Foray 285 · Essai 02 · Executed native-tool comparison
The latest row is not the decision history
The PDA platform kept the current change and its reasons after restart. It could not recover who made the earlier decision or when. An ordinary versioned ledger could.
A deliberately narrower test
The first essai tested a fictional winter fleet and distinguished genuine retained records from unsupported resource and financial answers. This trial asks whether the change-control module can reconstruct an earlier decision after correction.
Two invented histories differ in their first approver and first approval date. Both then withdraw the change and approve a revised plan. We entered the same substantive fields through the native MCP tools, closed the server, started a new process and retrieved each change again. The current upstream commit was unchanged since the first trial.
Null result for a PDA advantage on this job. The result does not assess every module, the hosted service or human usability. It does expose a specific boundary between durable current state and decision history.
Watch the histories converge
Switch the history and the point of retrieval. These are recorded native responses, not a simulation of what the tool ought to return.
Native current row
Versioned ledger for the same entered events
Exact native response at this point
Temporary evidence browser only. The selector changes which saved observation is displayed; it makes no PDA call and changes no record. Refresh restores the initial view.
What can a later reader recover?
Both methods received four entries per history. The ledger preserves the fields already supplied to the tool, plus their order. No timings, keystroke savings or advantage from a more polished interface are claimed.
| Scored question | Native change log | Versioned ledger |
|---|
With event order as the seventh question, four of seven answers survive in the normal current row and all seven in the ledger. Two current-row answers depend on prose. This small answer count is descriptive, not a general quality score. Poorly entered notes would make even the narrative recovery weaker.
Does the current approval supersede the first?
The frozen protocol asked about supersession. The score table substituted the order of three events. The native notes preserve that chronology, and the ledger records changes against the same request identifier. If the rule is “the latest approval after withdrawal governs”, both histories support that inference. If the requirement is an explicit link from this approval to the approval it supersedes, neither representation records that link. That stricter reading scores 3/7 and 6/7 respectively.
Either way, the current rows cannot tell us whether Alder or Birch approved first. The matched histories settle that question independently of a score.
Why this is a loss of information, not a missed keyword
After removing generated identifiers and creation/update clocks, the two retrieved current rows are identical in every field. Their earlier approvers and decision dates differ. A reader of those substantive current fields therefore cannot guarantee the earlier answer. Identifiers and write clocks distinguish the records but do not encode the missing decision fields. Inspection of the actual update implementation confirms that the latest approver/date replace the old values.
The notes append successfully. We credit them for the old and corrected availability basis and for the described transition order. We do not call the history wholly lost.
The other routes were checked
The native tool inventory includes history functions for other workflows, but no change-transition history query. The inspected change registry has no audit-chain wrapper; the isolated run created no native audit file. The SQLite store has a change_requests table and no separate change-history table. Generic server logs name tools, identifiers, timings and outcomes; they do not retain the overwritten decision payloads. The practitioner's guide describes transition-history behaviour more strongly than this inspected implementation provides.
An invalid status was rejected and the current row was unchanged. That successful validation check is separate from the missing history. Our observer transcript does retain every call: it supplies the evidence for this trial, and could underpin an external ledger, but it is additional machinery rather than a native retrieval capability.
The useful implementation requirement is now precise: preserve each decision event, its subject and predecessor, actor, effective date, reason and changed assumptions; retrieve them through the normal interface. Appended notes can supplement that history. They should not silently become its sole carrier. Whether that addition is worth adopting still requires charging intake, maintenance and actual retrieval effort.
Sources, reproduction and limits
Kitchenham, Pickard and Pfleeger(1995), especially the case-study planning discussion on p 55, supplies the discipline of an explicit hypothesis, baseline, variables and confounds. It does not establish this tool's quality. The question and comparator were recorded before execution.
Inspected PDA source commit · Change handlers · Practitioner guide.
All 17 requests/responses and runtime versions · Relevant native schemas · Baseline ledger · Assessment · Read-only reproduction checks. Native handlers were unmodified. Isolation changed only the database and audit destinations and sent structured logs to stderr. There were no external model calls. No private client data or production records were used.