Semantic project planning · Essai 19
What would make the required work possible?
An infeasible plan is a question, not an explanation. Which changes could make the promised work possible, and which of those changes survive a tighter access condition?
The earlier work generator derives installation, evidence and removal actions from required outputs. This essai searches one level above its plans: it enumerates changes to a small catalogue of available capabilities, without changing those outputs. It can give two different repairs, explain why both are minimal, and distinguish the cheapest repair now from one that remains usable after access narrows.
Two tunnel units still have to be installed
Both units need installation and inspection evidence. All temporary equipment must be removed at handover. Initially no setup capability is available. Restoring a shared gantry costs four illustrative units; restoring each unit's trolley costs one. These are hypothetical capability proposals, not approved purchases or verified site facts.
At open access, two trolley capabilities are the cheapest repair. If the same repaired model must also work at tight access, only the gantry repair survives. Reducing the requested scope to nothing would also make the planner return a plan. It is explicitly rejected as a repair of this End.
The explanation behind a repair
A minimal blocking cut is an inclusion-minimal set of setup capabilities whose joint absence makes the requested work impossible. With both units required and open access, there are two: gantry plus trolley A; gantry plus trolley B. At least one capability from each cut must be available. Restoring the gantry meets both requirements; restoring both trolleys meets them separately.
This is why “minimal” and “cheapest” differ. Minimal means no selected restoration can be removed while preserving feasibility. A single expensive shared capability and two cheap local capabilities can both be minimal. Cost then orders the alternatives; it does not erase the explanation.
From model reconciliation to goal-preserving repair
Chakraborti and colleagues search model changes that reconcile a human's expected plan with an agent's plan. Their distinction between concise explanation and changes robust to further model correction prompts a different question here: which explicitly permitted capability changes restore a plan without changing its goal?
This is our construction, not their algorithm. It uses the existing ontology-generated action model and searches all eight restoration subsets. For each it performs exact finite state search. A second access condition is checked separately: the same capability purchase must permit a plan in both conditions, but the executable sequence may differ. This is scenario robustness of a capability set, not a single online plan that tolerates an unexpected mid-execution closure.
Why the cuts and repairs coincide in this catalogue
Availability only adds possible setup actions; it removes no goal or physical precondition. Feasibility is therefore monotone in the available capability set. If the available set misses a blocking cut, it is infeasible. Conversely, any infeasible available set has an absent set containing a minimal blocking cut, by finite deletion. Hitting every cut is thus necessary and sufficient. The empty cut in closed access says that even all three setup capabilities cannot repair the declared access restriction.
The independent planning model reaches the same feasibility decisions for these 216 configurations. The ontology has no demonstrated planning advantage in this test. Its named work and evidence relationships give a reader a way to trace the result. A different repair changes an inconsistent commitment. Try locking the shared interface in the four-node example.
Independent checks and continuity
216 configurations cover requested scope, both access conditions and all previously available capability subsets. Each candidate is compared with a separately hand-authored PDDL domain and independent parser/searcher; 744 solved plans are replayed. Tests also check the cut/repair duality, minimality and the scope-waiver exclusion. All earlier ontology, measurement, handover and work-view applications remain separately reachable.
Strongest objection. A capability catalogue defines what this search can imagine. An access redesign, new method, failed inspection or different physical model could change every answer. Here an inspection action supplies evidence by stipulation; this does not replace the richer measurement-and-acceptance enquiry. No repair is a site fact, authority or engineering approval. The new result is an explanation of a bounded infeasibility, with the obligation preserved.
Inspect the original work model · Inspect measurement and acceptance.