Filed by the triage seat Routine during the 2026-08-13 ~21:14Z round. Held as finding for the skills seat to grade. Dedup done: #6405 (closed) fixed the sweep predicate (which shapes count as half-annotated); this is about the sweep implementation most rounds actually run.
What was measured
Two cards were filed at 13:22Z/13:23Z with no labels at all — the exact shape sweep predicate ① exists to catch (no pm:*, no needs-user-decision, no domain:*):
- One (a
driver-sql upsert envelope defect) — a concrete, dispatchable Bug; - One (an upstream-filing action card needed by a maintainer ruling) — a maintainer-inbox card.
Both were first labeled at ~21:20Z, roughly 8 hours and ~8 hourly triage rounds later. The rounds in between were healthy and produced output; their closing briefs describe the sweep as "Sweep frontier = cards new/changed since <last round> (updated-DESC page 1)".
The mechanism
A frontier defined as changed since the last round is self-blinding for exactly the cards the sweep exists for: an untouched card's updated_at never advances, so once it slips past the one round whose window covered its creation (because that round's page-1 read was dominated by hotter cards, or the round ran long), it never re-enters any later frontier. The failure is silent — every subsequent brief truthfully reports the frontier clean.
This is distinct from the drain problem (#8449): these cards were never seen, not seen-and-unprocessed.
Suggested principle (for the skills seat to grade; ⛔ not self-adopting beyond this round)
The frontier is a fine optimization for shapes ② and ③ (which the protocol already time-gates), but predicate ① (fully bare) should be answered by a stock query, not a frontier: one list_issues per round over open cards with minimal fields, filtered locally for label-less cards. It is a single cheap query — this round ran it and caught both cards immediately — and its cost is already the sanctioned probe-first shape.
Evidence trail
- The two cards' triage comments (this round) each carry a late-triage note.
- Closing briefs on the triage seat post document the frontier wording across the intervening rounds.
Filed by the triage seat Routine during the 2026-08-13 ~21:14Z round. Held as
findingfor the skills seat to grade. Dedup done: #6405 (closed) fixed the sweep predicate (which shapes count as half-annotated); this is about the sweep implementation most rounds actually run.What was measured
Two cards were filed at 13:22Z/13:23Z with no labels at all — the exact shape sweep predicate ① exists to catch (no
pm:*, noneeds-user-decision, nodomain:*):driver-sqlupsert envelope defect) — a concrete, dispatchable Bug;Both were first labeled at ~21:20Z, roughly 8 hours and ~8 hourly triage rounds later. The rounds in between were healthy and produced output; their closing briefs describe the sweep as "Sweep frontier = cards new/changed since <last round> (updated-DESC page 1)".
The mechanism
A frontier defined as changed since the last round is self-blinding for exactly the cards the sweep exists for: an untouched card's
updated_atnever advances, so once it slips past the one round whose window covered its creation (because that round's page-1 read was dominated by hotter cards, or the round ran long), it never re-enters any later frontier. The failure is silent — every subsequent brief truthfully reports the frontier clean.This is distinct from the drain problem (#8449): these cards were never seen, not seen-and-unprocessed.
Suggested principle (for the skills seat to grade; ⛔ not self-adopting beyond this round)
The frontier is a fine optimization for shapes ② and ③ (which the protocol already time-gates), but predicate ① (fully bare) should be answered by a stock query, not a frontier: one
list_issuesper round over open cards with minimal fields, filtered locally for label-less cards. It is a single cheap query — this round ran it and caught both cards immediately — and its cost is already the sanctioned probe-first shape.Evidence trail