| name | intent-ranked-regression-craft |
| description | >- Use when this capability is needed. |
Intent-Ranked Regression Craft
Minted 2026-06-05 from the operator's framing: "not mechanical — take all my
intents, use cases, logic, scenarios, interactions (past/present/future), make a
ranking system, research inside and outside, review it, then regress." This is
the value layer above the mechanical inter-wave regression (which checks is
everything wired?); this layer asks is what I care about most still served?
Why this exists (vs the mechanical sweep)
akos-inter-wave-regression.mdc runs 13 structural dimensions at ~equal weight.
That answers integrity, not priority. When attention is scarce (operator review
minutes; agent context budget), you want the highest-stakes surfaces checked first
and deepest. External practice converges here:
- FMEA / RPN (Jama, SixSigma, testRigor 2026): rank failure modes by
Severity × Occurrence × Detection; severity-first override — a catastrophic
surface is actioned regardless of composite score.
- WSJF (SAFe / Reinertsen): Cost of Delay = Business-Value + Time-Criticality +
Risk-Reduction; the right altitude for an initiative-portfolio regression.
- TIA + Predictive Test Selection (Microsoft, Meta, Fowler): bound the candidate
set by what changed, rank by fault-likelihood, keep the full sweep as a periodic
safety net (never delete it).
The model (SSOT in code)
akos/hlk_intent_ranked_regression.py — 7 intent tiers + N surfaces + ICS.
scripts/intent_ranked_regression.py --rank prints the ranked sweep order.
ICS = 3*intent_value + 2*time_criticality + 2*risk_reduction + 1*detection_gap
intent_value (1-5) = max value of the intent tiers the surface serves.
time_criticality (1-5) = does delay cost? (before first sale / incorporation).
risk_reduction (1-5) = blast radius if it breaks.
detection_gap (1-5) = FMEA Detection inverted — 5 means no always-on gate
would catch a regression here (so it deserves manual eyes); 1 means a strong gate
already does.
severity_first = FMEA Action-Priority override: existence-critical surfaces
(legal/fiscal IT-2, governance-integrity IT-3) with a known live gap lead the
queue regardless of ICS.
The loop (both seats)
- Distil the intent corpus (thinking seat): sweep operator-scratchpad,
RICE-ranked
OPERATOR_INBOX.md, USE_CASE_ARCHIVE.csv, open OPS_REGISTER
severities, INITIATIVE_REGISTRY. Update tier weights only with evidence — never
invent priority. Re-derive when the portfolio shifts materially.
- Score / re-score surfaces (either seat): edit the surfaces in the SSOT module;
--self-test must stay green.
- Run in ICS order (execution seat):
--rank gives the ordered probe list. Run
top-down; stop-early is allowed only when remaining surfaces are all low-ICS AND
green so far.
- Attribute every finding (thinking seat) — the load-bearing craft move:
- new = this session's diff caused it → rework-now.
- pre-existing = drift that predates this session → fix-now only if high-ICS +
low-effort + reversible (FMEA severity-first); else defer-OPS with owner.
- known-deferred = already has an OPS row + ETA → cite it, do not re-litigate.
- Disposition via the inter-wave 5-option enum (rework-now / forward-charter /
defer-OPS / accept-as-canon / escalate-to-blocker). Never ad-hoc.
- Mint findings + (if high value) harden the method so the next regression is
cheaper — that is this discipline eating its own tail.
Anti-patterns
- AP1 — Flat failure lists. A release-gate "9 failed" is not a regression result;
rank + attribute the 9, or you have only noise.
- AP2 — Blaming your own diff for inherited drift (or the reverse). Attribution
is mandatory; it is the difference between "I regressed X" and "X was already red".
- AP3 — Deleting the safety net. The ranked sweep supplements the full
mechanical sweep / release-gate; it never replaces it (TIA safety-fallback).
- AP4 — Inventing intent weights. Every tier value cites scratchpad / inbox /
OPS / registry evidence. No evidence, no weight change.
- AP5 — RPN-only blindness. Honour severity-first: an existence-critical surface
leads even if its composite ICS is not the maximum.
Worked example
docs/wip/planning/88-cross-area-ops-wiring-review-discipline/reports/intent-ranked-regression-2026-06-05.md
— the 2026-06-05 run: 3552 pass / 9 fail, attributed to 0 new regressions,
1 high-intent pre-existing drift fixed-now (the area_governance pattern-class
enum lag from I93), 8 known/deferred dispositioned.
Cross-references
- SSOT model:
akos/hlk_intent_ranked_regression.py
- Runbook:
scripts/intent_ranked_regression.py
- Draft rule:
.cursor/rules/akos-intent-ranked-regression.mdc
- Sister discipline (the safety net):
.cursor/rules/akos-inter-wave-regression.mdc
.cursor/skills/inter-wave-regression-craft/SKILL.md
- Disposition enum:
.cursor/rules/akos-inline-ratification.mdc
Source: FraysaXII/openclaw-akos — distributed by TomeVault.