project_round_5_gold_2026_04_25

Round 5 — Gold-hunting deep-dives complete (Saturday 25 April 2026)

After 4 rounds of T-programme research + spike-replacement, Rich commissioned a final batch with directive: “dig deep and find ways to make our eventual solution ultra modern and utilise all the best tools.” 7 parallel subagents dispatched (T35-T41); ~28,100 words; ~£100-150 cost; 48 new spike candidates surfaced + 9 strategic findings + 5 standards-engagement chess-moves.

9 strategic findings worth surviving compaction

1. Cedar cedar-for-agents v0.4.0 — rails already shipped (T35)

December 2025, Apache-2.0, primarily Rust. Three production crates: mcp-tools-sdk (parse MCP tool descriptions) + cedar-policy-mcp-schema-generator (auto-derive Cedar Schema from any MCP server) + cedar-analysis-mcp-server (expose Cedar’s analysis incl. SymCC AS an MCP tool to agents). mcp.inherit.com exposing cedar.evaluate() + cedar.analyse() + cedar.derive_counterexample() becomes a 3-day prototype, not months.

2. Cedar Symbolic Compiler cedar-policy-symcc is production-grade (T35)

Not research — shipped crate on crates.io with cvc5 SMT backend. Six verifiable properties (never-errors / always-allows / always-denies / subsumption / equivalence / disjointness) + concrete-counterexample synthesis on failure. Compiler itself verified in Lean. AI-vendor pitch + ISO/JTC 1 PAS narrative differentiator.

3. Cedar legal-tech first-mover CONFIRMED (T35)

Exhaustive search returned ZERO hits for Cedar adoption in estate / probate / fiduciary / legal-tech. INHERIT can claim “first formally-verified authorisation engine in fiduciary law” — defensible for 6-12 months.

4. Catala faith-tradition first-mover CONFIRMED (T36)

Multi-axis search returned ZERO prior art on Catala for Islamic faraid / Hindu HSA-1956 / halakhic distribution. INHERIT v2 Phase-1.5 faith-tradition Catala scope-suite is genuinely first-mover — publishable at ProLaLa / NLLP / Stanford CodeX FutureLaw / IETF / OASIS.

5. Catala competitive landscape COLLAPSED 2024-2025 (T36)

SMU CCLAW S$15M wound down 2025 (research handed to Legalese spinout); Blawx self-described “not production-quality”; Hammurabi dormant; OpenFisca lacks default-logic+IDE; Logical English academic-only. Catala is the ONLY entrant with all required properties (verification + IDE + Apache + production-track + LLM benchmark + active 2026 development). Competitive-risk LOWERED to LOW.

6. W3C Web of Things adopted LinkML as source-of-truth (T37)

W3C minutes 25 Sep 2024 confirm WoT WG selected LinkML “for having the best feature set.” WoT Thing Description 2.0 reached W3C Recommendation June 2025. First-class W3C-REC precedent — INHERIT-via-LinkML to ISO is well-trodden, not a leap. GigaScience canonical paper landed late 2025 (Hegde+Matentzoglu+Mungall, DOI 10.1093/gigascience/giaf152) gives ISO-credible peer-reviewed foundation.

7. InheritKit Authoring Studio £25-40m UK Y3 ARR opportunity — 18-month window (T39)

Verdict: PARTIAL leapfrog REAL but conditional. All component pieces exist (L4M pipeline 94.1% Valid Ratio + Defeasible Deontic Logic + LLMs + Anthropic Skills open standard Dec 2025 + TLAi+Bench + xpSHACL + text2shacl + Cedar VS Code + Catala LSP). Nobody has assembled them for succession+tax+faith. 18-month window before Anthropic / Harvey / Spellbook / CodeX spin-out fills it.

Conditional on shipping: (a) verification gates (Catala compiler + SymPy + Z3/SMT + LeMAJ cross-model judge + solicitor sign-off) / (b) Anthropic Skills marketplace publication (inheritkit-catala-skill, -cedar-skill, -faraid-skill, -hsa-skill, -halakhic-skill) / (c) InheritKitBench (TLAi+Bench-pattern benchmark) / (d) domain-specific skills NOT single universal skill.

Commercial sizing UK Year-3 conservative:

• Pro £499/seat × ~500 firms × 5 seats = £15m ARR
• Enterprise £4,999/firm × top-200 = £12m ARR
• AI-Vendor licensing £1-5m
• Total £25-40m UK ARR before international expansion

8. XTDB v2 immutable bitemporal SQL + crypto-shredding (T40)

Solves GDPR Article 17 + bi-temporal “what was the rule on date X” + immutable audit log in ONE primitive. Postgres wire-protocol compatible. v2 stable Q1-2026 with native EVICT. Saves ~30-50K LOC of custom infra. Architectural lock-candidate for Phase-1.

9. GPT-5.5 86% hallucination on AA-Omniscience — DISQUALIFYING for IW (T38)

Opus 4.7: 36%; Gemini 3.1 Pro: 50%; GPT-5.5: 86%. Even though GPT-5.5 wins on Harvey Legal at 91.7%, fabrication-when-uncertain is a probate-killer. Lock Path B’ on Claude Opus 4.7 / Sonnet 4.6 / Haiku 4.5 family — do NOT permit GPT-5.5 fallback in IW Phase-1.

Five standards-engagement chess-moves (T41) — Year 1 ~£15-30K total

1. OASIS Associate + LegalDocML TC observer — Q3 2026, ~£1,500-2,000. Choose Non-Assertion IPR mode (RAND would conflict with Apache+commercial). Highest knock-on (PAS-submitter route to ISO).
2. W3C VC WG + Agentic Web CG + Solid CG — Q3 2026, £6-15K (or £0 CG observer). VC 2.0 Render Method + Confidence Method shipping Sept 2026 must-influence window.
3. AAIF (Linux Foundation) Supporter — Q4 2026, £3-5K. AGNTCon + MCPCon Europe Sept 17-18 2026 Amsterdam — Rich should attend regardless of membership. Positions TT alongside Anthropic/OpenAI/Google/AWS/Cloudflare/Bloomberg.
4. BSI Member + UK mirror committees IST/41 + IST/33 + ART/1 — Q4 2026, £3-8K. UK national-body seat decisive for ISO PAS.
5. OAI engagement via Erik Wilde + Donnie Prakoso — Q1 2027, £0-10K. Moonwalk’s 2026 pivot to AI-agent readiness.

FHIR analogue timeline: OASIS TC 2027-2028 → BSI fast-track 2029 → ISO publication 2030. Tight but achievable. OpenAPI-as-precedent shows ISO is optional but adds AI-vendor credibility.

EU AI Act Aug 2 2026: INHERIT v2 schema (Apache) + InheritKit workflow (AGPL) NOT high-risk. InheritKit AI-augmented features (LLM probate decision-support) ARE high-risk under Annex III “administration of justice.” Recommendation: ISO 42001 conformance + CE marking Q1 2027; UK + non-EU launch can proceed earlier.

48 spike candidates surfaced (priority cluster summary)

HIGH priority (architecture-shaping):

• Cedar S1+S2+S4 cluster (~5 days wall-clock parallel) — mcp.inherit.com cedar-for-agents adoption + Sigstore-signed bundles + cedar-spec fuzz
• Catala 12 spikes including catala-mcp-server (3 wks) / catleg-uk/ch/eu (2 wks each) / Catala-NLLP-LLM-assist (3 wks) / Catala-Rust backend (3 wks)
• LinkML SPK-LINKML-OWL-DL-01 (10d) + linkml-store+DuckDB Phase-1 (7d) + schemasheets lawyer authoring (5d)
• IAS SP-IAS-01..05 (1-4 wks each) — Catala Skill v0.1 + InheritKitBench + Faraid Skill + Lean-4 verification + Anthropic Skills publish
• Modern infra: XTDB v2 crypto-shredding (5d) + WAC composition (5d) + Loro+Tiptap+Liveblocks IW editor (5d) + DOs facets multi-tenant (2d)
• Doc extraction: Path B” Mistral OCR 3 (3d) + Path L Local privacy tier (1 wk) + Spike L+ synthetic UK will corpus (2 wks defensive moat)

MEDIUM priority:

• Cedar S3 Catala→Cedar bridge (5-7d); S5 cedar-inherit Cursor extension (10d); S6 DASDT reference policy pack (5-7d)
• Catala bi-temporal extension (6 wks) / Catala+SHACL bridge (4d) / Catala-Blockly visual editor (4 wks)
• LinkML LLM authoring Skill (7d) / RDF round-trip property-test suite (5d) / alpha-component audit (3d)
• IAS variations (Lean-4 spike 2wks)
• Modern infra: TEE Phase-2 architecture sketch (2d) + MLS practitioner-chat prototype (3d)

LOW priority / deferred:

• Cedar S7 ZK feasibility (Phase-2)
• Modern infra Phase-3+ ZK proofs (Aleo/Noir) for consent
• Modern infra Phase-4+ FHE only when Zama hits 100k+ TPS
• LinkML HuggingFace dataset-card generator (5d)
• IAS Lean-4 verification spike (2 wks)

Cumulative T-programme + spike state after Round 5

• 41 T-files total: T1-T34 (Rounds 1-4) + T35-T41 (Round 5 deep-dives)
• All 41 research-complete
• 48 new spike candidates Round 5 + 22 from earlier rounds (15 active after spike-replacement, 6 closed-by-research) = ~70 spike-candidates surfaced across programme
• 17 cross-cutting candidates A-NEW-T-CROSS-01..17 (registered v2.24)
• 9 strategic findings Round 5 worth promoting to architecture-state
• 5 standards engagement chess-moves Y1 cost £15-30K
• Cumulative cost ~£300-450 subagent + Claude time across 5 rounds

What this enables for THE PLAN

The spike-portfolio is now sized: ~48 candidates with time-boxes + benefits. The plan needs to ruthlessly prioritise — most projects can budget 4-8 weeks of spike-time per quarter, not 48-spikes-worth. THE PLAN proposes a 2026-Q3 cluster (5-week wall-clock parallel programme) covering the highest-leverage spikes + an 18-month rolling roadmap.