| name | inter-wave-regression-craft |
| description | Use when running, dispositioning findings from, or wiring up the inter-wave regression sweep in this AKOS workspace. Codifies the craft for keeping the 13-dimension sweep healthy + the 5-option disposition enum + heuristic-evolution patterns. Wave R DIM-02 false-positive recovery is the worked example. Triggers on inter-wave regression, regression sweep, validate_inter_wave_regression.py, inter_wave_regression_sweep.py, DIM-NN finding, wave-close cadence, RegressionFindingRow, inter-wave drift. Pairs with .cursor/rules/akos-inter-wave-regression.mdc (the WHEN); this skill is the HOW. |
| metadata | {"author":"FraysaXII"} |
Inter-Wave Regression Craft
Codified at I86 Wave R close (drain7) from the operator's framing: regression sweeps need to be dispositioned thoughtfully, not mechanically. The craft turns the 13-dimension sweep from a noise generator into a structural-integrity check. Ratified by D-IH-86-CT alongside the parent rule akos-inter-wave-regression.mdc (D-IH-86-BK).
When to use this skill
Read this skill before:
- Running
py scripts/inter_wave_regression_sweep.py --wave-closing <wave-letter> at any in-scope wave close.
- Dispositioning any WARN or FAIL finding from a regression sweep via inline-ratify.
- Patching the sweep heuristic itself (any of the 13
_probe_dimension_N_* functions in scripts/inter_wave_regression_sweep.py).
- Authoring a UAT closure report's §3.3 (regression-class mechanical evidence).
This skill assumes you have already read akos-inter-wave-regression.mdc, which defines what the sweep is + when it fires. This skill is the craft layer on top of that rule.
Core principles
Principle 1 — Run the self-test before the full sweep
py scripts/inter_wave_regression_sweep.py --self-test exercises all 13 probe functions against a tiny synthetic fixture set and verifies they return well-formed RegressionFindingRow Pydantic models. If the self-test fails, the sweep itself is broken — fix the runbook before running the real sweep. The full sweep against the entire workspace takes 30-90 seconds; debugging a broken runbook by running the full sweep wastes that time per iteration.
Principle 2 — Inventory expected findings count BEFORE drafting the disposition gate
Before invoking the AskQuestion batch to disposition findings, compute the findings count from the JSON sidecar:
import json
findings = json.load(open("artifacts/regression-sweep-<date>-wave-<letter>.json"))["findings"]
print(f"PASS={sum(1 for f in findings if f['verdict']=='PASS')}; WARN={sum(1 for f in findings if f['verdict']=='WARN')}; FAIL={sum(1 for f in findings if f['verdict']=='FAIL')}")
Findings count drives the disposition strategy:
- ≤ 10 findings: fits in one
AskQuestion batch (up to 10 questions).
- 11-20 findings: splits into 2 batches (group by dimension; cite cross-references inline).
- > 20 findings: triggers the bandwidth-recovery pattern — halt + propose splitting the wave into Wave-N + Wave-N.5 (the dispositioning load is too high to land in one wave-close commit).
Principle 3 — Distinguish true gaps from heuristic false-positives BEFORE filing the disposition
The 13 probes carry heuristics that can produce false positives (the probe flags a gap that isn't actually a gap because the heuristic doesn't see the evidence the canonical commits to). The Wave R precedent is concrete: _probe_dimension_2_forward_charter_carryover flagged 10 forward_charters: entries as gaps because its token-extraction logic missed evidence in forward_charters_pending and linked_canonicals frontmatter. Spot the false positives by:
- Re-running with
--verbose to see the per-finding evidence trail.
- Spot-checking 3-5 findings against the actual canonical body — does the canonical genuinely lack the forward-charter resolution, or does the heuristic miss the resolution?
- If false-positive rate > 20 % of findings, halt and patch the heuristic before filing dispositions. Filing dispositions on noise wastes operator cycles.
The Wave R recovery shrunk DIM-02 false-positives from 10 to 4 by:
- Expanding evidence sources from "decision-log files only" to "filesystem walk + canonical body content + frontmatter sub-keys."
- Normalising token comparison to alphanumeric-only (
_alnum) so forward_charters: I-NN-... matches linked_canonicals: I_NN_....
- Adding stop-prefix filter (
OPS-, D-IH-, R-IH-) so non-initiative IDs don't get treated as forward-charter targets.
The patch is in scripts/inter_wave_regression_sweep.py lines ~280-360.
Principle 4 — Disposition via the 5-option enum, never ad-hoc
Per the parent rule §RULE 2, every WARN/FAIL finding gets dispositioned via exactly one of:
rework-now (FAIL on baseline dimensions; the finding must be fixed in this commit).forward-charter-next-wave (WARN on conditional dimensions; mint an OPS row or candidate file scoped to the next wave).defer-OPS (INFO; mint OPS row but don't gate on it).accept-as-canon (operator ratifies the drift was deliberate; append contra-precedent + decision-register row).escalate-to-blocker-tracker (mint _blockers/<slug>-tracker.md per akos-conflict-surfacing-and-blocker-trackers.mdc).
Ad-hoc dispositions ("we'll address it later" without OPS row OR decision row) violate the rule and create silent governance debt.
Principle 5 — Heuristic-evolution is a craft, not a bug-fix
When a probe heuristic produces false positives at rate > 20 %, the response is NOT just to file dispositions and move on. Patch the heuristic in the same commit (or the next commit in the same wave) so the next sweep doesn't repeat the noise. The patch must:
- Expand evidence sources (filesystem + canonical body + frontmatter sub-keys, not just decision-log).
- Improve token normalisation (alphanumeric-only comparison, stop-prefix filters, case-insensitive matching).
- Add unit tests in
tests/test_inter_wave_regression.py to lock in the heuristic against regression.
- Document the heuristic evolution in the parent canonical's evidence-base section.
The Wave R DIM-02 heuristic patch is the worked example: 4 commits across the wave (sweep run → false-positive analysis → heuristic patch → tests + re-sweep) landed as a clean tightening of the rule.
Per-dimension craft (13 dimensions, key gotchas)
| Dim | Probe focus | Common false-positive | Disposition default |
|---|
| 1 | decision_lineage | FK fails because decision row not yet appended (race condition between OPS row + DECISION row landing) | rework-now (add the missing decision row first) |
| 2 | forward_charter_carryover | Token-extraction misses evidence in linked_canonicals or forward_charters_pending | re-run after heuristic patch; if true gap, forward-charter-next-wave |
| 3 | validator_ramp_consistency | Threshold change without paired decision (often legitimate at INFO ramp; document in decision row) | accept-as-canon (after decision row) |
| 4 | canonical_csv_pair_completeness | Pydantic module missing OR Supabase mirror missing | rework-now (mints fail without full quartet) |
| 5 | sop_runbook_pairing | SOP exists but linked_runbooks: not populated in frontmatter | rework-now (1-line frontmatter edit) |
| 6 | uat_report_class_completeness | UAT report missing a §-class that compose_UAT requires for the initiative's audience tags | rework-now if class is mandatory; forward-charter if optional |
| 7 | render_trail_audience_match | External-tagged surface lacks paired render artifact | re-run validate_external_render_trail.py --strict; rework-now if confirmed |
| 8 | brand_baseline_register_match | CORPINT-internal token leaked into external surface | rework-now (translation back to external register) |
| 9 | cross_area_breakthrough_announcement | New PEOPLE_DESIGN_PATTERN_REGISTRY row without paired announcement evidence |
Pre-flight checklist (before each sweep)
- Self-test PASSES (
py scripts/inter_wave_regression_sweep.py --self-test).
- Wave-letter argument is correct (
--wave-closing R not --wave-closing Q).
- Wave is in-scope per parent rule §"When this rule applies" (≥ 2 canonicals OR ≥ 2 cursor rules OR canonical-CSV mint).
- Report path follows convention (
reports/regression-sweep-<YYYY-MM-DD>-wave-<letter>-close.md).
- JSON sidecar path is conventional (
artifacts/regression-sweep-<YYYY-MM-DD>-wave-<letter>.json).
- Findings count computed BEFORE drafting disposition gates.
- False-positive spot-check (3-5 findings inspected against canonical body) completed.
- Disposition enum (5 options) is the only mode used; no ad-hoc text.
- Operator-scratchpad will be drained in the same wave-close commit (DIM-12 PASS guaranteed).
- UAT report §3.3 reserves a row to cite this sweep + its disposition outcomes.
Anti-patterns (don't do these)
- AP1 — Sweep then move on without disposition. Findings without dispositions = silent governance debt; never lands.
- AP2 — Dispose by ad-hoc note. "Will fix later" without OPS or DECISION row = breaks rule + audit trail.
- AP3 — Skip the false-positive spot-check. Filing 20 dispositions on noise wastes operator cycles + creates fatigue.
- AP4 — Patch the heuristic in a separate wave. Heuristic + dispositions belong in the same wave so the next sweep is clean.
- AP5 — Treat all 13 dimensions as equal-weight. Baseline 7 always fire; conditional 6 fire only when applicable; ignoring the rule wastes time.
Recovery patterns
- R1 — Too many findings (>20): halt + propose Wave-N + Wave-N.5 split; don't try to disposition all in one shot.
- R2 — Heuristic broken: write a failing unit test that reproduces the false positive; patch the probe; re-run sweep; verify tests pass.
- R3 — Operator silent on dispositions: per
akos-inline-ratification.mdc Time-box recovery, auto-default after 24h+ silence only for reversible options (defer-OPS, forward-charter); never for accept-as-canon (irreversible) or escalate-to-blocker (high-blast).
Cross-references
Source: FraysaXII/openclaw-akos — distributed by TomeVault.
