TT Claude Memory

project_session_2026_04_18_v2_architecture_scorecard

Properties1
descriptionSaturday 18 April 2026. Produced 4-option architecture scorecard (B/C/D/E), D/E improvements memo, Eric's v0.5 book letter, Option F handoff prompt. Established D→E phased recommendation as pre-F baseline; Option F to be produced in fresh session with strict library-grounding barrier. Committed 4 commits, all pushed.

5 min read

Session date: Saturday 18 April 2026, long session producing the architecture-decision spine for INHERIT v2.

What was produced

1. 4-option architecture scorecard + 4 option deep-dives

  • docs/superpowers/specs/2026-04-18-inherit-v2-architecture-options-scorecard.md (master scorecard, 353 lines / ~6,800 words, v1.0, PDF in Downloads)
  • docs/superpowers/specs/2026-04-18-inherit-v2-architecture-options/option-{B,C,D,E}.md (deep-dives, 3,618–5,735 words each)
  • Headline scores: D 66/80, E 60/80, B 55/80, C 51/80
  • Recommendation: D→E phased path over 18 months (D core months 1-6, + Catala months 7-12, + fine-tuned LLM months 13-18)

2. D/E improvements memo (library-grounded, 6,665 words)

  • docs/superpowers/specs/2026-04-18-inherit-v2-architecture-options/improvements-memo.md
  • Top 5 proposed changes (LOC-ordered): SQLite+sqlite-vec Year-1 substrate, OpenAI fine-tuning API until ARR > £500k, Catala-only counterfactual harness (drop OpenFisca), event-sourcing + MVCC + three-timestamp (Kleppmann DDIA 2e), Khononov bounded-Catala-module discipline
  • Revised totals: D 185k → 163k LOC (ratio 1:3.1 → 1:2.7, score 66 → 68), E 247k → 205k LOC (ratio 1:3.7 → 1:3.0, score 60 → 62)
  • PDF in Downloads

3. Eric’s v0.5 book letter (4,678 words)

  • docs/superpowers/specs/foundation-rework/reading-corpus/recommended-books-v0.5-options-DE.md
  • 6 2025-2026 Kindle-convertible book recommendations, ~£208 total
  • Core shelf: Brousseau & Sharp LLMs in Production, Godoy fine-tuning guide, Gullí Agentic Design Patterns, Iusztin & Labonne LLM Engineer’s Handbook
  • Supporting shelf: Alto AI Agents in Practice, Raschka Build a LLM From Scratch, Lanham AI Agents in Action 2e MEAP
  • Free open-access companions catalogued for Gap 4 (formal methods) and Gap 5 (legal KGs) where 2025-2026 books were thin
  • Rich acquired 4 of the 6: Godoy, Gullí, Raschka, Brousseau & Sharp — PDFs + EPUBs in ~/off-github/library/new/ at session close, waiting for Option F session to index
  • PDF in Downloads

4. Option F handoff prompt (v1.1)

  • docs/superpowers/scoping/option-F-and-library-expansion-handoff-prompt.md
  • Strict 3-track sequence (index 4 new books + download ~14 Gap 2/4/5/6 articles → MANDATORY library read + evidence digest → Option F drafting)
  • Convergent-vocabulary probe required across expanded 49+ indexed corpus
  • Every architecture claim in option-F.md must cite a ULID-tagged library quote
  • Rich’s verbatim directive: “absolutely whatever technology you think is best, with no bias towards software we are currently using”

5. 4 Option-C-batch books read in this session

Tier + ULID counts in notes files:

  • Bratanič Essential GraphRAG (tier 3, 25 Q-BRAT-*)
  • Negro Knowledge Graphs and LLMs in Action (tier 2, 29 Q-NEGR-*)
  • Tamò-Larrieux AI and Law (tier 2, 40 Q-TAMO-*)
  • Raieli Building AI Agents (tier 3, 27 Q-RAIE-*)

6. W3C spec bundles indexed as reference

  • w3c-sparql11 — SPARQL 1.1 Overview + Query Language (216pp)
  • w3c-jsonld11 — JSON-LD 1.1 + Framing + Primer + Learn + landing (294pp)

Key findings load-bearing for future sessions

Convergent-zero finding (Option C disconfirmation)

Across 1,733 combined pages of 2025 specialist literature (Bratanič + Negro + Tamò-Larrieux + Raieli), the following scored zero mentions: SHACL, JSON-LD, Oxigraph, Jena, Stardog, GraphDB, Blazegraph, Virtuoso, Akoma Ntoso, Protégé, FOAF, competency-question methodology. Akoma Ntoso zero even in a specialist legal-informatics book. Option C’s full stack is convergently disconfirmed in 2025 mainstream.

What the 2025 literature DOES endorse

Neo4j/Memgraph LPG (3-book convergence), Pydantic/JSON Schema (schema-in-prompt), RDF/OWL as ontology-PUBLICATION format only (Negro’s handleVocabUris: IGNORE), GraphRAG, LoRA/QLoRA on 7-70B (Raieli — solo-feasible), LangChain/LlamaIndex/Haystack, Catala DSL with worked inheritance-tax examples as its own corpus, OpenFisca (30+ countries production), ACE Attempto controlled language, ELI/ECLI IRIs, SaaS/MaaS/DaaS/RaaS commercial framework.

Critical caveats shaping D and E

  • Mirzadeh 2025 (cited by Raieli ch 11): o1-class frontier LLMs fail on perturbed formal reasoning. LLM-as-oracle is disqualified; E’s three-canonical-layer design (LLM / Catala / KG) follows from this.
  • MCP is pre-mature in 2025. Raieli’s 707pp body mentions MCP once (back-cover marketing only). TT must build its own MCP server; cannot rely on ecosystem readiness.
  • Llama 3.1 is commercially excluded for Options D and E: Community Licence’s “no training non-Llama on Llama outputs” clause poisons the DaaS revenue line to frontier AI vendors; 700M-MAU clause creates acquirer-diligence leverage against TT. Mistral Nemo 12B (Apache 2.0) is the recommended base model.

6-family decomposition works in all options

Adopt regardless of paradigm choice. Earlier session proposed Catalogue as sub-family of Assets (5 top-level families instead of 6) — carry that convention forward.

What’s pending (for the fresh-session Option F handoff)

  1. Index the 4 new books in ~/off-github/library/new/ (Godoy, Gullí, Raschka, Brousseau & Sharp)
  2. Download + render + index ~14 free articles (Hogan, Shimizu-Hitzler, Keet, Lean 4 × 2, Practical Alloy, Learn TLA+, Catala POPL 2021, OpenFisca, ELI, ECLI, Nay Stanford paper, MCP spec + roadmap)
  3. Mandatory Track 3 Step 0 library read before any Option F drafting — produce evidence digest
  4. Write option-F.md with zero bias toward currently-used TT tech, every claim ULID-cited
  5. Update master scorecard to v1.1 with F column added and recommendation re-examined
  6. Auto-PDF to Downloads; commit in 4 sequenced commits; push

Handoff prompt at docs/superpowers/scoping/option-F-and-library-expansion-handoff-prompt.md (committed + pushed). Paste from ## THE PROMPT to ## END OF PROMPT into a fresh cd ~/testatetech/docs-strategy && claude --dangerously-skip-permissions session.

Git state at session close

Remote == local at be63cee. 4 commits on main today:

  • 21f8808 — scorecard + 4 option files
  • c3573e7 — library catalogue v1.0 → v1.17
  • 42146ea — Option F handoff prompt v1.1
  • be63cee — improvements memo + Eric’s v0.5 book letter

New memories created this session

  • feedback_no_current_stack_bias_for_inherit_v2.md — Vercel/Supabase-not-assumed directive
  • project_options_d_and_e_synthesis.md — D/E shape details + “be brave” directive (updated during session with Catala + OpenFisca + Attempto + ELI findings)
  • project_session_2026_04_18_v2_architecture_scorecard.md — this file

Open architectural questions for Option F to resolve

  1. Does Option F displace D→E on the 16-criterion scorecard, or does it come in behind?
  2. Which substrate does F pick if Neo4j is explicitly not a default — XTDB, Datomic, SurrealDB, SQLite+extensions, something else?
  3. Which runtime language — Rust, Go, Elixir, Gleam, OCaml, something else?
  4. Does F use MCP or does it pick an alternative agent protocol?
  5. Does F retain Catala + OpenFisca or propose a different formal-methods stack (Lean 4, Alloy, TLA+)?
  6. Is Mistral Nemo retained or does F pick a different base model?
  7. Does F’s LOC estimate come in under D’s 163k / E’s 205k post-improvements? What’s F’s maint ratio?

These are the questions the fresh session will answer.

Graph View

Backlinks