refined-prompt v3.13 → v3.14 released — Wide bundle 5 improvements (C1+C2+C3+F4+F6)
Released: 2026-05-17T17:50 BST (directive lastmod). Commit 077b5e1 on docs-strategy origin/main. Pin-drift hook PASSED on first commit attempt — C3 YAML-safety discipline applied during authoring.
Trigger
Q-027 ξ.η’★ first v3.13 production use 2026-05-17 surfaced 3 directive gaps that v3.13 step 7 per-flavour audit + step 7.5 named-persona pass did NOT catch at criterion-set level — caught only via Rich’s explicit critique (“INHERIT is an international standard”) + Rich’s explicit meta-scorecard request comparing Lock-A / Lock-B / Lock-C / Lock-D options. Plus F4 + F6 carry-over from v3.12 (Q-022 review 2026-05-05) deferral. Rich-locked wide-bundle scope 2026-05-17T17:35 BST after surfacing options.
What shipped — 5 additions
C1 — Step 7 CASCADE-COMPLIANCE SUB-AUDIT (UNIVERSAL across all Qs)
New meta-scorecard criterion c-cascade-compliance 0-5 scored on whether candidate option’s sub_clarifications_locked implicitly contradict any prior-locked Q’s sub_clarification_locked. Cascade-violations LOCK-BLOCKING — two resolution paths:
- (a) revise candidate’s sub_clarifications to honour upstream cascade (typical)
- (b) author Q-X.Y amendment FIRST per Lock-C path, then lock candidate as next sequential A-N
5/5 = no implicit upstream amendment; 0/5 = silent upstream amendment required.
Closes: Q-027 meta-scorecard pattern where Lock-B (override → ξ.ζ’ dropping faith hook Day-1) would have silently violated Q-024 Φ.η cascade (which locked faith-advisory-layered NULLABLE Day-1 per A-115). Without explicit meta-scorecard, Lock-B would have scored higher numeric Wsum (4.94 vs 4.82) and been recommended.
C2 — Step 7 INTERNATIONAL-SCOPE SUB-AUDIT (inside v3.13 CORRECTNESS-FLAVOUR AUDIT for FR flavour)
For cross-jurisdiction modules (8 of 9 v2 modules — Wills / Probate / Trusts / Assets / Commerce / Transfer / Delegation / Credentials; Core may not load), EVERY faithful-representation-of-law criterion MUST name the international scope, NOT a single national statute.
Wrong: Wills Act 1837 5-class typology fidelity
Right: Cross-jurisdictional typology fidelity (UK&W + Scotland T&S(S)A 2024 + Switzerland CC/CO + faith pillars Phase-1.5+ per Q-024 Φ.η)
If a candidate criterion names ONLY a single national statute, REWRITE before scoring.
Closes: Q-027 first v3.13 production use E&W-creep recurrence. v3.13 step 7.5 named-persona pass DID catch the gap at persona-projection-time (persona P5 Jurisdiction-fidelity (c)-class objection) but step 7 per-flavour audit didn’t flag the criterion-set itself.
C3 — Step 1 SAVE+COMMIT LASTMOD-PROSE YAML-SAFETY DISCIPLINE
When authoring arch-state amendment narrative inside the lastmod scalar value, NEVER use : (colon-space) as a prose separator. Use — (space em-dash space) instead.
Reason — lastmod scalar is unquoted YAML value; inner : triggers yaml.safe_load: mapping values are not allowed here and pre-commit pin-drift hook BLOCKS commit. Timestamp colons (e.g. 17:30 BST) safe; breaker is colon-FOLLOWED-by-space anywhere in semantic prose.
Closes: Q-027 ξ.η’★ lock 2026-05-17 hit pin-drift hook block twice; comprehensive fix replaced 14 instances. v3.14 prevents recurrence by codifying the discipline at authoring time.
F4 — Step 7 CLOSING-QUESTION BOUNDARY TEST (deferred from v3.12)
Folds project-CLAUDE.md discipline locked 2026-04-27 into directive body. Before authoring any closing-question alternative:
- For each alternative about to be named, identify the topic-keyword
- Run
grep -liagainst T-files + spike files + indexed/ library - If matched, extract actually-evaluated alternatives + cite verbatim
- If no match, tag as “speculative — not yet researched; do not select without first running a spike”
- NEVER present prior-knowledge alternatives as if they were research-grounded
Closes: 2026-04-27 anti-pattern (Hetzner/OVH speculative naming when T54 actually evaluated cloud-iam.com + loginfactor.com). Discipline was in project-CLAUDE.md but not folded into directive — each Q-cycle relied on memory rather than directive enforcement.
F6 — Step 6h RESEARCH-COMPLETENESS SELF-SCORE (deferred from v3.12)
Mandatory ≤5 min before emitting Research Sweep Summary. 6 criteria:
| ID | Criterion | Reference |
|---|---|---|
| RC1 | Tier-2 pgvector query run + top hits surfaced | v3.9 / v3.11 |
| RC2 | ≥3-5 parallel web tracks run (5-7 for foundational substrate-architectural) | v3.11 |
| RC3 | ≥1 T-file per topic-mapping table | [[feedback_check_t_files_first_for_any_inherit_v2_work]] |
| RC4 | Substrate-completeness pre-flight gate explicit | v3.12 #1 |
| RC5 | Maturity vocabulary 3 sub-modes applied (STABLE / CANDIDATE / TRANSIENT-INFERENCE-NOTE) | v3.12 #5 |
| RC6 | Jurisdiction-symmetric substrate (≥1 substrate evidence per Phase-1 jurisdiction) | Q-015 λ.η + Q-026.1 + [[feedback_e_w_creep_recurrence_q027_first_v3_13_production_use]] |
Each 0/3/5 (not done / partial / full). <3/5 on any criterion flags research-thin → extend-or-spike before locking.
Closes: Q-027 first v3.13 production use research-thin RC6 defect (4 of 5 web tracks UK-only). RC6 ≥3/5 threshold check pre-step-7 would have caught it at first-pass.
Diagnostic significance
Q-027 first v3.13 production use revealed that two NEW v3.13 mechanisms (step 7 per-flavour audit + step 7.5 named-persona pass) caught the E&W-creep at persona-projection-time but NOT at criterion-authoring-time. v3.14 closes that 1-pass-removed gap.
Meta-pattern: directive amendments come in pairs across versions, with downstream amendments closing gaps the prior version’s upstream amendments couldn’t catch. v3.14 is the v3.13-gap-closer; predict v3.15 to be the v3.14-gap-closer when a new failure mode surfaces.
Cost-benefit
| Addition | Time-cost per Q | Catches |
|---|---|---|
| C1 cascade-compliance scan | 2-4 min | Silent upstream-Q cascade-violations (Q-027 Lock-B pattern) |
| C2 international-scope sub-audit | 1-2 min | National-statute-anchored FR criteria (E&W-creep recurrence) |
| C3 YAML-safety discipline | 0 per-Q (authoring discipline only) | Pin-drift hook commit blocks (Q-027 lock 14-defect fix) |
| F4 closing-question boundary test | 2-4 min | Speculative alternatives presented as research-grounded |
| F6 research-completeness self-score | 3-5 min | Research-thin sweeps escaping to step 7 |
Total cost ~+5-15 min per Q. Acceptable per Rich’s preference for thorough discipline over speed.
Scope + reversal
- No reversal of v3.0..v3.13 locks
- Q-001..Q-027 unaffected — v3.14 affects Q-028 onwards (first v3.14 production use Q-028 Phase-3 ζ-Q3 Wills.SubstitutionClause class shape MVP per A-124)
- Step 6 sub-letters 6a-6g unchanged; 6h NEW
- Step 7 framing structure unchanged; per-flavour audit gets 2 new sub-audits; closing-question gets boundary test; step 7.5 named-persona pass unchanged
- Step 8 DECISION QUALITY CHECK unchanged
- Step 1 SAVE+COMMIT gets C3 sub-rule
First production use
Q-028 = Phase-3 ζ-Q3 Wills.SubstitutionClause class shape MVP per A-124 (third and final foundational rule for Wills module Phase-α MVP; substrate-readiness HIGH from Q-026 ψ.η + Q-027 ξ.η’★ patterns).
Substrate documents
- Directive:
~/testatetech/docs-strategy/docs/superpowers/specs/2026-04-29-multi-phase-audit/refined-end-of-turn-directive.mdv3.14 (commit077b5e1) - Companion v3.13 release:
[[refined_prompt_v3_13_released_named_persona_review_pass_2026_05_17]] - E&W-creep drift codification:
[[feedback_e_w_creep_recurrence_q027_first_v3_13_production_use]] - T-file citation discipline (F4 substrate):
[[feedback_actively_use_t_files_in_scorecard_authoring]] - Cascade-discipline forerunner: A-124 anti-E&W-creep + Q-024 Φ.η Day-1 commitment + Q-015 λ.η 3-jurisdiction Phase-1 lock
- v3.12 deferral source:
[[feedback_v3_12_release_9_improvements_2026_05_05]] - Q-027 lock memory:
[[project_zeta_q27_locked_xi_eta_prime_wills_bequest_class_shape_mvp_2026_05_17]]