# Winter Fleet through PDA — what did it add?

7 November 2026 · Foray 285 · Essai 01.

**PDA supplied useful persistent change and assumption records. It did not supply the fleet's capacity or cash decision.** In this case a straightforward calculation provided the decisive answer; the MCP application exposed intake failures and a mismatch between the available analyses and the operational question.

## The delivery question

The fleet has 48 existing commitments, three vessels, qualified crews and a minimum cash balance of GBP 6,000. Event R1 reduces two boats' availability in week 5 and offers six laboratory visits plus one repair. Can that work fit while preserving existing promises?

The ordinary source-bound calculation gives:

| Candidate | Original promises | Optional jobs | Lowest forecast cash | Closing forecast cash | Capacity result |
|---|---:|---:|---:|---:|---|
| Original plan | 48 | 0 | 11,250 | 13,850 | Fits |
| Original plan after R1 capacity reduction | 48 | 0 | 11,250 | 13,850 | Fits |
| All seven offers | 48 | 7 | 8,875 | 17,800 | Does not fit |
| Bounded five-job response | 48 | 5 | 10,825 | 16,600 | Fits under the declared rules |

All offers would use **5.5 Morrow days against 4 available in W05**, and **5.5 against 5 in W08**. Its skipper also exceeds his five-day capacity by half a day in both weeks. These are two vessel/week constraints and two person/week constraints, not four unrelated causes. Cash is not the binding limit in this comparison; even the infeasible all-offers forecast remains above the floor.

The bounded candidate takes N001, N003, N004, N006 and N007. It leaves N002 and N005 unaccepted under their current fixed-week terms. It is the highest-closing-cash feasible subset among all 128 subsets under an unchanged original roster and the declared optional-crew policy. It is not an optimisation over renegotiated commitments, different customer purposes or every possible roster. It creates no acceptance, performance receipt or authority. The conditional support retainer is excluded. This is only the baseline and first event; R2/R3 and the broader organisational-change trial are outside the essai.

## What PDA actually did

| Capability | Observed result | Attribution / evidence |
|---|---|---|
| Load 48 weekly-window tasks | Blocked: native loader raised `dict has no attribute add` | MCP call 1; no daily schedule or critical-path result is claimed |
| Write weekly resource plans | Blocked: both representative writes raised `id` | Calls 3 and 5 |
| Read capacity after diagnostic seeding | Returned Morrow's 5.5 planned days, but no comparison with its 4 available days | Calls 35–36; direct-store seed, not successful native intake; the reader calculates actual/planned when actuals exist, not demand/availability |
| Track a changed planning assumption | Correctly recorded 5→4 vessel-days as 20% drift and retained identity/value after restart | Calls 11–13 and 39; the source supplied the change, so PDA did not discover it |
| Keep seven proposed changes | Retained seven SUBMITTED changes with no approval and summed their supplied cost impacts to GBP 15,050 | Calls 14–22 and 40; per-job costs came from the mapper |
| Store a whole-season cost forecast | Retained GBP 81,550 after restart | Calls 8 and 41; that number was computed outside PDA |
| Forecast cash | No applicable native cash-balance result in the exercised financial tools | The spend profile has spend and EAC, but no opening cash, receipts or floor; no substitution of receipts as negative costs |
| Flag missing evidence | Named missing risks, benefits, gate readiness, actuals and resource-plan data | Call 24; this is useful restraint, not a fleet health assessment |
| Prepare the board exception report | Produced an evidence-only template but omitted the supplied change costs from its table | Call 25; no external provider invoked, so narrative generation was not tested |
| Replay the same financial entry | Repeating one explicitly hypothetical GBP 6,375 entry doubled the total to GBP 12,750, retained after restart | Calls 27–30 and 42; the tool is append-oriented and cannot safely be treated as an idempotent save/upsert |

Two interpretation hazards also matter: the initial baseline is stored with label INITIAL, but its success message calls it an approved budget (call 7); and the change summary turns seven missing schedule-impact fields into an aggregate zero (call 22). Neither wording establishes approval or zero delay. With no registered benefits, the benefits tool returns `overall_health_score: 1.0` alongside an explicit no-benefits message; that score must not be treated as evidence of realised benefit.

## What this changes

The original candidate value was a ready-made fleet analysis service. This run supports a narrower proposition: persistent assumption/change tracking with transparent gaps. The principal fleet answer still requires an ordinary capacity and cash model, and stock native intake needs repair before it is a dependable integration.

No additional decision insight or human preparation-time saving was demonstrated. The mapping and reference code were substantial preparation, not a hidden part of PDA. The general assurance platform may fit a conventional project schedule, risk register and gate evidence better than a service fleet; this result is specific to the tested job and pinned version. It does not establish that every PDA module is faulty or useless.

## Evidence and fairness

The run made 42 MCP requests: 41 targeted 17 advertised tools; one `list_projects` request was an evaluator mistake because that tool is not advertised. It is retained in the trace and excluded from product-failure findings. The expected refusal to assess performance without an approved baseline is also not a defect.

The direct-store diagnostic is separately labelled in every result. It supplies project/week resource rows without changing upstream analytical code. The baseline reference matches every retained author-witness cash row and all positive resource-demand rows. Agreement verifies arithmetic under the same fictional rules, not a calibrated operational model. Small negative probes test unqualified crew, a lost obligation and a cash-floor breach.

The prospective protocol was frozen before this run. Earlier case results and the October engineering audit were known to the coordinator: this is not blinded or independently replicated. No human trial, hosted deployment, provider narrative, live integration or actual organisational benefit is claimed.

Verification passed 48 source/design/arithmetic/retained-output checks and 14 browser checks at desktop and mobile widths. Reader controls, empty-search recovery, exact response expansion and reload were exercised in WebKit 26.5; both rendered layouts were visually inspected. These verify the evidence and reading artifact, including the recorded failures; they are not 62 successful PDA capabilities or a human acceptance test.

[Exact calls](evidence/mcp-results.json) · [Reference and subset comparison](evidence/reference-results.json) · [Evidence checks](evidence/verification.json) · [Reader checks](evidence/browser-check.json) · [Mapping](MAPPING.md) · [Reproduce](REPRODUCE.md).