clay-borg/evidence/CB-EV-0022-collect-the-rulings.md

# CB-EV-0022 — collect the rulings

CB-WP-0026 T05. Tier S (structural S — applies rulings inside an existing
capability; chaos d8=6 → no override). Declaration 9 of chaos window 2.
Closed 2026-08-05.

**Delivered:** ten rulings recorded in `§Underdetermined`, five
`provisional: true` flags lifted, four scenario schema fields, the
`encodes_u_item` declaration, and an empty finding queue.

---

## 1. How long the answers sat, and what noticed them

**Two days**, and the thing that noticed was **the register's first run** —
not a human, not the adversarial review.

`ground-game` ruled all ten U-items on **2026-08-03** (GROUND-WP-0002 T05),
every one confirmed. On the same day CB-RES-0007 reported *"0 of 10
ruled."* Two days later CB-WP-0022 built the register recording them as
`reported`, and the adversarial review — which found four other things —
did not catch it either. `make design` did, on its first execution.

**This is the symmetric failure nobody designed for.** CB-WP-0022 was
shaped end to end around *we send findings and nobody reads them*: the
four-day unread inbox is quoted in the declaration, the survey, the ADR
and the spec. The mirror case — *they answer and we do not collect it* —
appears in none of them.

It is arguably the worse of the two. An unread message is visible as
silence; a collected-but-unapplied ruling looks exactly like work in
progress.

## 2. Did any confirmed default fail to match the kernel? No.

This was the control that mattered. Every ruling was a *confirmation* of
what we told `ground-game` we simulate — so a red scenario would have
meant **we described our own behaviour incorrectly to them**, a defect in
our report rather than in their ruling, and the most serious class
available since it would be about our own code.

```
make sim → 26 passed, 59 rules covered
```

**No scenario went red.** The five confirmed defaults are the behaviour
implemented. That is the strongest single result here and it is a
negative: nothing was wrong.

## 3. The mapping gap was not what the survey said, and I reproduced the defect writing it

CB-RES-0007: *"six of the ten already have provisional scenarios."*

**Measured: one.** And the interesting part is how the other nine were
lost.

I wrote two mappings from the `covers:` lists and **both were wrong**:

| claimed | why it was withdrawn |
|---|---|
| `gr-a04-bond-support` → U1 | asserts consent is **required**; U1 asks **when** the target accepts |
| `gr-d05-darvo-reverse` → U5 | exercises the **unrejected** REVERSE; U5 is the **rejected** one (GROUND—ND) |

Both were plausible from `covers:`. Neither survived reading the
description. **These are the third and fourth instances of this exact
defect** — a link that looks right from metadata, asserted without
checking what the artifact exercises — and the first two reached
`ground-game`.

So the answer to *"do the four unlinked scenarios encode U-item defaults
at all?"* is: **at least two of the four do not**, and the survey's "six of
ten" was not an under-documented truth. It was wrong.

`encodes_u_item` is now a declaration a scenario makes or omits, and
`design.py` asserts the file names what it claims.

### The check's own first version was the same looseness

Written as `grep -lE "\bU<n>\b"`, it went red the moment two scenarios
recorded *why they do not* encode U1 and U5 — reporting `['U1','U2','U5']`.

**A mention is not a claim.** That is precisely the imprecision that let
*"six of the ten have provisional scenarios"* stand unchallenged for five
days: someone grepped for U-item strings and counted hits. The check now
asserts on the declaration.

## 4. The queue reached 0

```
QUEUE (open findings)          (none)
open, lacking a reproduction   0     target 0
closed (log)                  12     [U1..U10, F11, F13]
with a resolving reproduction  3/12 = 25%
notes                          2     F12, F14
```

**`9 → 0`**, the number CB-WP-0026 T04 named. First evidence that
ADR-0012 D5's lifecycle is real rather than drawn: findings entered a
state, moved through it, and left the queue.

**25% reproduced must not be read as a failure.** Nine U-items closed by a
**ruling**, and a ruling is not an artifact. The metric is now honest
about something the survey's 33% concealed: most of our findings close
because someone answered them, not because anything demonstrates them.

**What the lifecycle could not express** — the honest gap: `applied` is
defined as *the source changed and the provisional default was deleted*.
Here the rulings **confirmed** our defaults, so nothing in the rules moved;
what changed is that the flags came off. The state fits, but the
definition had to be read generously. If a future ruling *overturns* a
default, `applied` will mean something materially different from what it
meant today, and D5 does not distinguish them.

## 5. Two things the schema caught

**`ScenarioFile` is `deny_unknown_fields`**, so five scenarios failed to
parse until `ruled`, `ruled_by`, `ruled_note` and `encodes_u_item` were
declared in the Rust struct. A corpus that accepted unknown metadata would
let a typo'd `encodes_u_iem` sit forever claiming nothing — and this
pass's whole subject is claims nobody checks.

**`record.rs` had to set them explicitly.** A recorded game is evidence of
what happened, not a claim about an undetermined rule. Filling the fields
via `..Default::default()` would have been shorter and would let a
recording silently inherit a U-item claim, pointing a reproduction at a
finding it has nothing to do with.

## 5b. Two gates went red for the best possible reason

Lifting the last five provisional flags broke two positive controls, and
**both were broken in the same way**: they asserted against live repo data
instead of constructing the condition they test.

**`rule-coverage.py`** required `bool(prov)` — *at least one provisional
item must exist*. That guard was the right instinct (a control that passes
vacuously is worthless) wired the wrong way. With nothing provisional, it
went red. It now builds a fixture, asserts the missing-owner case is
**caught**, and reports the live count as a diagnostic — because **there is
no number of provisional items this project should have.**

**`design-baseline.py`** pinned *"the measured baseline is 2 of 6"* and
reported `1/6`. Its "six provisional defaults" row **globbed
`provisional: true` at run time**, so the dated snapshot was never a
snapshot — it drifted with the repo. C2 dismantled this tool four hours
earlier and missed this: a hand-maintained dict with one dynamically
computed row is worse than a fully hand-maintained one, because the
recomputed row silently disagrees with the date in the header.

Frozen to the literal list it measured on 2026-08-03, and **removed from
`make self-tests`** — that target is *"a positive control for every
reporting tool"*, and per ADR-0012 D8 this is no longer a reporting tool.
Leaving it wired in meant a superseded instrument could fail the build.

**The pattern across both**: a control that reads the world it is meant to
audit will eventually audit a world that has moved. Neither was caught by
review; both were caught by the world moving.

## 6. The register moved, and ADR-0012 D2 gave way to loadability

`specs/GroundRules.md` crossed the ~400-line limit and `loop-lint`
required a split. ADR-0012 D2 said **"no new file"**, so this is a
deviation and is recorded as one.

**D2's substance holds.** Its argument was *one register, not a second
mechanism competing with the first* — and `specs/FindingRegister.md` is
that same register, moved, still driving off the `provisional`/ruling
machinery, still the only one. What was traded away is the literal "no new
file", which was D2's *implementation*, not its reason.

It also resolves an awkwardness D2 named itself: *"a finding about the
engine's behaviour sits in a document about the game."* Now it does not.

## 7. Chaos window 2

**Declaration 9 of 12.** Structural S, d8 = 6, no override. Recorded per
§Loop tiers even though it changed nothing.

**Window 2 still has no override to evaluate** — nine declarations, zero
8s. The retirement condition (*retire if an override changes nothing twice
running*) cannot be assessed, and at d8 the expected count over twelve
declarations is 1.5, so this is unremarkable rather than evidence of
anything.

## 8. Cost

CB-WP-0022's cost, by re-running the instrument:

```
CB-WP-0022-T05   $ 6.43
CB-WP-0022-T01   $ 4.17
CB-WP-0022-T02   $ 3.56   (the separate reviewer's own spend is in the subagent tree)
CB-WP-0022-T04   $ 1.92
CB-WP-0022-T03   $ 1.01
```

**The chain did not snap this time** — CB-EV-0019 §4 predicted it might,
having found `cb-cost.py --slug CB-WP-0020` aborts for want of retained
transcripts. CB-WP-0022 is recent enough to still be in the window. **The
bound CB-EV-0019 asked for is still owed**; this pass is evidence that the
rule works for a pass one step back, not that it works generally.

## Open after this pass

- **Nine U-items are `applied` with no reproduction.** That is recorded,
  not hidden, but it means nine rules rest on a ruling nobody can re-run.
  Cheap to fix incrementally: each needs one scenario naming its item.
- **`applied` conflates *confirmed* with *overturned*** (§4). It will
  matter the first time a ruling goes against us.
- **The cost-chain bound** (CB-EV-0019 §4) is still unwritten.