custody), #16 (mutation scoring), #21 (knowledge gold-set scoring): v0-plan pieces, scheduled after v0.

  • 24 (trial-result cache and drift sentinel) and #27 (rubric judges, PPI): P4 frontier pieces, each gated.

ADR 0004 says the next piece that doesn't fit is a signal to cut. Each PR of this batch looks for cuts first and records them here. The cuts so far:

  • stats/__init__.py re-exported stats.core (14 lines of duplication); callers import stats.core (#17).
  • applies_when matched source and repo with two special cases; one glob rule over the runs columns replaces both and covers task_id and variant too, for a few lines less (#17).

  • 13 found no cut in what it touches; it moved task digest --write's logic into tasks/format.py so sealing

    reuses it instead of adding a third copy.

  • 16 reuses stats compare, pass_hat_k_suite, wilson, the spec SQL compiler and experiment run's cost gate

    rather than adding its own; no cut found in specs/.

  • 21 reuses the custody manifest, canaries and canary-scan for the gold set instead of a second sealing scheme;

    no cut found.

  • 24 lets load_trials take a run set, so holdout score and mutate score drop their own loops over experiment

    ids (two lines); no larger cut found.

  • 27 reuses cohen_kappa from #21, the model-concreteness rule and the custody-style digests; no cut found.

The batch ends at 3,785 counted lines, under rule 3's 3,800. The next piece (#25, #26, #28-#32) starts its own ADR and should begin with #23's cleanup and the real-client question.

Cuts considered and not taken:

  • The local runner's real-client path (claude-code@<version>: stream parsing, credentials, run budget; about 55 lines in runners/local.py). It is unqualified, but ADR 0004 hands that call to #23 once ai-dlc#223 merges; cutting it here would reverse #50 hours after it merged.
  • Code in ported/: shrinking it changes nothing counted, and moving flightdeck code there is the dodge CONTRIBUTING forbids.
  • Compressing reviewed code to fit: ADR 0003's option B, already rejected.

Options

Option What it is For Against
A. One ADR per PR Each issue argues its own raise Smallest steps Six ADRs saying the same thing; reviewers lose the total
B. One ADR for the batch, amended per PR Each PR adds its measured size to the table below and moves the limit by exactly that One place shows the total and every cut; no headroom is granted in advance The ceiling moves several times in a day
C. A separate budget for frontier (P4) code v0 stays at 2,800; #24 and #27 count elsewhere Keeps the v0 number A second number is still a raise; the boundary between v0 and frontier code is soft

Decision

B.

  1. Every PR of the batch (#17, #13, #16, #21, #24, #27) first looks for cuts and records them above, then measures its net counted lines with scripts/loc_budget.py and adds a row to the table. LIMIT in scripts/loc_budget.py moves to the new total, with no headroom.
  2. Frontier code counts like v0 code. A UI, a server or a database stays out regardless of room (ADR 0003).
  3. If the batch's total passes 3,800 counted lines, stop and revisit: either cut a piece or split the package.
  4. Any piece outside this batch needs its own ADR.
PR Issue What it adds Net counted lines Limit after
(ADR 0004) #18 the local runner's adapter 2,800
#51 #17 tasks/aidlc.py and task import-aidlc (+131); the any path operator and task_id/variant scope in specs, after the applies_when cut (-4); the stats re-export cut (-14) +113 2,896
#52 #13 tasks/custody.py: manifest, seal, canary scan, score once, briefs, and the holdout CLI (+204); write_digest moved to tasks/format.py so the CLI, the importer and sealing share it (+1); registration (+2) +207 3,103
#53 #16 specs/mutate.py: operators, plan, run through the cost gate, kill scoring and the survivors' report (+167); stats compare no longer fails when no trial has a known cost (+1); registration (+2) +170 3,273
#54 #21 tasks/knowledge.py: seal, citation classes, correctness by raters, the paired report and the 15-point rule (+130); cohen_kappa in stats/core.py, which #27's calibration reuses (+10); custody.save creates its directory (+1); registration (+2) +143 3,416
#55 #24 runners/cache.py: key, lookup, plan reduction, reuse rows, cache drop and the drift sentinel (+148); --cache in experiment run and digest stamping in both backends (+21); load_trials reads run sets and reused trials once each, and stats compare reports them (+12); schema and registration (+3) +184 3,600
#56 #27 judges/: rubric versions and digests, label sheets, calibration with weighted kappa and its interval, judge-label storage and the PPI estimate line (+141); stats/ppi.py, PPI++ and stratified (+36); the decision.yaml rubric gate (+5); schema and registration (+3) +185 3,785

Consequences

  • The total and its history live in one table, so the cost of each post-v0 piece stays visible next to its cuts.
  • ported/ keeps growing uncounted (#17 adds ai-dlc's contracts and suite checks); loc_budget prints it beside the total.

  • 23 should re-check the real-client path and the importer once ai-dlc#195 removes ai-dlc's copy.

Ratification (2026-10-18)

Ratified with a cap. The batch is closed at 3,785 counted lines, and rule 3's 3,800 is withdrawn: LIMIT stays 3,785.

  • Until #20's report is posted and its ADR is merged, no ADR that raises the limit is accepted.
  • A counted line added before then is paid for by a counted line cut, starting with #23's cleanup.
  • New feature work (#25, #26, #28-#32) waits for #20.
  • The batch's pieces run on real data before anything is added to them.