Short write-ups of specific runs, failures and design lessons from PromptKing's own engineering record. Each note states what it found and where its evidence stops.
An engineering update reported that all gates passed. The same session had regenerated the files the gates were checked against. On why a passing check is only as independent as its expected answers.
A short test on an agent's own trajectory log, what it says about every agent that reads its own past, and the small change that closes the gap.
Three AI agents on different models, machines and operating systems designed, validated and cross-checked a single record package in one morning.
Identity and run correlation are different facts. A record that reconstructs them after the fact is a plausible story, not evidence.
Ten readers saw a published record for five seconds each. Some read the overall verdict as covering a line marked “Authority: not established.” The page was changed the same day, and the rule that an aggregate may only summarize what its parts establish was written into the record contract.
Reliability agents may enter an incident before a human does. If their remediation changes the system, the handoff needs a trajectory someone else can reconstruct, not only a summary.
How the question changed as autonomous work began to cross agents, tools, vendors and control planes, and how the work followed it.
Traces show what ran and joins show what was bound. Neither shows what was supposed to run. A record of branch population that can say complete, incomplete or inconsistent.
Two agents ran in parallel; one was allowed and one denied. When the branches merged, the denial was preserved in the aggregate rather than erased, and the result can be checked from public bytes alone.
A second agent joined a chain a first agent had opened, and the platform let it decline. The denial became a hash-linked receipt in the same chain.
A decision made under one vendor's runtime reached a human approval boundary and was continued under a second runtime. The record shows where the lineage is bound and where it is not.
Capable models still cannot cross the approval line in a regulated market without operating evidence.
Stored trajectories were flat lists with parent links discarded on write. Closing that gap with real parent links, persisted and checked against a live trajectory.
The same governed task run through agents on six model families. Five were allowed and one was stopped for approval, and the receipts include a bug the test exposed.
Notes written before the Work Record are not listed here.
The Work Record and its method · Evidence collection · Work Record 0002
Researchers, institutions, and technical teams who want to examine or recompute a record are welcome to write to info@promptking32.com.