Skip to content
Archive
No. 012GovernanceCase studyAwaiting review

Verifiers That Cannot Fail

A taxonomy of silent failure from thirteen incidents in one multi-agent operating stack

Verification code fails in a way application code does not: it fails by reporting success. Across thirteen incidents recorded in one multi-agent operating stack between August and October 2026, every failure that cost time arrived looking like a pass, a clean exit, silence, or a confident number. We sort the incidents into six classes: a skipped check read as a pass, the right check on the wrong target, a proxy checked in place of the behavior, a result accepted on plausibility, output that no one receives, and a check that mutated what it inspected. Each class has one cheap defense. The sharpest instance is a purpose-built verifier that printed PASS because it could not read its own threshold, while the defect it existed to catch sat in the same output. We give seven rules the operator's notes record as the fix for each class, and state what thirteen incidents from one operator can and cannot establish.

Date
OCT 2026
Read
7 min
Depth
Depth: DeepDeep
Cluster
Governance
Department
Operations
Provenance
Operator incident notes, Aug to Oct 2026 · draft v0.1
Authors
Lenox Paris, Third Signal Lab
About 7 min left

Thesis

A verifier fails by reporting success.

Application code that breaks produces an error or a wrong output. A verifier that breaks produces a pass, a clean exit, or silence. Across thirteen incidents in one operating stack, every failure that cost time arrived looking like success.

01 · Argument

Six classes cover every incident, and each has one cheap defense.

A skipped check read as a pass. The right check on the wrong target. A proxy checked in place of the behavior. A result accepted on plausibility. Output no one receives. A check that mutated what it inspected.

Evidence

The sharpest case: a verifier that printed PASS because it could not read its own threshold.

The check existed to catch advisors truncated at a token cap. A missing dependency kept it from reading the cap, so it skipped the comparison and printed PASS while a truncated advisor sat in the same output. It now prints UNCHECKED and exits non-zero.

Diagram · Incident map

Wrong target holds four of the thirteen incidents, a proxy in place of the behavior holds three, a skipped check and plausibility hold two each, and silence and mutation hold one each.

Signal 01Weight 5 of 5Proposed
“A check that cannot run must fail loudly, never pass quietly.”
Signal 02Weight 4 of 5Proposed
“Silence is not health. A result no one receives is not a result.”

Tension

Invisible operation and observable verification pull in opposite directions.

An agent worth forgetting is also one whose checks no one watches. The remedy is not visibility everywhere. It is a heartbeat that stays quiet only when everything is confirmed and goes loud when it cannot be.

Implication

Distrust a clean result until the check has been seen to fail.

Every incident was found by a person who checked a clean result by hand. The seven rules move that habit into the tooling.

Sources

The public MoA field notes, the public skill-routing README, and the operator's private working notes from 2026-08-05 to 2026-10-11.

  1. 01Third Signal Lab (2026-08-05). MoA Council Qualification: field notes. Public, third-signal-skill-packs repository. (2026)
  2. 02Third Signal Lab (2026). Skill Routing Economics. Public, third-signal-skill-packs repository. (2026)
  3. 03Third Signal operator working notes, 2026-08-05 to 2026-10-11. Private. (2026)