Design the :calibrate protocol #42

New Issue

Closed

opened 2026-07-14 16:52:05 +00:00 by jared · 1 comment

jared commented

2026-07-14 16:52:05 +00:00

Owner

Map: #31
Blocked by #40

Question

Define the calibration loop: sampling vs every-file (sample-and-extrapolate to directory rules), cheap-model categorization fan-out, strong-model (Opus/Fable) confirmation step, retest budget/stopping criteria, and how confirmed rules land in the global rulebook.

Map: #31 Blocked by #40 ## Question Define the calibration loop: sampling vs every-file (sample-and-extrapolate to directory rules), cheap-model categorization fan-out, strong-model (Opus/Fable) confirmation step, retest budget/stopping criteria, and how confirmed rules land in the global rulebook.

jared added the

wayfinder:grilling

label 2026-07-14 16:52:05 +00:00

jared referenced this issue

2026-07-14 16:52:37 +00:00

Assemble the spec + ADR set #46

jared self-assigned this 2026-07-14 17:38:07 +00:00

jared referenced this issue

2026-07-14 18:12:20 +00:00

Decide deletion confirm-gates and temporary-tier durations #43

jared commented

2026-07-14 18:12:45 +00:00

Author

Owner

CONSTRAINTS from #43 (Decide deletion confirm-gates and temporary-tier durations), two things :calibrate must carry:

DETERMINISM-PROMOTION NUDGE (new design principle, mechanics are yours). #43 split delete-once-served into served_when_path (deterministic, scanner-proven -> may delete silently) and served_when (free-text, classifier-judged -> ALWAYS forced to confirm). When a rule's signal is the subjective kind, the tool must not merely downgrade it: it names the subjectivity and recommends a concrete structural convention (an archive/ bucket, a status: shipped frontmatter key) that would let the rule graduate to served_when_path and go silent. Hygiene should push projects toward structurally-obvious maintenance; confirm-fatigue is the incentive to fix the convention. Where that nudge surfaces (check report vs :calibrate run) and whether it may propose the rule edit itself are :calibrate's to design.
The per-rule confirm: true escape hatch is HUMAN-SETTABLE ONLY. A model-proposed rule may not set it — :calibrate can only ask a human to. Rationale: left to a model it gets used too liberally and everything drifts back to always-confirm.

CONSTRAINTS from #43 (Decide deletion confirm-gates and temporary-tier durations), two things :calibrate must carry: 1. DETERMINISM-PROMOTION NUDGE (new design principle, mechanics are yours). #43 split delete-once-served into served_when_path (deterministic, scanner-proven -> may delete silently) and served_when (free-text, classifier-judged -> ALWAYS forced to confirm). When a rule's signal is the subjective kind, the tool must not merely downgrade it: it names the subjectivity and recommends a concrete structural convention (an archive/ bucket, a status: shipped frontmatter key) that would let the rule graduate to served_when_path and go silent. Hygiene should push projects toward structurally-obvious maintenance; confirm-fatigue is the incentive to fix the convention. Where that nudge surfaces (check report vs :calibrate run) and whether it may propose the rule edit itself are :calibrate's to design. 2. The per-rule confirm: true escape hatch is HUMAN-SETTABLE ONLY. A model-proposed rule may not set it — :calibrate can only ask a human to. Rationale: left to a model it gets used too liberally and everything drifts back to always-confirm.

~~jared referenced this issue 2026-07-14 18:13:27 +00:00~~

Wayfinder map: lifecycle-aware doc hygiene (ruthless os-doc-hygiene) #31

jared referenced this issue

2026-07-14 18:17:52 +00:00

jared closed this issue

2026-07-14 18:18:33 +00:00

jared referenced this issue

2026-07-14 19:02:26 +00:00

Pick calibration project #1 and success criteria #45

jared referenced this issue

2026-07-14 19:04:35 +00:00

Pick calibration project #1 and success criteria #45