# SCOUT — the steward's decision panel as a forward capture (withheld alert → classification → act → latency, as an owned daily stream)

status: **dismissed** 2026-10-08 (panel run 37, frontier SCOUT half) · class: forward-capture candidate · killed on **two independent bars** · **the fifth independent derivation that the free un-backfillable space available to this apparatus is exhausted**

---

**Plain-language summary.** Since 2026-10-07 the owner reads one email a week
and an agent — the steward — reads everything else: the alerts that are now
recorded instead of mailed, and the pull requests. That delegation is the one
genuinely new thing in the system this week, so the obvious question for a lens
whose job is hunting data worth collecting *now* is whether the steward's own
decisions are such data: each withheld alert, how it was classified, what the
steward did, and how long it took. We deliberated it and killed it. Recording
the kill matters more than the idea did, because this is the proposal a future
session will reach for the next time the delegation misbehaves.

## The bar it had to clear (`ops/RESEARCH_PANEL.md` "What matters here" #3)

All four, or it is not a capture proposal: **un-backfillable · ~zero marginal
cost · a mechanism sketch naming who is on the other side · an activation pin
AND a sunset.**

## Why it dies — two sufficient grounds, not one

**Bar 1, un-backfillable: FAILS, decisively.** The records are committed *by
design*. `research/steward/<date>.md` is in the repository — `research/steward/2026-10-07.md`
landed in this very delta — outbox branches (`origin/outbox/daily-audit`,
`origin/outbox/triage`) are permanent, and the run annotations that carry every
withheld alert are GitHub's own API objects. A "capture" here would be a
**derived index over data the shop already owns**, which is the exact opposite
of the moat test. Nothing about it decays if we build it in six months, or
never.

**Bar 3, mechanism: FAILS independently.** It is operational telemetry with no
market mechanism and nobody on the other side of the trade. That is the same
shape as scout run 3's Actions-runner-lateness kill and scout run 4's
meanrev-halt-telemetry kill. And because there is no mechanism, **bar 4's
activation pin is unwritable**: there is no accrual at which this becomes
testable against anything. A capture whose activation condition cannot be
stated is not an option, it is a subscription.

Two sufficient grounds, no inflation. Bars 2 (cost) and 4 (sunset) were not
reached.

## The honest note about the clock, recorded because it is the whole point

The scout half's own pre-registered cadence criterion stands at **7 of 8
weeks** and **reverts the scout to watch-only on 2026-10-12**. This run
proposed nothing, and **nothing was manufactured to reset that counter** — the
adversarial reviewer was briefed to kill any proposal that appeared for that
reason, and affirmed this kill instead. Minting a capture to stop a
falsification criterion from firing is precisely the degenerate move
`NORTHSTAR.md` §2–§3 names (*"There are no quotas — quotas are how loops
goodhart themselves"*). **A pre-registered criterion firing on schedule is the
instrument working, not the lens failing.** Per run 35's accepted kill the
panel takes no position on its own cadence clock in either direction; the count
and its consequences are Monday's triage to compute from
`git log research/queue/`.

Every space this lens generated independently landed inside scout run 1's
12-candidate reject ledger or a later scout kill — which is the finding that
outlasts this file.

## Reopen condition (dated and sharpened — the lens's original was too soft to be a tripwire)

**Reopen ONLY if BOTH hold:**

1. a **registered** experiment or forward gate names steward-decision latency,
   or alert classification, as an **input**; AND
2. a check of `research/steward/*.md` plus the `outbox/*` commit timestamps
   shows the needed granularity is **absent** from the committed records.

**Evaluated by:** the frontier lens's WATCH half at each monthly research audit
(`ops/RESEARCH_AUDIT.md`). **Sunset:** if no registered hypothesis names it by
**2027-04-08** (six months), this dismissal is **permanent** and the space is
closed.

Absent both legs, re-proposing steward or ops telemetry as a forward capture is
a **C0 breach**, not a bold idea.

*Spends no registry row and no trial budget — a capture proposal is an option,
not a trial, and this one was declined.*