What we tested, and what came back.
Explore nine published evidence entries, including a six-agent policy run and a cross-machine fixture reproduction. Four featured results show what we challenged, what happened and why it matters.
AUG 8, 2026DEMO ORGANIZATION
A denied action left a record.
A test sub-agent's action was denied by policy, and the denial was kept as a receipt linked to two earlier receipts.
- CHALLENGED
- Whether a delegated agent’s refusal leaves evidence, or only its successes do.
- DOCUMENTED SCALE
- One test sub-agent and one denial, linked to two earlier receipts. An engineering verification run in PromptKing’s demo organization, not customer production traffic.
- OBSERVED
- The denial was kept as a receipt, hash-linked into the same chain.
- WHY IT MATTERS
- A reviewer can see what was stopped, not only what went through.
Does not establish: universal enforcement. Test agents in PromptKing's demo organization; the public receipts do not show which runtime they ran in.
Read the record →SEP 19, 2026FIXTURE · LOG-FORMAT TEST
A rewritten log was detected.
We rewrote two steps of a synthetic log made with a third party's agent tool. That tool's own commands accepted the change; a hash held outside the log flagged it.
- CHALLENGED
- Whether an append-only log can be rewritten without anyone noticing.
- DOCUMENTED SCALE
- A six-step synthetic log made with a third party’s agent tool; two steps rewritten in place at the same byte length. The fix was tried on a four-step log.
- OBSERVED
- The tool’s own commands accepted the rewritten history. A SHA-256 held outside the log flagged it, and a hash chain pointed to the changed step.
- WHY IT MATTERS
- A separately held fingerprint exposed a rewrite that the log tool's own checks accepted.
Does not establish: anything beyond a log-format and command-line test. No agent was observed rewriting its history, and the fix was tried on a four-step log.
Read the test →SEP 21, 2026FIXTURE · INTERNAL EXERCISE
A green report was less than it seemed.
A working session had regenerated the answer files its own check compared against, so “all gates passed” said less than it appeared to.
- CHALLENGED
- Whether “all gates passed” meant the new verifier still matched the old one.
- DOCUMENTED SCALE
- Five preserved fixture assessments. Four mutation cases were testable and passed; two were explicitly not testable.
- OBSERVED
- The repair had regenerated its own answer key. With the original files restored, the pinned incumbent and the portable verifier both matched all five.
- WHY IT MATTERS
- A green status is only as strong as whatever supplied the expected answer.
Does not establish: anything beyond one verifier profile in a fixture-only internal exercise. It describes one verifier profile, not PromptKing as a whole; its evidence package is held internally.
Read the note →SEP 22, 2026FIXTURE · REPRODUCED ELSEWHERE
The same fixture results on another machine.
A founder-dispatched external AI runner executed the portable verifier twice on each of five fixtures, on a different operating system and Node version. Repeat runs were byte-identical.
- CHALLENGED
- Whether the same fixtures give the same results on a different machine.
- DOCUMENTED SCALE
- Five fixtures, each run twice, on a different operating system and Node version.
- OBSERVED
- Repeat runs were byte-identical.
- WHY IT MATTERS
- Repeatable results across environments are a step toward checks that other reviewers can reproduce.
Does not establish: independent third-party validation. The package the runner used has not been checked byte for byte against PromptKing's own fixtures, and the result is scoped to the current homepage record. The run report is private today; a redacted copy is in review.
See the entry →Scale, observed results and limits are taken from the published entries.
