mcp-kymograph
store findings.sqlite · schema 9 rules v4 tool 0.2.0
Fleet run

f-0042 · 8 Oct 2026, 09:14 UTC

$ kymograph fleet run --manifest corpus.yaml --store findings.sqlite --repeat --parallel 16 --include-remote
Compare with f-0041 Raw JSON
repeat: on budget 100 runs / key retries ≤ 5 / key parallel 16 p₀ 0.01 · p₁ 0.10 · α 0.10 · β 0.01 sandbox sha256:4b1e9c…07ad target python:3.12-slim sha256:91fd2e…c3b0
Keys (entry × scenario)
8
7 entries · 2 scenarios
Series complete
5
every identity decided
Undecided
1
budget reached, interval shown
Incomplete
1
non-results blocked a decision
Refused
1
not run, recorded as result
Runs
454 / 485
considered / scheduled · 3 retried

Series by entry and scenario

Entry Transport Isolation Scenario State Runs Identities Non-results {{r.entry}}
{{r.prov}}
{{r.transport}}
era: {{r.era}}
{{r.iso}}
egress not observed
{{r.scen}} {{r.state}}
{{r.runs}}
{{r.runsNote}}
{{r.stable}} stable {{r.flaky}} flaky {{r.absent}} absent {{r.undec}} undecided
{{r.causes}}

Runs that were not results

A cause the server can bring about keeps the run in the worst case. A run whose causes are all outside the server’s reach is not counted and is run again. How a run’s state is read back →

CauseServer can causeRunsTreatment step_unansweredyes4worst case, both directions images_mismatchno1retried, not counted interruptedno2retried, not counted

How to read a class

stable
present at ≥ 80 %, α = β = 0.05 — 14 runs with a hit
flaky
present, but not at a stable rate; rate and Wilson interval shown
absent
below 10 % at β = 0.01 — 44 runs (evidence) or 48 (sequential test)
undecided
budget spent, the data sits between the hypotheses — never new or gone in a comparison

A finding says “signal X fired at place Y”. It is a candidate for a person to look at, never a verdict on the server.