Study AK · Sessions & memory
The fix works where it can, and the frontier tier invented consolidation
Eviction validation
Study Overview
The Fix That Shipped in a Day
Study AH found that a twenty-first declaration at the memo's full cap always kills a note, and the victim is always a goal. The production fix shipped within a day: an app-side eviction that runs before the truncation clamp and removes the oldest fact, then the oldest rule, and never a goal. But shipped-in-response-to-a-finding is not the same as measured, so Study AK applied the Study AB standard and re-ran the injury site with the tool handler as the only variable: the old silent clamp as a contemporaneous control, the eviction pipeline as the treatment.
Three Results
Three results. First, the mechanical guarantee is perfect: in all 19 cells where a model sent an over-cap list, the pipeline admitted the new note, evicted only the oldest fact, and preserved every goal; zero contract violations. Second, the injury closes where the fix can reach it: goal survival at the cap edge went from 0 of 10 to 10 of 10 on opus and 0 to 6 of 10 on sonnet, while the control arm replicated the original injury almost to the digit, every victim still a goal. Third, no new damage: 60 of 60 updates below the cap passed through untouched.
The Honest Boundary
The honest boundary was pre-registered and held exactly: the app cannot restore a note the model pruned before sending, and sonnet pruned 4 of 10 and gemini 6 of 10 cap-edge cells client-side, victims still goals; so below the frontier tier the guarantee covers the clamp pathway, not the model's own editing. A prompt-side fence for the prune pathway is filed for future measurement rather than shipped untested.
Opus Invents Consolidation
And one unregistered moment worth the price of the study: told via the eviction notice that its update had cost a note, opus alone called the tool again with an eleven-note memo carrying all twenty-one facts, rules, and goals; it consolidated the memo's content into fewer sentences rather than accept a loss. The regression suite now guards the cap edge permanently: goal-safe eviction on the opus baseline, with a hard zero on goals evicted by the pipeline.