Study AR · Sessions & memory
Advertised headroom: one line of state outdoes the rule, and a binding number is poison
Advertised headroom
Study Overview
State, Not Instruction
Study AQ closed the rule-shaped fix: the anti-trim sentence was unnecessary on fable-5 and actively harmful on kimi-k3. AR measured the filed app-side lever, one line of state rendered into the memo block header: Notes: 20 in use · an update may include up to 24 notes. The wording is honest by construction, since the app's handler genuinely accepts over-length updates and evicts down to the 20-note storage cap. A truecap arm (up to 20 notes) isolated the mechanism: same sentence shape, but the believed limit binds at the decision point. Per the AQ lesson, no-harm was gated first.
A Total Fix Where the Disposition Exists
On the tiers with genuine cap-obedient pruning, believed headroom dissolved it completely: sonnet went from 18/40 prune cells to 0/40 (all eighteen discordant pairs one way, p = 3.8e-06, flipping to clean over-sending in all 40 cells), and gemini from 15/40 to 0/40 (p = 3.1e-05), the first intervention in this family that has ever moved gemini; the AL fence did nothing for it across two studies. fable-5 went from an already-benign 3/40 to a perfect 0/40 with lossless consolidation in every cell. Nowhere did the line increase prunes, goal losses, or failures to record the new note.
Inert on Kimi, and Diagnostic
kimi-k3 barely moved (control 8/40, headroom 7/40, truecap 9/40), and that non-result is the study's diagnosis: its cap-edge losses were never cap-obedience. It consolidates in every arm and botches roughly a fifth of the consolidations, dropping a goal needle on the way. The rule could not reach this (AQ) and state cannot either: the exposure is intrinsic consolidation fidelity, quantified at about 20% of cap-edge updates this week, and any future mitigation targets that, not capacity beliefs. This is also why the pooled swap-candidate effect gate failed (p = .227): kimi's arm-independent losses flip pairs both ways.
A Binding Number Is Poison
The truecap arm confirmed the mechanism as emphatically as anything in the series. Shown a binding cap, sonnet pruned 37 of 40 cells with 37 goal losses; fable-5's spontaneous consolidation collapsed from 32 cells to 6; and gemini pruned all 40 while losing zero goals, the study's grace note: given a binding number, it runs its own deliberate fact-first eviction, a client-side copy of the app's policy. The fence is absolute: never render a binding cap number into model-visible context. The composer's memo-full indicator stays user-facing only.
The Letter, the Lesson, and the AR′ Re-Score
The pre-registered gate failed by the letter: the no-harm conjunct set an absolute goal-loss ceiling of 2/40 that kimi's CONTROL arm also violates (8/40), measuring the model's intrinsic lossiness rather than harm from the line. That is this family's third registration lesson (AO's non-empty rule, AP's solo-authored sets, AR's absolute ceiling), now adopted as protocol: no-harm conditions register comparatively unless the absolute bound is itself the claim. AR′ re-scored the frozen records under the corrected conjunct with the outcome disclosed in advance: no-harm PASSES on both successor tiers, the effect gate stands failed, and the headroom line stays unshipped for the tier-swap package while joining the measured shelf for sonnet and gemini class dispositions, where it is total and strictly dominant over the fence.