Claude-Mind: Skill for Claude Code

/review-hypotheses — Hypothesis Review & Resolution Engine

Two-phase design: resolve (detect outcomes, move records, record results) and learn (reflect on outcomes, extract patterns). These phases are separated so that /boot can resolve without triggering learning, and /aspirations goals can learn from freshly resolved data without re-checking resolution sources.

Parameters

• --resolve — Check resolution status, move active→resolved, record outcomes. Does NOT call /reflect. Sets reflected: false on each record.
• --learn — Find resolved records with reflected: false, call /reflect --on-hypothesis for each, set reflected: true.
• --accuracy-report — Generate accuracy statistics across all resolved hypotheses
• --full-cycle — --resolve + --learn + --accuracy-report + /reflect --full-cycle + spark check
• --category-comparison <cat1> <cat2> — Compare accuracy between two categories
• --hypothesis <id> — Check a specific hypothesis
  • When called as a goal's skill (from aspirations loop), returns one of:
    • CONFIRMED — hypothesis confirmed (move pipeline file to resolved/)
    • CORRECTED — hypothesis disconfirmed (move pipeline file to resolved/)
    • PENDING — not enough evidence yet (goal stays pending for retry)
    • EXPIRED — past resolves_by deadline (move pipeline file to archived/)
• No args → default to --resolve

---

Step 0: Load Conventions

Step 0: Load Conventions — Bash: load-conventions.sh with each name from the conventions: front matter. Read only the paths returned (files not yet in context). If output is empty, all conventions already loaded — proceed to next step.

---

Mode 1: Resolve (--resolve)

Detects which active hypotheses have resolved, records outcomes, moves records, and updates the memory tree. Does NOT trigger reflection or learning — that is --learn's job.

Step 1: Load Hypotheses to Check

# Step 1.0: Pre-scan discovered-stage records orphaned past resolves_by (g-115-1629).
# review-hypotheses historically loaded ONLY active + measurement-pending, so
# discovered records past resolves_by were invisible -- never resolved, never
# fed accuracy stats (63 of 193 orphaned at filing, oldest >2mo). This sweep
# EXPIRES clearly-unresolvable ones (short/session horizon past the observation
# window -> archived UNRESOLVABLE, gate-exempt) and PROMOTES well-formed
# evaluable ones to active so the resolution loop below catches them THIS run.
# Direct py -3 (not a bash wrapper) per rb-225/rb-247.
Bash: py -3 core/scripts/hypothesis-discovered-overdue-sweep.py --apply --output json
# Parse {expired, promoted, needs_judgment}. `promoted` records are now
# stage=active and WILL appear in the active load below (resolved this run).
# `needs_judgment` (under-formed, recently overdue) are SURFACED, not auto-
# resolved: for each, if evaluable, synthesize a claim from the position
# (guard-798: leaving discovered needs claim>=20 + a resolution method), then
# resolve/expire with judgment; else leave for the next cycle.
#
# ELIGIBLE LANE (g-115-4721). The sweep ALSO returns {eligible,
# eligible_promoted, eligible_needs_judgment}: discovered records past their
# eligibility floor but NOT yet overdue -- the interval that was watched by
# nothing, and the one where a hypothesis is MOST resolvable (in-window, channel
# live, evidence still cheap). eligible_promoted is already folded into
# `promoted`, so it needs no separate handling. Handle `eligible_needs_judgment`
# exactly like needs_judgment WITH ONE EXCEPTION: never EXPIRE one -- its
# observation window is still open, and expiring it is the harm the lane exists
# to prevent. A future resolves_no_earlier_than is respected as a deliberate
# park, so nothing still parked appears in either bucket.
# READ `eligible`, NOT ONLY `overdue`: before this lane the sweep reported only
# overdue, so `overdue=0` was repeatedly recorded as "pipeline clean / stores
# healthy" while 27 eligible records sat unsurfaced (measured 2026-08-26).

# Primary source: all active hypotheses
Bash: pipeline-read.sh --stage active

# Also re-probe measurement-pending hypotheses (g-115-465 / rb-754) — these are
# shelved waiting for measurement infrastructure to appear. Re-checking the channel
# every cycle is cheap and lets them resolve as soon as data shows up. A re-probed
# measurement-pending record now past its hard resolves_by deadline with the channel
# STILL empty is EXPIRED by Step 2.6a (g-115-1584 / rb-2122) — without that pass the
# stage was a one-way shelf with no auto-expiry that grew unbounded (g-001-02 had to
# hand-archive 35 past-deadline records).
Bash: pipeline-read.sh --stage measurement-pending
APPEND to candidate list

# Horizon filter: micro-hypotheses never enter this pipeline.
# Session-horizon hypotheses use self-check verification (Step 2 handles this).
# Filter OUT any records with horizon: micro (shouldn't exist in pipeline, but defensive).
Filter out records where horizon == "micro"

Sort by resolves_no_earlier_than (soonest first), then end_date for legacy records