Failure cascade calibration v4

5 October 2026. This is a bounded balance revision in response to the actual v4 playthrough, whose six unattended faults and extreme heat/pressure were still easy to recover from. The numerical comparisons below are deterministic simulation replays, not a newly completed browser playthrough or final difficulty approval.

The first warning still allows a measured response. After its existing 18-second escalation grace, an unresolved incident now grows along a capped convex curve rather than an indefinitely gentle straight line. Its multiplier never exceeds six. Crew correction, quarantine, targeted purification, save migration, analog observance snapshots and the 18-second recoil travel grace retain their existing behavior. A curse does not cause instant death.

The cascade now follows the physical connections. A neglected recoil witness on the same circuit multiplies an existing mechanical obstruction's load by 1.35. Hot motive bearings wear faster under return pressure; heat erodes cooling seals and tightens breech clearances. Sustained heat beyond 78 degrees and pressure beyond 74% increasingly damage the hull, and the two together add a bounded shared load. A crossing into the coupled overload band emits an explicit physical warning. Lower readings remain routine maintenance territory.

This gives the short religious responses a useful planning role: spending 1.4 seconds reseating a loose cooling witness can preserve the circuit while delaying a reagent-consuming intake rite. It does not clear the separate mechanical obstruction. Choosing the causal cooling response before another lower-risk task arrests three wear paths at once. Halting motive demand remains an available veto, trading journey time for reduced heating while faults and enemy fire continue.

The late recovery comparison uses the full seven-second traditional coolant rite, which is slower than the condition-led physical sequence. It demonstrates that recovery remains possible while the earlier structural loss survives the halt; it does not claim all human strategies produce that exact value.

For a separate critical replay, both strategies halted, spent the same oil/incense/seal, performed a coolant rite and purged the vox, and used the same seventeen seconds including a six-second travel/attention interval. Cooling first finished with 87.8 hull and 55.9 coolant health; visiting the vox first finished with 79.8 hull and 44.5 coolant health. The order therefore matters without adding an arbitrary extra reagent penalty. A free recoil response earlier in the same watch preserved roughly 21 additional coolant health and eleven pressure points by the comparison time, while the underlying obstruction remained unresolved.

Three focused regression cases establish accelerating neglect, recoverable late triage with lasting cost, useful task order and the free attention tradeoff. The existing competent-maintenance campaign still reaches victory across seeds 1, 41, 1984 and 4096, with three replacements, bounded supplies and more than forty hull. These automated routes include decision delays but do not prove human walking, perception or uninterrupted crisis handling. The next acceptance is an actual uninterrupted watch against this revised curve.