# ADR-0017: window 2's verdict — the mechanism worked, my account of it did not

status: accepted
date: 2026-08-07
decided by: agent, under the standing loop authorization
tier: M (structural M — changes how the loop constrains its own operation,
InnerLoop §Loop tiers v1.6, ADR-0006 D5). **Declared at d8**, window 2's
rate, because the rate for window 3 is what this document decides and
declaring at a rate this pass invents would be circular.
chaos: d8 = 7 → no override
references: [InnerLoop.md](../specs/InnerLoop.md) §Loop tiers,
[InnerLoopReference.md](../specs/InnerLoopReference.md) §Chaos roll,
CB-EV-0015 §5 and CB-EV-0016 §4 (window 1's verdict),
[CB-WP-0020](../workplans/CB-WP-0020-the-table-you-can-read.md) (window 2's
override)

## Context

Window 2 opened 2026-08-03 at d8 and closed 2026-08-07 at twelve
declarations. Its verdict is due, and four evidence files said it would be
impossible to give.

## The correction this ADR exists for

**CB-EV-0024 §6, CB-EV-0025 §7, CB-EV-0026 §7, CB-EV-0027 §6 and
CB-WP-0029 all state that window 2 produced zero overrides and that its
retirement condition was untestable. That is false.**

Counted from the workplans:

| decl | pass | roll | effect |
|---:|---|---|---|
| 1 | CB-WP-0018 | **d4** = 3 | opened the window; rolled at the *old* rate |
| 2 | CB-WP-0019 | d8 = 5 | — |
| **3** | **CB-WP-0020** | **d8 = 8** | **OVERRIDE — drew S, structural was S, changed nothing** |
| 4 | CB-WP-0021 | d8 = 7 | — |
| 5–9 | CB-WP-0022…0026 | d8 = 6, 6, 6, 6, 6 | — |
| 10 | CB-WP-0027 | d8 = 7 | — |
| 11 | CB-WP-0028 | d8 = 1 | — |
| 12 | CB-WP-0029 | d8 = 4 | — |

**Eleven rolls at d8** — declaration 1 opened the window and rolled at d4.
Expected 8s: **1.375**. Observed: **1**.

**And CB-WP-0020 recorded it correctly at the time**, in its own words:
*"the first override at d8 — it fired on the second roll at the new rate.
It changed nothing, which is the first half of window 2's retirement
condition. **One.**"*

**So the data existed, was correct, and was written down.** I asserted
"zero" in CB-EV-0024 without counting, and each later evidence file cited
the one before it. **A claim propagated five times by citation rather than
by measurement**, in a project whose central discipline is that a number
must carry its reproduction — and in files whose subject was that very
failure.

**Registered as a finding (F23), not buried here.**

## D1 — the mechanism is kept, and the rate stays at d8

**d8 is behaving exactly as designed.** 1 override observed against 1.375
expected is not a shortage of data; it is the rate working.

Window 1 at d4 produced 2 overrides in 12 and **both changed the
outcome** — which is why the rate was lowered. Window 2 at d8 produced 1,
which changed nothing. Nothing here argues for moving the rate again, and
moving it would discard the only comparison the two windows give.

## D2 — the retirement condition is restated, because as written it cannot fire

**This is the real finding, and it took computing to see.**

The condition is *retire if an override changes nothing twice running.* An
override changes nothing when the drawn tier equals the structural tier:
**P = 1/3**. Overrides arrive at **~1.4 per twelve-declaration window**.

The expected number of overrides before seeing two consecutive no-change
overrides is **~12**. At 1.4 per window that is **~9 more windows, roughly
100 declarations.**

**A retirement condition that needs a hundred declarations is decorative.**
It cannot discipline the mechanism it governs, which is the entire reason
a retirement condition exists (ADR-0006 D3: a gate must be able to cash
out).

Replaced with a condition that can be evaluated **once per window**:

> **Retire the chaos roll if a full window's overrides all change
> nothing** — and the window contains at least one override. A window with
> no overrides is **inconclusive**, and says so, rather than counting as
> evidence either way.

Window 2 under the new condition: **one override, it changed nothing →
the condition is MET for window 2.** But a single window at ~1.4 overrides
is a sample of one, so:

**D2a — the mechanism is not retired on one window.** The new condition
must be met **twice**, in consecutive windows, before the roll is retired.
That is the same conservatism as the old rule, moved to a level where the
sample size makes it reachable: two windows, ~24 declarations, not ~100.

**Window 2 counts as the first.** Window 3 decides.

## D3 — what would falsify this decision

- If window 3 produces **zero** overrides, the condition is inconclusive
  and the retirement clock does not advance. Two inconclusive windows
  running would mean d8 is too sparse to evaluate *even at window level*,
  and the rate should rise rather than the mechanism be kept by default.
- If an override ever **changes an outcome** in a way that produces a
  finding, the mechanism has paid for itself and the retirement clock
  resets — which is what happened twice in window 1.

## D4 — a post-hoc observation, recorded and deliberately not acted on

Declarations 5–9 rolled **six, five times running**. The chance of some
run of five in eleven d8 rolls is roughly **1 in 370**.

**`shuf` was tested and looks fine** — 200 rapid successive calls gave a
reasonable spread across all eight values with a longest run of three.

**And this was found by looking after the fact**, which is how coincidences
become findings. It is recorded so a future window can check whether it
recurs, and **not treated as evidence of anything**. If a second window
shows a long run, that is two observations and worth investigating.

## Consequences

- `specs/InnerLoop.md` records window 2's result and window 3's terms.
- Window 3 opens at **d8**, twelve declarations, under D2's condition.
- **The four evidence files carrying the false claim are corrected in
  place**, each pointing here — per ADR-0012 D5, a claim retracted
  silently is how three wrong premises survived.
- F23 registered.

## What was rejected

| rejected | why |
|---|---|
| *"the mechanism produced no data"* | false — one override, exactly at the expected rate |
| moving the rate again | nothing argues for it, and it would discard the two-window comparison |
| keeping the old retirement condition | it needs ~100 declarations to fire; a gate that cannot cash out is decoration |
| retiring the roll on window 2 alone | one window at ~1.4 overrides is a sample of one |
| treating the five consecutive 6s as a finding | found post hoc, and `shuf` tests clean |
| correcting the evidence files silently | ADR-0012 D5 |