DEFERRED 2026-05-01T18:30 per phase-fit reversal. This memory file records the original Π.θ ★ lock decision but the lock has been DISCARDED. AI agent stack revisit was prematurely framed as Phase 3 work; correct phase is Phase 5 (IAS+IK §8 open Q’s). The original content below is preserved for forensic continuity + re-use when Phase 5 opens. Do not act on this memory’s recommendations as if locked.
Outcome: Rich-locked Π.θ ★ on 2026-05-01T09:30 BST per Rich directive (“Run any spikes… next most relevant & impactful question”).
A-numbered amendment: A-142 (Tier-1 arch-state v3.27 → v3.28).
Reverses:
- A-113 IAS substrate-component: LangGraph 1.0 self-hosted + 5-agent loop (Drafter→Verifier→Judge→Reviewer→Recorder per T39+T47)
- Q7’ F: 6-agent IK loop with Scholar-Reviewer extra
- cedar-for-agents-v0.4-only authorization stack
New locked configuration:
- Decomposition — L4M 4-stage architecture unified IAS + IK per arXiv 2511.21033 (Statute Knowledge Building → Dual Fact-and-Statute Extraction → Solver-Centric Adjudication → Judicial Rendering); empirically 0.75-0.78 precision on specific-provision legal-verdict prediction; nearly 2× precision on general-provision vs next-best.
- Framework — Anthropic Agent SDK programmatic (replaces LangGraph 1.0); A-113 sovereignty preserved (Cloudflare Workers + per-firm DO topology); NOT Anthropic Managed Agents (currently lacks VPC peering / private endpoints — Apr 8 2026 launch state).
- Distribution — 9 Anthropic Skills published per T39+T47: 4 stage-Skills (knowledge-building / extraction / adjudication / rendering) + 5 jurisdiction-Skills (catala / cedar / faraid / hsa / halakhic); cross-vendor portable (Claude Code + ChatGPT/Codex CLI + Cursor + Gemini CLI per Anthropic Agent Skills open standard Dec 2025); first-mover slot in succession-rule-authoring Skills marketplace (zero competitors per T47).
- Authorization — cedar-for-agents v0.4 + Microsoft Agent Governance Toolkit (April 2 2026; MIT-licensed; OWASP 10 agentic AI risks; deterministic sub-millisecond enforcement).
NO reversal of: A-113 sovereignty discipline (still self-hosted Cloudflare; still DO-per-firm); A-114 Cedar+Catala substrate; A-141 SSSOM verificationTier annotation; T1 MCP architecture; Cloudflare AI Gateway + Haiku 4.5 default model.
Sensitivity: 8-option scoring; Π.γ 4.96 (math-best) vs Π.θ ★ 4.92 (Rich-pick); margin 0.04 well below 0.20 threshold. 5/8 perturbations split between top-2. Status-quo Π.α 4.00 only wins under “status-quo discipline” perturbation. Π.θ wins on commercial-moat + first-mover-Skills + AI-vendor-procurement + tooling-ecosystem + acquirer-narrative.
Methodological observations:
- FIRST instance of question REPLACED mid-flight (original ζ-Q15 t-box layering → new ζ-Q15 AI agent stack); codification candidate
feedback_question_supersession_via_inline_spike. - FIRST ζ.2 multi-lock reversal in single decision (reverses A-113 IAS substrate + Q7’ F + 6-agent IK loop simultaneously); codification candidate
feedback_lock_reversal_under_external_evidence_pressure— when 4+ major external developments accrue post-lock, lock-revisit produces non-status-quo outcome at high probability. - BREAKS FIFTH-CONSECUTIVE-CORE Rich-pattern that ζ-Q14 anticipated; ζ-Q15 (REVISED) Π.θ ★ is genuine-innovation pick (richer-option-vs-leaner-option pattern), not CORE-vs-Hybrid pattern.
- Spike-driven-reframing validates “spikes must not be delayed” discipline — 30-min inline spike (3 web searches + library re-read) surfaced 4 lock-reversal candidates worth more than the original ζ-Q15.
- 16 of 17 ζ.2-related decisions remain convergent.
5 NEW richard-tasks:
- Skills publication ordering (faraid + halakhic first per T47 white-space; UK + Scots + Swiss next; HSA after; per-stage Skills parallel)
- Microsoft AGT integration spike (~2-3 days; Phase-1 sprint S2 or S3)
- L4M paper deeper read + reference implementation hunt (Phase-3 sprint S2; does open-source code exist? license? port to Catala/Cedar substrate)
- Anthropic Agent SDK durable-state evaluation (Phase-1 sprint S2; LangGraph durable state + checkpointing equivalent in Anthropic SDK?)
- Anthropic Managed Agents Phase-1.5+ re-evaluation trigger (date-triggered ~Q3 2026; if VPC peering + private endpoints land, re-evaluate Managed Agents adoption for IK MCP server only — IAS A-113 sovereignty preserved regardless)
NEW R-NEW-AI-VENDOR-LOCK risk: Anthropic Agent SDK = vendor lock to Anthropic; mitigation = Skills cross-vendor portability hedges (Skills run on OpenAI Codex / Cursor / Gemini CLI without rewrite). Risk register update pending v1.3.
Spike file: ~/off-github/library/projects/inherit/T-ai-agent-stack-spike-2026-05-01.md v1.0 — 5 web-research targets + library re-read of Biswas-Talukdar + Gulli + 5 T-files cross-checked.
Cascade artefacts: original ζ-Q15 cascade file moved current-questions/ → _deferred-questions/ with deferral metadata; new ζ-Q15 (REVISED) ANSWERED file in answered-questions/; arch-state v3.28 A-142; IK arch-state v1.2; IAS arch-state v1.3; INDEX v2.18; this memory file; active-work-log update.
Phase 3 status: 7 Q asked + answered + cascaded (ζ-Q9..ζ-Q15-REVISED); 1 Q deferred (ζ-Q15-original t-box layering → Phase 5); ζ-Q16+ candidates surfacing.
NEXT: ζ-Q16 candidate = partner-firm legacy data migration architecture (zero prior coverage gap; affects MLP P-ζ M16-M18 deliverable; structural complement to Π.θ which addresses NEW authoring not legacy migration).