Expand description
Online-learned per-extension compression-threshold deltas (#538, EFF-1).
Closes the quality feedback loop that the static LANGUAGE_THRESHOLDS
table and the savings-only bandit cannot: bounces (compressed read followed
by a full re-read) and edit failures after compressed reads push the
entropy threshold DOWN (compress less), while clean compressed reads and
wasted full reads push it UP (compress more). The learned delta is additive
on top of the static base table, hard-clamped so the base stays the safety
anchor, and decays toward zero daily so stale lessons fade.
Neuroscience analogue: dopaminergic active forgetting — eviction policy is learned from outcome signals, not hardcoded (SleepGate 2603.14517).
Structs§
Enums§
Functions§
- export_
state - Process-global: clone of the full learner state for export (#550).
- flush
- Process-global: flush the buffer to disk (call from shutdown paths).
- learned_
delta - Process-global: learned additive delta for the extension (0.0 in warmup).
- merge_
state - Process-global: merge a foreign learner state in and persist (#550).
- record_
signal - Process-global: record a quality signal for the extension of
path. - report
- Process-global: report lines for ctx_metrics.
- snapshot
- Process-global: machine-readable snapshot for the dashboard (#548), sorted by extension.