TT Claude Memory

feedback_universal_production_pipeline_sequence

Properties1
descriptionRich-directive 2026-05-02 codifying the canonical 3-layer production pipeline for jurisdiction-specific INHERIT v2 content. STEP 1 (SEED) — fully utilise existing v6.6 hand-authored material before any LLM step (it's depth-4.6/5 with statute citations + enum-typed terminology + temporal-rule pattern across 21+ jurisdictions); STEP 2 (LLM-REFINE) — only AFTER STEP 1 use LLMs to research weaknesses, fill gaps, and emit Catala scopes / Cedar policies / SKOS classifier schemes; STEP 3 (partner-REVIEW) — pass refined output to a jurisdiction-partner-firm for production-grade legal review. Skipping or reordering steps WASTES the rich SEED material we already have and weakens the partner-review signal.

5 min read

Rule: For any jurisdiction-specific INHERIT v2 content (rules / schemas / classifier values / fixtures / temporal lookups), the production pipeline is strictly sequenced:

  1. SEED — fully utilise existing v6.6 material first. v6.6 (~/openinherit/code-standard/) is a high-quality 80%+ SEED across 21 dedicated jurisdiction-or-tradition extensions (UK-E&W, Scotland, Ireland, US, Canada, AU/NZ, EU-26, Switzerland, India, Japan, China, Hong Kong, Singapore-Malaysia, UAE, Israel, Brazil, Latin-America, Africa-customary, plus Islamic / Jewish / Hindu faith extensions) plus 32 jurisdictions in reference-data/jurisdiction-profiles.json, plus 11-jurisdiction tax-thresholds with multi-decade history. Per S1 spike outcome 2026-05-02 (~/off-github/library/projects/inherit/T-spike-eps-iota-S1-v66-seed-audit-2026-05-02.md): sample mean depth 4.6/5, statute-cited (IHTA 1984 / AEA 1925 / IFPA 1975 / HSA 1956 / ZGB Art 457-640 / etc.), enum-typed (22-value Islamic heir enum, 15-rank asaba hierarchy, etc.), temporal-rule pattern universal. Do NOT skip the SEED step. Reading the existing extension is faster + cheaper + more accurate than an LLM cold-start.

  2. LLM-REFINE — only AFTER SEED utilisation. Use LLMs to (a) research weaknesses in the SEED, (b) emit Catala scopes / Cedar policies / SKOS classifier schemes derived from the SEED’s rule-text + statute citations, (c) populate the remaining 10-15% per jurisdiction (not 20% — depth 4.6/5 means SEED is stronger than the working assumption). Per S1: cost-per-jurisdiction revises DOWNWARD when REFINE step has SEED handles to elaborate against vs starting from scratch. The LLM step is the SECOND layer, not the first.

  3. partner-firm-REVIEW — only AFTER REFINE. Pass refined output to a jurisdiction partner-firm (e.g. Manchester Law Partnership for E&W; future partner for Scotland; future for Switzerland; etc.) for production-grade legal review. The maturity: draft|candidate|stable signal in the SEED already self-reports per-extension confidence — partner-REVIEW is the signal that moves it to stable. Without REFINE having run first, partner-REVIEW is reviewing v6.6-as-is which is not the production artefact.

Why: Rich-directive 2026-05-02 BST after S1 spike outcome surfaced the depth + breadth of v6.6 material:

“there is a huge amount of valuable material revealed by S1, and we need to ensure that our plans utilise it, before using LLMs to then research any weaknesses, and finally passing to a jurisdiction partner to review”

Two failure modes this rule prevents:

  • LLM cold-start (skip step 1) — generates plausible-but-shallow output that reinvents what v6.6 already has, costs more LLM budget, and is harder for partners to review because it lacks the statute-citation handles partners use to anchor review.
  • SEED-direct-to-partner (skip step 2) — partners review v6.6 as-is which is not the production artefact (no Catala scope code, no Cedar policy code, no SKOS TTL); wastes partner time on the wrong layer.

How to apply:

  1. Before dispatching any spike or task that touches jurisdiction content: read the relevant ~/openinherit/code-standard/v3/extensions/<juris>/<juris>.json first. Capture verbatim citations + enum values + temporal-rule references. THEN frame the LLM-REFINE step against those handles, not against an empty prompt.

  2. In spike planning (e.g. ε.ι derisking suite Phase B/C/D spikes that operationalise the universal-pipeline): explicitly state which v6.6 extension is the SEED for the work. Example: S4 (UK&W NRB pilot) seeds from v3/extensions/uk-england-wales/uk-england-wales.json (352 lines, depth 5, IHTA 1984 cited). S5 (FIBO SSSOM) seeds from reference-data/local-term-mappings.json (4027 lines).

  3. In cost modelling: calibrate LLM-REFINE budget to depth-4.6/5 SEED quality — that is, ~10-15% LLM work per jurisdiction × per artefact-type, not the looser 20%. If a Layer-4-style cost story is being scored, the SEED inventory should be cited as evidence the cost-per-jurisdiction is closer to the lower end of any range.

  4. In partner-recruitment + partner-engagement (e.g. MLP / future jurisdiction partners): the work product going to partners is REFINED output (Catala scopes / Cedar policies / SKOS classifiers derived from the SEED), not v6.6 itself. The SEED + REFINE outputs go side-by-side in the partner-review packet so partners can see both the source material and the operationalised form.

  5. In ζ-Q3 ε.ι lock decision pathway: S1 spike outcome (KILL CONDITION NOT MET on all 3 legs) is direct evidence that ε.ι Layer 4 (universal-production-pipeline) cost story is sound — proceed with ε.ι lock unless S2-S10 surface contradicting evidence. Cost story revises DOWNWARD given depth-4.6 vs depth-4.0 working assumption.

Boundary tests (when this rule fires STRONGLY):

  • ✓ Authoring per-jurisdiction Catala scope code (S3, S4, S6, S8 in ε.ι derisking suite)
  • ✓ Authoring per-jurisdiction Cedar policy code
  • ✓ Authoring SKOS classifier schemes for jurisdiction-specific taxonomies
  • ✓ Populating jurisdiction-keyed reference-data tables (tax-thresholds extensions, form-requirements, etc.)
  • ✓ Cost-modelling for ε.ι Layer 4 / Year-2+ jurisdiction expansion
  • ✓ Partner-firm engagement scoping (what work product goes to partners)

Boundary tests (when this rule does NOT apply):

  • Substrate-architectural decisions (LinkML / Cedar / Catala / SSSOM substrate choices) — those are jurisdiction-agnostic
  • Tooling / DevOps / CI gate decisions — those don’t touch jurisdiction content
  • Brand / commercial / acquirer-narrative decisions — those don’t touch jurisdiction content

Codification trigger: Rich directive 2026-05-02 BST after S1 spike outcome (T-spike-eps-iota-S1-v66-seed-audit-2026-05-02.md) surfaced 21+ jurisdictions × depth-4.6/5 SEED material — directive made the sequence explicit (“UTILISE → THEN LLM → THEN partner”) so plans don’t waste the SEED.

Related memories:

  • project_zeta_q3_eps_iota_S1_2026_05_02.md — S1 spike outcome record
  • feedback_actively_use_t_files_in_scorecard_authoring.md — analogous discipline for T-file material in scorecards (reads first, cites verbatim, uses as starting point for LLM work)
  • feedback_always_check_library_indexed_first.md — analogous discipline for the TT research library (~99 books)

Graph View

Backlinks