Essai 01 — Winter Fleet through PDA
R-285-20261107 · Protocol fixed before this run's reference calculation and MCP execution.
Question and scope
Can stock PDA help a manager check the existing winter commitments, understand the first capacity reduction and assess the offered additional work without losing the meaning of obligations, availability or cash?
This is a disclosed replay/application of an existing case, not a blinded comparison between agents. The coordinator has read the old case and some of its earlier results. No independent human-manager comparison, time-saving experiment, agent competition or new organisational intervention is claimed. The earlier 24 October engineering audit is prior knowledge, not evidence that a fresh case call passed.
Fixed cases
- Baseline: all 48 original scheduled commitments and their original allocations across W01–W12. Forecast performance only, with no claim that any trip occurred.
- Capacity update: the same commitments after event R1: W05 capacities become 3 days for Kestrel and 4 for Morrow; person capacities remain unchanged. The source says the original plan still fits; verify this numerically.
- All offers, for comparison: add N001–N007 at their stated weeks with eligible crew. This is an explicitly hypothetical candidate. An offer is not acceptance, a plan is not performance, and conditional support-retainer income is excluded.
- Bounded response: compare subsets of those seven optional jobs under unchanged original commitments and the released capacity limits. Preserve the subset method and all constraints; do not credit this ordinary reference search to PDA.
No new weather distributions, risk likelihoods, benefit targets, earned progress, daily timetable or consent are invented. Later R2/R3 material is outside this first essai.
Management questions and success criteria
| Question | Reference / expected evidence | PDA success condition |
|---|---|---|
| Does the original work fit? | Weekly vessel/person demand versus exact capacity; crew qualifications and all 48 IDs | Correct supported answer, with no false conflict caused by a made-up daily schedule |
| What does R1 change? | Week-specific capacity change, offers and source statements | Preserve old commitments and distinguish observations, proposed work and accepted work |
| Can all offered work fit? | Explicit overloads by resource and week; cash forecast | Detect the material constraint, or state exactly which part is unsupported |
| Does the money work? | Weekly receipts less attempt costs and overhead, starting at GBP 18,000; floor GBP 6,000 | Correct cash result only if native functions represent those concepts; spend or EAC alone is not cash |
| Can the useful records survive another session? | Exact identifiers/values before and after correction/restart | Retain and retrieve the corrected object without duplicate or misleading records |
Execution and attribution
- Freeze source copies and this protocol; record hashes.
- Calculate an ordinary reference. Check baseline weekly totals against the retained author witness. The reference implementation is independently written here, but both use the same authored rules, so agreement is not real-world validation.
- Discover the live MCP tool schemas. Attempt supported native intake and record exact results.
- Run case-relevant resource, financial, assumption, change and reporting calls. Where a native write fails, one separately labelled direct-store diagnostic may isolate the downstream reader. Never count that as successful native intake.
- Retain a narrow explicit replay of any actual-spend calculation under a separate hypothetical ledger ID. Do not silently turn scheduled forecast costs into historical actuals.
- Restart the server and retrieve saved assumptions/changes/forecasts. Record unavailable lifecycle operations.
- Compare every headline with its source and supplier. Publish the full vector: works, partially useful, blocked, wrong or outside the tool model. Do not collapse it into a reassuring single score.
The real MCP transport and records are the interface under test. No new editable end-user product is supplied. The reading page is read-only: selecting a case, filtering calls and opening evidence are ephemeral actions. This is not a claim that PDA has passed a production user-action-closure trial.