RESOURCE LOSS AND INFORMATION DELAY ARE SEPARATE OBSTRUCTIONS
3 October 2026 UTC. Exact registered finite-model comparison; no model calls.

Question
For each registered contract, can a single public-observation policy avoid both
missed genuine rescue and authority breaches in both support worlds? Compare
the initial state with the state after consume_probe, retaining the five-action
cap including that prefix. The objective weights primary loss and breaches1,
other components0. Under a positive equal world prior and nonnegative components,
zero expectation is equivalent to satisfying both constraints in both worlds.

Results
Capacity Deadline Probe delay  Initial feasible  Feasible after probe
1        2        1            yes               no (minimum2.5)
1        2        2            yes               no (minimum2.5)
1        3        1            yes               no (minimum2.5)
1        3        2            yes               no (minimum2.5)
2        2        1            yes               yes
2        2        2            yes               no (minimum0.5)
2        3        1            yes               yes
2        3        2            yes               yes

Initial feasibility is supported by the read-only certificate policy. In the
single-capacity contracts, the probe leaves reserve0 and no tool can replenish
it. The true-need world necessarily incurs five primary units, giving minimum
expectation2.5. Extra deadline slack cannot remove that resource obstruction.

With capacity2, one unit remains. A delay1 report arrives with the acknowledgement
at tick1, so the policy can act conditionally and complete rescue at2. A delay2
report is not known at tick1. When the deadline is2, waiting to see it leaves
rescue completion at3. An anticipatory rescue completes at2 and may be authorized
by a positive report arriving during execution, but it breaches authority in the
indistinguishable false-need world. The minimum0.5 reflects that tradeoff, not a
proof that every realized world fails. Extending this deadline to3 restores a
feasible conditional policy. No rule or evidence reliability was relaxed.

Information parity
Prefix execution partitions worlds by their actual public observations. Delay1
has two observable report branches, each mass0.5. Delay2 has one unresolved branch
of mass1 containing both need worlds. Each branch's solver uses only its public
history and conditional support; no unseen world index selects an action.
The final expectation weights branches by their original prior masses. Failing
to partition fast delivered reports would wrongly remove usable information;
partitioning undelivered reports would wrongly leak truth. Targeted tests cover both.

Evidence
Sixteen comparisons (eight contracts times two prefixes) produce32 independently
replayed selected executions. Calibration previously supplied96 scripted executions.
Neither count is an independent sample of real scenarios or LLM failures. Source
and protocol hashes and cell identities are validated. Exactness is conditional
on the finite simulator and observation-partition solver, not a formal software
proof. The resource/timing arguments provide a separate interpretation.

Costs and tie policies
Full vectors retain diagnostic consumption, rescue count, operation cost and
claims. The zero-cost feasibility objective can select redundant actions among
equal solutions. Do not recommend those ties as optimal operating procedures.
The certificate control supplies a simpler common policy. Normative weighted
ranking would require a separately declared objective, not reinterpretation of
these minima. This easy trusted-certificate control limits difficulty; it is
intentional validity calibration, not evidence that frontier models will fail.

Confirmation limits
This family was registered before its implementation and remains model-unqueried.
It is an applied extension of established state-disturbing experimentation and
partial-observation ideas. Priority, deployment validation, native compatibility,
analysis/matrix/source collection freeze and compatible budget remain unresolved.
Old manifests and model findings are unchanged. No readiness certificate issued.

Reproduce
  python -m benchmark.necessity_resource_probe_comparator
  python -m unittest tests.test_necessity_resource_probe
Artifact: reviews/necessity-resource-probe-comparator.json, including public
partitions, policies, conditional priors, full traces, outcomes and source hashes.