Status. Implemented and measured on this box, 2026-08-06, under the best-effort development profile the 9F-A ruling authorized. Everything below carries strict_comparison_eligible: false, enforced by the code rather than promised by this memo.
response_coverage reads one episode. Aggregating a split has two defensible answers, and they disagree:
pooled sum the numerators, sum the denominators, divide
-> every unit of rubric weight counts once
mean average the per-episode ratios
-> every episode counts onceOne heavy fully-answered task and nine light unanswered ones:
pooled weight coverage 50%
per-task mean 10%
pooled signal coverage 10%Either is a correct answer to a different question. A reader shown one of them and told it is "the coverage" has been misled about the other, so both are reported and neither is named `coverage` — a key by that name would be read as the answer and make the others decorative. That is the same refusal coverage.py already makes between weight and signal count.
S1 constructs the divergence rather than hoping for it, and pins the exact pair. S11 pins a case where the two coincide (77.5% both), so S1's gap reads as a property of skew rather than as noise.
Unreadable tasks are counted and excluded. A task with no receipt — the agent would not launch, admission refused the fixture — stays in tasks, is absent from every ratio, is listed by id, and makes the split ineligible. Dropping it silently would inflate coverage by discarding exactly the tasks that went worst; scoring it zero would invent an adjudication that never happened.
`error` episodes stay in the denominator (the 9E ruling, applied). An episode whose verifiers errored has a receipt, a reading and a real denominator; excluding it would let a split raise its coverage by breaking verifiers, which is the shape of every erasure route in this repository. S3 measures it as the difference between two splits identical but for the error episode — "it is included" is a claim about a denominator, and a denominator is what a reader cannot see.
`None` means no denominator, never zero. Reporting zero would be the strongest possible claim from the least possible evidence — the same inference-from-absence the containment slices spent two rounds removing.
S5 refuses to certify a mixture of a certified and a best-effort episode. That is the laundering path: two episodes aggregated into one figure that looks publishable and inherits the stronger label from neither. The profiles are listed so the mixture is visible, the arithmetic is unaffected, and strict_comparison_eligible is false whatever the ratios say.
An unclassified profile is refused too, not assumed fine — the same rule execfacts.strict_comparison_eligible applies one level down.
| verifiers wired | tests / identity | reward | weight coverage | |
|---|---|---|---|---|
| A | all five | fail / fail | 0.55 | 100% |
| B | three | not_applicable / not_applicable | 0.55 | 55% |
reward.score gives a fail and a not_applicable the identical 0.0 (§6a), so both receipts post the same number, the same status: ok, the same validity: valid. Nothing in the reward can tell them apart.
In A the missing 0.45 is a verdict — two verifiers looked and said no. In B it is a gap — nothing was ever asked. That is the entire justification for response_coverage existing, demonstrated rather than asserted.
S10 asserts the rewards are identical first. If they differed, the pair would be showing that a score does distinguish them, which is the opposite claim, and the law would be proving the reverse of its own name.
S11 aggregates the pair: pooled 77.5%, {"identity": 1, "tests": 1} unanswered, and a reward mean that cannot move because the two episodes score the same. That gap is the deliverable of items 11 and 12 together — a split-level number a reward mean cannot produce, over evidence a reward mean cannot see.
split_coverage is a pure function: no I/O, no clock, no engine, no identity. S7 reads that off the parse tree rather than trusting the docstring. It mints no id and is never stored, for the reason response_coverage and ComparisonV1 are not: every input is already hash-bound inside the episodes it came from, so storing the aggregate would move ids to record something those bytes already determine.
S9 — the one law that composes with real sealed bytes rather than hand-built numbers — looked for the episode bundle at a path that does not exist, and then for a manifest member called reward_spec when the manifest calls it reward. Each time it skipped, which is how a law quietly stops testing anything.
Both were mine, and both were visible only because a skip is reported as a skip. The packet gate would have caught them regardless: G6 forbids skips with the engine present, so the packet would have been REJECTED rather than shipping a law that tested nothing. That is the backstop doing exactly what it was added for after the 9D J15 rejection.
test/test_splitcoverage.py 11 passed, 0 skipped, 0 failed
full battery 805 passed, 0 skipped, 0 failed (36 files)
packet traaviis-11-12-pre.zip
d6cb637a7b79591334fd3b0adebd5c76c79508f74934bc296a436f434a2cab8a
G5 636 passed, 169 skipped, 0 failed
G6 805 passed, 0 skipped, 0 failed
ACCEPTEDAgainst 9F-A: G5 625 → 636, G6 794 → 805.
13 v2 cutover certified backend
14 [&]code integration certified backend
9F items 2-7 a host that can isolate the control planeTwo things worth ruling before the cutover, neither urgent:
The v2 capability vector names `memory_max_bytes` and `pids_max` as sealed fields. Those are values a certified backend sets, and this host cannot set them, so they are deliberately not stubbed into the vector. A placeholder null would be indistinguishable from a measured absence, which is the distinction UNKNOWN was added to preserve.
`compare_episodes(strict=…)` still defaults to `False`. Correct today — the entire corpus is best-effort and a True default would refuse all of it — but a caller who forgets the flag gets a comparison of best-effort evidence with no complaint. The flip belongs to the cutover; it is recorded here so it is not discovered as a surprise then.