FIELD GUIDE 02 — THE CHECKLIST, IN THE OPEN
The 16 checks, explained
Every BlameGraph audit runs the same 16 weighted checks, in the same order, against the reconstructed handoffs of a run. Findings are charged in blame units against the step or handoff responsible, and the weights are policy — yours to set. Same trace, same engine version, same policy: same verdict, every time. This page is the public reference for what each check examines. See them run on a real trace in the free audit.
Empty payload received
CriticalThe sender produced fields; the receiver got an empty payload. The run continued anyway — on nothing. When a whole payload vanishes, this one finding replaces the per-field drop charges for that seam.
Field dropped at handoff
HighFields the sender produced are absent from what the receiver got. Charged once per seam, not per field — and scaled by the loss: losing everything a seam sent costs far more than losing one field of many.
Field invented
MediumA field enters the record untracked: present in the run's final output, never sent by any step. Computed work product along the way is normal; a source fact that appears from nowhere at the end is not.
Value mutated across handoff
HighSame field on both sides of the seam, different value. A pure sign flip (−12 → +12) displays as critical. On fields a policy marks critical — money, ids, account numbers — a mutation counts double.
Type changed across handoff
HighSame field, different type on each side: a number arrives as a string, an object as a scalar. The value looks familiar; every comparison downstream quietly changes meaning.
Field arrived null
HighThe sender produced a real value; the receiver got null or an empty string for the same field — and treated "nothing" as an answer.
Payload truncated
Low · corroboratingThe received payload is under half the sender's size and at least two fields went missing — the signature of a cut-down extract. It rides along as a corroborating note on a seam already charged for its dropped fields.
Schema / key mismatch
Low · corroboratingThe sent and received payloads barely share any top-level keys. The two sides of the seam were not describing the same shape of data — a corroborating note on the seam's field findings.
Unresolved agent disagreement
CriticalTwo different agents independently asserted different values for the same field, the conflict reached the final output, and no step ever resolved it. Blame lands on each conflicting agent. A divergence already explained by a handoff mutation is counted once — as the mutation.
Retry / loop storm
HighThe same work keeps bouncing: an agent's repeats, a handoff's repeats, or a step's folded tool calls carry substantially the same payload, so the task comes back around instead of moving on. An agent running three times on materially different work is a queue, not a loop — and is not charged.
Cost outlier
MediumA step cost at least 5× the median cost of the other steps in the run — an outlier against the run's own distribution, not a global price list.
Duration outlier
LowA step took at least 5× the median duration of the run's other steps. Where the run waited is where the run's design usually hides an assumption.
Output never consumed
MediumFollowing the declared input_from links: no later step declares this step's output as its source, and none of its fields turn up in a later step's input. Work was produced and never used — the chain took a different path than its paperwork says.
Reviewer checked the wrong source
HighA review step validated a value that differs from the value at the origin of its own declared source chain. It checked a mutated copy, not the real source — and signed off anyway. The most dangerous pass in any run, because it converts a fault into a verified fault.
Final value with no source
MediumA field in the final output that no earlier step produced under the same field path with the same value. Provenance is tracked by path, value and source step — a value that only ever existed under a different name does not source it.
Unresolvable input_from reference
MediumA step declares an input_from pointing at a step that does not exist in the run (or at itself). The declared source cannot be resolved, so the audit records it as a provenance finding against that step — and never quietly treats an assumed link as a declared one.
Two honest footnotes. The audit can also surface two zero-blame notes — "nothing auditable" (a run whose steps carry no payloads at all, so no handoff could be checked) and an advisory that a chat-message payload changed wording across a seam, which takes a semantic check rather than a field comparison. Both are deliberately not counted among the 16: no blame units, no effect on the score's arithmetic beyond the audit saying, out loud, what it could not judge.
How the score is built
Each finding charges blame units to the step or handoff the evidence points at. Units total into the run integrity score, and the blame list shows where they landed — step by step, with the field, the values, and the seam. Weights and severity thresholds live in a policy you control: tighten what a failure costs your organisation, mark your critical fields, and the same trace re-scores under your rules. What never changes is the evidence — the values the record actually holds.
Sixteen checks. One trace. The verdict names the step.
RUN THE FREE AUDIT →