clay-borg/evidence/CB-EV-0009-inspectable-table.md

CB-EV-0009 — the chaos roll fired, and what it cost

CB-WP-0011 T03. Measured 2026-08-02 at HEAD after T02. Pass kind product.

This is the first pass in which the chaos mechanism changed a tier, so the first question is the one the calibration window exists to answer.

---

1. Did tier S produce a worse outcome than tier L would have?

No — and the reason is not "tier L was unnecessary". It is that the roll forced a decomposition that was better than the one I had planned.

The declaration:

structural tier  L   (INTENT stage 1 creates a new capability port)
chaos            d4 = 4  → override fires
override roll    shuf -e S M L → S

Before the roll I had the shape of the tier-L pass in hand: a survey (CB-RES-0006) whose leading constraint would have been AM-4a's 3,750 lines of headroom — 246,250 of a 250,000 target, measured — followed by an ADR choosing a 2D toolkit and standing up cb-render-api / cb-render-null.

Tier S has no survey and no ADR, and the hard gate — no implementation code for a capability exists before its ADR is committed — is not a tier weight and does not roll down. So the two rules met head-on, and the resolution was to split the declaration: take the part of stage 1 that creates no port and adds no dependency, and leave the port to its own declaration and its own roll.

What that split surfaced is the finding this pass acted on:

cb-play already had a renderer. It had had one since stage 0. It was showing 24 of the 41 leaf paths a populated GroundView carries.

A tier-L pass would have opened by surveying 2D toolkits. It would have been surveying how to draw a table whose text renderer was silently dropping the DARVO state machine, the GROUND practice, the scoring mode, the Focus tokens, the discard pile, per-seat protection, and every part of the outcome except the headline. The port was the wrong first question, and nothing in the structural tier derivation could have said so, because the trigger fires on "creates a capability port" — a property of the plan, not of the code.

Stated against my own prior. In the turn before the roll I recommended tier L and said the AM-4a headroom "should lead the survey". The roll deleted the survey and the pass was better for it. One data point, and it is one data point — but it is the first evidence the CHAOS gate has produced in six declarations, and it points the way the mechanism's defenders hoped and I did not expect.

What tier L would have caught that this pass did not. Being honest in the other direction: this pass made a real interface change — cb_game_runtime::replay::open, extracted so the inspector and the replay gate share one bundle reader — with no review and no ADR. It is small, dev-only, and flagged in its commit per the chaos limits, and I believe it is right. But "I believe it is right" is exactly what a tier-L review exists not to accept, and that is the cost side of the ledger.

Carried, not concluded. Six of twelve declarations used, one override. The retire condition in gates.toml asks whether an overridden tier ever produces a different outcome than the argued one. It just did. That is recorded as the gate's first caught entry, and the window stays open — one favourable fire is not a calibration.

2. The field-coverage gap, before and after

	leaf paths shown
before (stage-0 renderer)	24 of 41
after	41 of 41, plus 1 declared omitted

Measured, not counted. The old renderer was spliced back in beside the new fixture and run. That matters, because the first two figures I wrote for this — "42 of 43, up from 24", in the T01 commit message — were counted by hand and both were wrong. The trusted-arithmetic class, in a pass whose entire subject is a gate against unverified claims.

The measurement also corrects itself in the honest direction: outcome.winners.* shows as missing under the new token but was being rendered, in a different format. So 16 fields were genuinely absent, not 17. That is the number to quote.

The one deliberate omission is players.*.hand — null for every seat but the viewer, where the absence is what hand N card(s) renders.

The gate caught a field its own author had missed, on its first run, before it had ever been committed: players.*.hand was in neither list. That is the entry worth having in M-D1-MUT's ledger, more than the three mutations I designed on purpose.

3. What no test could have caught, and why

Worth writing down as a class, because it is new here:

Every assertion a renderer test naturally makes — the output mentions the round, the output is non-empty, P2's hand does not appear — is satisfied by a renderer that shows a third of the state.

This is the harness-does-nothing class in presentation form. The harness runs, the assertions are real, and they are all satisfied by the defect. What breaks it is asserting over the shape of the input rather than the content of the output: walk the serialized view, require every leaf path to be classified, and make silence cost a build.

Paths, not keys. problem occurs under a DARVO target, a GROUND choice and a Selection; a key-set walk would let one of the three vouch for the other two.

4. Cost, shape, and the meta budget

pass	kind	responses	cost	$/response
CB-WP-0006	meta	158	$57.22	0.362
CB-WP-0007	meta	21	$7.95	0.298
CB-WP-0008	product	134	$17.38	0.123
CB-WP-0009	meta	46	$11.31	0.246
CB-WP-0010	product	26	$4.08	0.157
CB-WP-0011	product	45	$4.23	0.094

The cheapest pass per response yet recorded, against a previous best of 0.123. Two mechanical causes, both boring and both worth keeping:

• it opened immediately after a compaction (§ below), and
• every task had a cargo test between it and being wrong.

The CB-WP-0009 figure moved from $6.73 (as reported in CB-EV-0008 §4) to $11.31. That is not a correction: CB-EV-0008 was written during CB-WP-0009, and the responses after it are attributed to the pass they belong to. A pass cannot measure its own final cost, and reporting one mid-pass will always read low.

Session shape

metric	this pass	previous pass opened at	target
SH-1 mean context	147,808	338,852 [HARD]	≤ 200,000
SH-2 p90 context	149,262	339,342 [SOFT]	≤ 300,000
SH-3 batching	0.0% [SOFT]	0.0%	≥ 20%

Both context metrics went from breach to comfortably inside, and the lever was one /compact before the pass opened. That is now the second time the same lever has produced the cheapest pass on record (CB-WP-0008 was the first). Two observations are not a law, but the prediction is cheap and falsifiable: the next pass opened above the SH-1 hard line will cost more per response than 0.123.

SH-3 stays at 0.0% against a 20% floor and remains unfalsified and unremedied. It has now read 0.0% for five consecutive passes.

Meta budget

58% of the trailing three, against a soft 25% — up from 45%, during two consecutive product passes.

That is not an instrument defect, but it is a property worth naming: the budget is a cost share, so two cheap product passes move it less than one expensive meta pass moved it up. CB-WP-0009's $11.31 is 58% of the $19.62 the window holds, and it leaves the window on the next pass.

Falsifiable prediction: if the next pass is product, the trailing-3 meta share drops to 0%, because CB-WP-0009 will be the pass that rolled off. If it does not, the windowing is wrong in a way neither CB-EV-0007 §3 nor CB-EV-0008 §1 found.

No product work was displaced by meta work in this pass — nothing meta was opened. The number is over the line and reported under the rule that requires reporting it.

5. Open

• Stage 1 is not shipped. This is its inspectable half. INTENT's stage-1 line stays open; the port, the visualization and drag-to-propose are untouched.
• The port declaration is still owed, structurally tier L, with its own chaos roll, and with AM-4a's 3,750-line headroom as its leading constraint. That number is unchanged by this pass — nothing was added to the shipped runtime.
• CHAOS has its first caught entry and stays open to 2026-09-30.
• GATE-REVIEW still has none, one pass older.
• SH-3 at 0.0% for five passes. Either the floor is wrong or the behaviour is, and neither has been argued.
• cb-play is now two tools in one binary. Play and inspect share a renderer and nothing else. If a third mode arrives, that is the second use, and the split should be reconsidered then rather than now.