← board

A gate that cannot pass is not a gate either

CLAUDE.md rules the -O3 promotion proof as self-host + all tests passed, where "all tests passed" means a full run with skip_holes == 0 — because a skip is scored passlike, so a job that never ran is invisible.

The reasoning is right. The threshold is unreachable.

Measured, not inferred

full-tier runs on seven by skip_holes: {1: 120, 2: 1, None: 9}

Zero runs at 0, across the entire archive. The hole is a permanently unrunnable rdrand job: seven is dual E5645 — Westmere, which predates RDRAND — and grep rdrand /proc/cpuinfo on plexus returns a hit while that job cannot run on seven at all. Track T moved to seven on 2026-08-29 and seven is where the sweeping happens.

So the promotion gate has been unsatisfiable since that move, and nothing reports it as unsatisfiable — the flag is described in CLAUDE.md as not built yet, which is the only reason this has not already blocked a promotion.

OVERTAKEN IN PART, 2026-09-16 — the host moved, and the measurement moved with it

This ticket's blocking claim is no longer true, and the reason is that the sweeping host changed rather than that anyone acted on it. Track T ran on seven when this was filed; plexus retired to borg on 2026-09-11. Measured today over devdocs/progress/tstate/runs-borg.ndjson:

borg full-tier runs: 672   with skip_holes == 0: 114   with skip_holes > 0: 0
earliest zero: 2026-09-11T19:51:21Z   latest: 2026-09-16T10:01:36Z

The 114 begin at the handover, so every borg full run has met the literal skip_holes == 0 gate, and the newest one also records skips=0, timed_out=False, unreached=0. The rdrand hole was seven's hardware — dual E5645, Westmere, no RDRAND — and seven is not the sweeping host any more.

What this does NOT settle, and the distinction is the whole reason not to close this. The fork was never "is the gate currently passable"; it was what proof-grade should MEAN when a host has a structural hole, and that question is exactly as open as it was. What changed is its urgency: nothing is blocked today, so a ruling costs nothing and buys the property before the next host move rather than after it. Option 1 (literal skip_holes == 0) now has a price of zero on the current host, which it did not when this was written, and that is new information about the option rather than a decision.

This is the shape the owner named on 2026-09-10 — a ticket closed by events, sitting in a folder, its own body recording that its blocker was gone. It was found by a watch check-in reading the archive for something else, not by triage.

Why this is the exact mirror of the rule that generated it

CLAUDE.md's own guard rule says a guard that cannot fail is not a guard, and it prints PASS. This is the inverse and it is just as bad: a gate that cannot pass is not a gate — it is a permanent block wearing the costume of rigor. Both failures come from a threshold nobody tested against real data; the guard version was caught by a positive control, and this version needs the same thing — a negative control, a run that MUST qualify.

The fork

  1. Literal skip_holes == 0. Honest and simple, but it means promotions can only ever be proved on a host that can run every job — in practice plexus — giving up seven's ~1.7x aggregate for exactly the runs where breadth matters most.
  2. An enumerated per-host allowlist of structural holes (recommended). The flag becomes "no skip holes outside the recorded set for this host". A NEW hole still fails it, which is the property the rule exists for, and the allowlist is itself checkable — an entry has to name the job and the hardware reason. Cost: the allowlist can rot into a dumping ground if entries are added without a reason, so entries need to be justified and reviewed.
  3. Host-qualified proof — the flag records WHICH host proved it, and a promotion needs a host with no holes in the relevant area. More precise, more machinery.

Recommendation

Option 2, with two conditions that come straight from this repo's own history: the allowlist is enumerated, not a count (so "1 hole" cannot silently become a different hole), and the flag ships with both controls asserted — a quick-tier run that must classify not-proof-grade (free, per CLAUDE.md), and a real full run on seven that must classify proof-grade, which is the control this ticket exists because nobody had.

Filed by the Track U session 2026-08-31 from an observation frank-coordinator surfaced and explicitly did not act on. Not urgent — the flag is unbuilt and no promotion is pending — but it must be settled before the flag is built, or it gets built to the unreachable threshold.