Round 6 — Harvard-mode ambiguity-elimination complete (Saturday 25 April 2026)

After Round 5 surfaced 9 strategic findings + 48 spike candidates + £25-40m UK Y3 ARR opportunity, Rich commissioned: “adopt the mentality of a time-served Harvard research lab engineer who has spent 25 years perfecting their researching and anticipatory skills. Go out there and uncover hidden gold and pearls of wisdom that dramatically reduce the uncertainties ahead of us — leave no stone unturned.”

8 parallel deep-dive subagents dispatched (T42-T49); ~38,000-42,000 words; ~£200-300 cost; covered all 60 items in THE PLAN’s Waves 1-6 + Wave-S. Zero retroactive disruption to locked architectural decisions; substantial reframing of 12 spike-readiness assumptions.

Top-5 decision-flipping findings

1. Cedar/Catala structurally INCOMPATIBLE for general translation (T42 + T43 confirmed independently)

Cedar: order-independent + default-deny + binary outcomes. Catala: priority-ordered defaults + arbitrary-typed outputs. Spike 5.5 (Catala→Cedar compilation) DROP-OR-NARROW. Only narrow boolean access-control sub-fragment maps. Architectural implication: Catala derives jurisdictional rules; Cedar evaluates a separate permission-layer. Different abstraction strata.

2. “Path A WASM” was misnomer — actually js_of_ocaml today (T43)

@catala-lang/french-law ships JavaScript only; NO .wasm files. wasm_of_ocaml port to InheritKit is green-field engineering choice (de-risked by Tarides + Jane Street Feb 2026 optimisations: 3.9× speedup, 2-8× over js_of_ocaml). Spike 1.2 reframed: measure js_of_ocaml today; wasm_of_ocaml port becomes Phase-1.5 spike (call it 5.14).

3. 🚨 Cultural-alignment partner-firm cost £240-450K/year — NOT in budget ledgers (T49)

Z7-FAITH α_max lock requires 7 mandatory reification-mitigations: per-rule provenance + non-authoritative clause + bilingual translation-audit (Arabic/Sanskrit/Hebrew separate) + women’s-rights counter-expert review + runtime-feedback loop + deprecation affordance + candidate-status publication. Each mitigation has staffing + recurring-cost implications. £80-150K per pillar per year × 3 pillars = £240-450K total Phase-1.5 run-rate. Not currently in v2 budget ledgers — biggest single budget surprise of programme.

4. Standards-engagement Y1 budget HALF Round-5 estimate (T48)

Round-5 said £15-30K; Round-6 verified actual: ~£6-8K (OASIS Associate $1, 975 + W 3 C S M E$ 1,905-2,500 [NOT £6-15K] + AAIF Associate $0 v ian o n - p r o f i t /£8 - 12 K S i l v er + A GN T C o n$ 1,000-1,500 + BSI £0-500). OAI is FREE for TT (sub-10-employee rule waives LF prereq). Frees £10-20K headroom for OIDF member-org tier ($2K) + reference impl development.

5. IAS leapfrog confidence raised to MEDIUM-HIGH (T47)

Round-5 LOW-MEDIUM verdict; Round-6 escalated to MEDIUM-HIGH. Anthropic Skills format spec locked + minimal; OpenAI adopted same format Dec 2025; cross-vendor portability real; lawvable/awesome-legal-skills has 16 legal Skills with ZERO succession/faraid/HSA/halakhic — white space confirmed. Anthropic legal plugin Feb 2026 explicitly NARROW (contract review/NDA only). Window: 9-12 months high-leverage / 12-18 months conditional. Highest-probability window-closer: Anthropic Cowork v2 plugin (6-9 months).

Per-cluster top-3 findings (8 clusters)

T42 Cedar

Bundle weight v4.10.0 = 4.20 MB raw / 1.42 MB gzipped / 1.21 MB brotli (12% of free Workers limit)
OPA-vs-Cedar bundle architectures non-comparable (OPA = loader+per-policy; Cedar = constant engine) — Cedar wins for INHERIT 200-500 policy set
RFC #112 wide open: 1 comment, 0 reactions — INHERIT’s “executor-PoA expires on death-of-grantor” is THE killer fiduciary application

T43 Catala

v1.1.0 clerk run JSON I/O turns catala-mcp-server from 4 weeks → ~1 week
Cedar/Catala structurally incompatible (DROP-OR-NARROW Spike 5.5)
Path A is js_of_ocaml today; wasm_of_ocaml is green-field upgrade path

T44 LinkML

gen-owl already has OWLProfile {dl, full} enum — Spike 5.13 reduced 10d → 3-5d
gen-shacl is most-improved generator 1.10→1.11 (PRs #3098, #3095, #3061, #3090, #3278, #3291, #3293)
🚨 linkml-store is pre-1.0 with explicit instability warning — Round-5 wrong to call “scaling backbone”; DOWNGRADE to sidecar

T45 Doc-extraction

🚨 Mistral OCR 3 INDEPENDENT score 79.75 OmniDocBench (NOT marketing 94.89%); below open-source SOTA
Path B” cost-at-scale £100/mo at 10k wills (down from £230) — OCR not cost driver
🚨 Apple Foundation 3B structurally underweight (~44% MMLU) — pre-process+redact only, NOT primary will-extraction

T46 Infrastructure

🎯 Pydantic AI v1.85.1 outperforms LangGraph 44% lower P95 / 5× fewer errors / 2.7× lower tokens
🚨 DOs facets STILL BETA (Round-5 said GA — wrong); SQLite-in-DO write-fail at 10 GB
🚨 Wasmtime CVE-2025-62711 in 38.0.0-38.0.3 — pin ≥38.0.3 or 39.x

T47 IAS Leapfrog

MEDIUM-HIGH leapfrog confidence; window 9-12 months high-leverage
Halakhic = LARGEST white space (zero computational rule-engine anywhere)
Highest-probability window-closer: Anthropic Cowork v2 plugin (6-9 months); mitigation = STEP UK / ISC London / Beth Din London partner-firm relationships months 1-9

T48 Standards-engagement

“DASDT” was Round-5 misnomer — actual is OIDF DADE CG + eKYC&IDA WG; Saxe + Maler + Kiser co-chair
AAIF Associate tier = $0 for non-profits; TT could spin up “Open Inheritance Foundation” affiliate
🎯 No standards body owns “Digital Estate Schema” — TT could propose OASIS TC + chair from formation at $1,975/yr Associate cost

T49 Faith-traditions

Library coverage strongest of any T-file dimension (Hallaq + Saiman + Bhattacharyya-Panda all deep-read with ULID-cited quotes)
Hindu HSA = HIGHEST cultural risk (HCUK partner depth thinnest); Jewish MEDIUM-LOW (700-year le-halakhah-ve-lo-le-ma'aseh disclaimer convention); Islamic MEDIUM
🚨 Cultural-alignment partner-firm cost £240-450K/year for 3 pillars NOT in budget ledgers

12 spike-readiness changes from THE PLAN v1.1

Spike	Was	Round-6
1.1 Cedar bundle measurement	2 days	1 day (already measured)
1.2 Path A WASM	2.5 days	🟡 PIVOT — measure js_of_ocaml today; wasm_of_ocaml port = Spike 5.14 Phase-1.5
1.7 DOs facets multi-tenant	”GA Q1-2026”	🚨 STILL BETA — ship with shared-table fallback
2.1 SPK-LINKML-01 LITE	7 days	✅ READY (with Biolink 5-flag set)
2.2 linkml-store + DuckDB	”scaling backbone”	🟡 DOWNGRADE / DEFER — Postgres-primary + DuckDB sidecar
2.6 WAC composition	Cedar+Catala+LinkML	🟡 REDUCE — Cedar+Catala only; LinkML host-side
2.7 Loro+Tiptap+Liveblocks	”experimental”	✅ STABLE v1.0 since 30 Apr 2025
2.8 Code Mode SDK	1 week	4-6 days (Anthropic+WorkOS confirm 32-81% real workloads)
5.4 catala-mcp-server	3 weeks	~1 week (v1.1.0 JSON I/O)
5.5 Catala→Cedar	5-7 days	🔴 DROP-OR-NARROW (structural incompatibility)
5.13 gen-owl-dl	10 days	3-5 days (DL flag exists)
5.8 Apple Foundation primary tier	”primary will-extraction”	🟡 PRE-PROCESS+REDACT only

Plus:

1.5 Path B” adoption rationale: cost/EU-residency only (NOT accuracy)
Path B” cost-at-scale £100/mo at 10k wills (NOT £230)
3.5 Hindu HSA Skill: SLIP-TRIGGER Phase-1b (highest cultural risk; partner-firm depth thinnest)

Budget surprises (both directions)

Line	THE PLAN v1.1	Round-6 actual	Delta
Standards engagement Y1	£15-30K	£6-8K	-£9-22K saved
W3C SME tier	£6-15K	~£1.5-2K	-£4-13K saved
AAIF	£3-5K	$0 (non-profit) or £8-12K (Silver)	path-dependent
Path B” OCR @ 10k wills/mo	£230/mo	£100/mo	-£1,560/yr saved
🚨 Cultural-alignment partner-firm protocol	NOT IN LEDGER	£240-450K/year	+£240-450K/year (NEW)

New commercial opportunities surfaced

OASIS TC for “Digital Estate Schema” — propose + chair from formation at $1,975/yr (no body owns this gap)
“Open Inheritance Foundation” non-profit affiliate — AAIF Associate $0 access
Vercel skills.sh launched 20 Jan 2026 — NEW formal Skills marketplace TT can list day-1
DASDT-track Cedar reference policy pack — co-publish with OIDF DADE Talking Tour 2026
InheritKitBench Phase-A UK E&W (~25 statute→Catala/Cedar pairs) — 3 weeks artefact for ISO submission

Hidden gold (Round-6 specific)

Sefaria.org Mishneh Torah Hilchot Nachalot CC-BY — open-licensed primary source for halakhic rule authoring
UUM Malaysia Nafed thesis — recruit-or-coauthor target for Indonesia/Malaysia faraid expert systems
Singapore Syariah Court online calculator — state-grade Shafi’i test-oracle for IAS Phase-D validation
Exeter “Material Culture of Wills” project (Leverhulme 2023-2027) — first publicly-available will corpus; open dialogue post-project release
Sourcemeta uses MCP JSON Schema 2020-12 default — strongest external validation of INHERIT schema-primary architecture
XTDB v2.1 ATTACH data-mesh primitive — partner-firm cross-queries with no ETL
DOs facets + per-tenant SQLite = “GDPR-by-architecture” — right-to-be-forgotten = DROP DATABASE
AGENTS.md in linkml/linkml — Claude-Code-friendly contributor guide; TT can contribute upstream PRs at high acceptance rate

Anti-portfolio additions

Catala→Cedar general compilation (structural incompatibility)
linkml-store as primary INHERIT data plane (pre-1.0)
Mistral OCR 3 chosen on accuracy (79.75 < SOTA)
Apple Foundation 3B as primary will-extractor (~44% MMLU)
LinkML→Rust / Go interop Year-1 (gen-rust missing constraints; gen-go zero production users)
LayoutLMv3++ A/B vs Path B’ (no LayoutLMv4; multimodal-LLM has overtaken)
langgraph-prebuilt 1.0.2+ (silent ToolNode.afunc breakage — pin <1.0.2)
Wasmtime 38.0.0-38.0.3 (CVE-2025-62711 — pin ≥38.0.3 or 39.x)

Cumulative T-programme state after Round 6

49 T-files all research-complete (T1-T34 baseline + T35-T41 Round-5 + T42-T49 Round-6)
~70-80 spike candidates surfaced; ~12 reframed by Round-6
Cumulative cost ~£500-750 across 6 rounds
0 retroactive disruption to locked architectural decisions
2 architecture-state amendment candidates added (Catala/Cedar incompatibility lock; cultural-alignment partner-firm protocol budget line)

Next actions (THE PLAN v1.2)

Drop Spike 5.5 Catala→Cedar
Add Spike 5.14 wasm_of_ocaml port (Phase-1.5)
Update Spike-portfolio table with 12 readiness changes
Add £240-450K cultural-alignment line to budget
Add 5 new commercial opportunities to Section 3
Update Wave-S budget to £6-8K
Add 8 anti-portfolio additions to Section 6

TT Claude Memory

Explorer

project_round_6_harvard_complete_2026_04_25