[StepenkoAnatoli]

Two gaps closed, and the second was found by asking the repository what it still admitted
to rather than by remembering.

1. Five documents said something that had stopped being true

ADR-0031, ADR-0033 and the delivery corpus all still said the runner-side install path
"has never executed." It has — twice (runs 35600022797, 35608301287), both with
install the vendor CLI and the CLI is a major this adapter supports succeeding.

Corrected in the live documents. The dated records are left alone: the preliminary note
and live-predicate-verification were true the day they were written and say so in their
titles, and a record that is quietly revised stops being a record.

Worth naming as a kit gap rather than only fixing: nothing detects a claim overtaken by
events. corroboration now sees single-sourcing; nothing sees "this says X never
happened, and X happened."

2. ADR-0035's own open question — answered, uncomfortably

That ADR closed with a section headed "What is NOT measured, and is therefore not
claimed": what a signed-in user sees, as against the anonymous caller it had probed.

Collected through the collector itself, dispatched via bin/collect-remote.mjs (the
agent path) as run 35620262684:

"You must be logged in to a GitHub account to view workflow run information, including
for public repositories."

Viewing logs, searching logs, downloading logs and downloading artifacts each require
only "read access to the repository."

On a public repository read access is universal. So any person with a free GitHub
account can open a collector run, read the logs — which echo the dispatched topic in the
env: group the runner prints — and download the artifact, which is the entire corpus.

The 401 and 403 ADR-0035 measured are refusals of anonymity, not of strangers. The
measurement was accurate about what it measured; the population it measured is not the
population that matters.

What changes, and what doesn't

• The naming rules are demoted, not discarded — they cover the anonymous surface and
  nothing past it.
• retention-days: 7 was already set and now has a recorded reason: it bounds the
  window and removes nothing. The comment says so, instead of leaving seven days looking
  like a tidy default.
• Staying public stands. That was decided on plan grounds — a free private repository
  has its environment secrets and variables ignored, so going private breaks the collector
  rather than protecting it. Untouched. What changes is the caveat: not "the exposure is
  limited to shape" but "the exposure is total to anyone with an account, and the remedy
  is a paid plan, not a toggle."

The kit under its own new check

The corpus reports four corroboration warnings — three single-source, one
one-voice, all docs.github.com — and they are left standing rather than argued
away. This is GitHub describing its own permission model: authoritative, and not a second
opinion. The check is right and the research is worth having, which is exactly the
judgement ADR-0036 said it would leave to a reviewer.

Verification

Committed at docs/decisions/2026-09-21-public-run-visibility/ per ADR-0030, ledger
included.

nested   PASS  0 blocking, 4 warning(s), 12 passing
handoff  OK — 3 ledger entries, chain verifies
repo     829 passed, 0 failed · preflight PASS

Cost: 4 credits.

🤖 Generated with Claude Code