TT Claude Memory

project_zeta_q2_xi_eta_g_plus_phase_1_graphrag_locked_2026_05_02

Properties1
descriptionSecond Q re-asked under refined-prompt v3.0.1 post-Full-Reversion. A-127 (re-issued) registered in arch-state v3.12. ξ.η = G+ commercial-discipline preserved (narrow Catala E&W IHT+intestacy+attestation; Cedar v4.10; LinkML 1.10; Rust/Go SDK) + graph-RAG committed Phase-1 (Postgres + pgvector + hybrid retrieval). 7 options scored under κ.θ-LITE 12 architecture criteria. Margin 0.05 baseline / 0.10 post-WR-Finding-1-adjustment sensitivity-borderline; 4 perturbations favour ξ.ε G+ / 3 favour ξ.η / 1 ties. Rich-divergence from `feedback_v2_ik_ias_build_plan_sovereign` commercial-velocity-priority — chose graph-RAG-from-Day-1 per WR Finding 1 (Techment Hybrid RAG production-table-stakes 2026). Phase-1 re-baseline 5mo→6mo / £75K→£90K / £20K→£24-28K Y1. R-042 MITIGATED; R-NEW-PHASE-1-GRAPHRAG-EXECUTION ADDED (33→34). 4 NEW richard-tasks #209-#212.

5 min read

Lock: ξ.η G+ Commercial-scoped + Phase-1-graphRAG synthesis

Locked at: 2026-05-02T03:50 BST per Rich confirmation

Margin: ξ.η 4.55 vs ξ.ε G+ baseline 4.60 = -0.05 baseline / +0.10 post-Finding-1-adjustment — sensitivity-borderline; below 0.20 threshold; Rich-divergence from feedback_v2_ik_ias_build_plan_sovereign commercial-velocity-priority pattern

7 options scored under κ.θ-LITE 12 architecture criteria (per ζ-Q1 A-126 frame):

  • ξ.α B JSON Schema enhanced 2.90 (devil’s advocate light)
  • ξ.β C Semantic Web purist 2.85 (devil’s advocate full; convergently disconfirmed across 1733 pages of 2025 specialist literature per scorecard §F)
  • ξ.γ D Pragmatic hybrid KG-canonical 3.85
  • ξ.δ F Maximal F v1.3 4.20
  • ξ.ε G+ Commercial-scoped + Year-2+ uplift 4.60 (numeric leader; PRE-REFINED-PROMPT pick reference; strong fallback)
  • ξ.ζ Graph-RAG-Native 3.35 (NEW 6th option per WR Finding 1)
  • ξ.η ★ G+ + Phase-1-graphRAG synthesis 4.55 Rich-locked (NEW 7th option; synthesis combining G+ commercial-discipline + WR Finding 1 production-table-stakes evidence)

Locked ξ.η commits the architecture to:

  1. G+ commercial-discipline elements PRESERVED — narrow Catala scope (IHT + intestacy + attestation E&W only Phase-1) + Cedar v4.10 + LinkML 1.10 + Rust/Go SDK + LLM narrow-adjunct (Cloudflare AI Gateway + Haiku 4.5 default; £1,500/mo ceiling per feedback_phase1_llm_scenario2_lock)
  2. Graph-RAG committed Phase-1: Postgres + pgvector + hybrid retrieval (semantic + lexical) over corpus + extracted-rule-text. NOT graph-RAG over Catala AST (out-of-scope per Q-002 §9.2 pre-mortem mitigation 2). NOT Loro CRDT or Loro-graphs Phase-1.
  3. Year-2+ uplift roadmap UPDATED: Alloy / TLA+ / Carneades / LUIMA / cross-tradition expansion remain deferred Year-2+; graph-RAG NO LONGER deferred (pulled forward to Phase-1)
  4. Phase-1 timeline re-baseline: 5 mo (G+ baseline) → 6 mo (G+ + Phase-1-graphRAG)
  5. Phase-1 build cost re-baseline: £75K (G+ baseline) → £90K (G+ + Phase-1-graphRAG; +£15K + 1mo Paul-time-equivalent)
  6. Y1 post-launch run-cost re-baseline: £20K/yr → £24-28K/yr (+£500/mo graph-RAG inference)
  7. 5-year cumulative cost (acquirer-ready scenario): £215K (G+ Y2-uplift-needed) → £245K (ξ.η no-Y2-graph-RAG-uplift-needed)

Cost-adjusted differential ξ.η-vs-ξ.ε: +£55K cumulative 5y baseline → +£30K acquirer-ready (after R4 retrofit-avoidance) → net ~£15-30K cumulative 5y = ~£3-6K/yr for acquirer-DD-parity-from-Day-1 instead of Year-2-retrofit-bet.

WR Finding 1 grounding (Techment 10-RAG-architectures-2026): “Hybrid RAG consistently outperforms vector-only because it captures both semantic meaning and exact terminology — critical for legal/financial/regulatory domains.” Production-table-stakes for legal-tech 2026; pressures graph-RAG inclusion in Phase-1 vs deferred Year-2+ uplift.

R-042 acquirer-narrative-ceiling risk MITIGATED under ξ.η — was added under PRE-REFINED-PROMPT G+ permanent; now downgraded (graph-RAG-from-Day-1 closes “where’s your knowledge graph?” DD question per stakeholder-sanity-check §9.4 acquirer-DD row). Demonstrates refined-prompt evidence-strengthening can DOWNGRADE risks added in prior locks.

R-NEW-PHASE-1-GRAPHRAG-EXECUTION risk ADDED (33→34): 30% probability per pre-mortem #1 (Phase-1 graph-RAG add slips Q4-2027 ship by >2mo); mitigation = time-box graph-RAG Phase-1 to vector-store + Postgres-pgvector + simple-hybrid retrieval; fallback to ξ.ε G+ Year-2-uplift if time-box exceeded by 50%.

8 sensitivity perturbations: 4 favour ξ.ε G+ baseline (#1 commercial-velocity ↑; #5 pure timeline-priority; #6 drop graph-RAG criterion; #8 κ.γ fallback frame) / 3 favour ξ.η (#3 acquirer-narrative ↑; #4 WR Finding 1 evidence honoured; #7 add Hybrid-RAG-table-stakes criterion) / 1 ties ξ.η/ξ.δ (#2 AI-vendor attestability ↑). No clean winner under sensitivity testing; decision pivots on whether WR Finding 1 is treated as material evidence (favours ξ.η) or as forward-looking trend (preserves ξ.ε G+ baseline). Rich CHOSE material-evidence interpretation.

**4 NEW richard-tasks 209-212**:

  • #209 Graph-RAG Phase-1 substrate spike (1-week time-box per pre-mortem #1)
  • #210 Year-2+ uplift roadmap update in option-G.md v1.2 → v1.3 (graph-RAG removed from deferred list)
  • #211 Phase-1 build cost re-baseline £75K → £90K + Path A funding capacity re-validation with Jana
  • #212 R-042 acquirer-narrative-ceiling risk register update (probability + impact reduced; status mitigated-via-A-127)

Methodological observations:

  1. SECOND Q under refined-prompt v3.0.1 — 10-step protocol executed cleanly again
  2. First Rich-divergence under refined prompt — Claude pick ξ.η matched lock; PRE-REFINED-PROMPT had Rich-divergence G+→F (margin 0.43); refined prompt had no divergence (Claude pick = ξ.η; Rich confirmed ξ.η)
  3. Cost-row enhancement Rich-requested AND Rich-acted-on: Rich saw cost-adjusted differential ~£15-30K cumulative 5y / ~£3-6K/yr and chose acquirer-DD-parity over commercial-velocity-priority — first time refined prompt’s evidence-strengthening overrode feedback_v2_ik_ias_build_plan_sovereign Rich-pattern
  4. Codification candidate: feedback_refined_prompt_evidence_overrides_rich_pattern_when_acquirer_DD_at_stake — when refined-prompt brings concrete cost-data + 2026-current production-table-stakes evidence + R-042 mitigation argument together, Rich-pattern of commercial-velocity-priority can be overridden in favour of acquirer-DD-parity-from-Day-1
  5. Sensitivity-borderline 0.05-0.10 margin honestly flagged + Rich-override capability preserved; Rich CHOSE Claude pick rather than override — pattern observation: refined prompt successfully aligned recommendation with Rich-judgement on this Q despite sensitivity-borderline
  6. R-042 mitigation via A-127 reverses prior PRE-REFINED-PROMPT-era risk addition — demonstrates how refined-prompt evidence-strengthening can DOWNGRADE risks added in prior locks
  7. Convergent decision pattern continuing: 2 of 2 refined-prompt re-ask decisions chose synthesis options frontloading rigour over lighter alternatives (κ.θ utility-tree synthesis at ζ-Q1; ξ.η G+ + Phase-1-graphRAG synthesis at ζ-Q2). Pattern emerging.

NO REVERSAL of: A-126 κ.θ frame (ξ.η scored under κ.θ-LITE); A-1..A-125 amendment registry; MQ-000 ζ.2 scope; A-113 sovereignty discipline (DO-per-firm topology preserved); A-114 Cedar+Catala substrate; A-125 Phase-1.5 21-jurisdiction expansion (orthogonal to ξ-selection); brand-architecture v1.16; master plan v1.10 (will re-synthesise at Phase-5 closure); per-repo BUILD-PLAN.md v1.4-1.6 (will refresh at Phase-5 closure); 2026-04-30-architecture-rederivation-design v1.1 §16; refined-prompt v3.0.1; feedback_phase1_llm_scenario2_lock.

NO new SKOS classifier scheme; NO new A-21 CI gate; CI gate count UNCHANGED at 19 (graph-RAG-related gates may emerge at Phase-3+ module-authoring time).

NO cross-module primitive count change (still 27 = 25 §2.2 + 2 §2.3 pending). NO module structure change (still 9). Phase-1.5 stress-test cell count UNCHANGED (21-jurisdiction expansion per A-125; orthogonal to architecture-option).

Phase 2 status post-lock: 2 Q answered/cascaded (ζ-Q1 + ζ-Q2) / 18-30 budget / 7-11% complete; pre-build-start cumulative 2/100-150.

NEXT: ζ-Q3 = Phase 2 — Shape X angle (i) asset taxonomy. Per audit-record §5 re-ask sequence S-3. PRE-REFINED-PROMPT pick was ε with 9 i-ζ classes (BFO 2020 Continuant rooting); fresh re-derivation under refined prompt; uses κ.θ-LITE 12-criterion frame projected onto top-level-asset-classification surface.

Graph View

Backlinks