Architecture · vault v2 · 2026-05-16

Seven hierarchies.

How the vault thinks — loading tiers, compression, co-loading, the pipeline, an entry’s layers, the confidence ceiling, and the query cascade.

HIERARCHY 01

Loading tiers

Distance from working memory. Inner tiers are always present; outer tiers load only when triggered.

DISTANCE FROM SESSION 0 far TIER 0 · ROOT Always loaded CLAUDE.md principles.md SESSION START EVERY TIME 2 files TIER 1 · meta/ Load on trigger CO-LOADED PAIR structural-recommendation.md structural-trap-catalog.md visualization .md canvas-conv .md reflection .md STRUCTURAL TRIGGER HTML / CANVAS / FAILURE 5 files TIER 2 · wiki/ Queried per need ○ index.md ○ log.md ○ overview.md ○ entities/ ○ concepts/ ○ sources/ ○ analyses/ QUERY / INGEST ~200 files TIER 3 · raw/ Immutable · never modified · PDFs / images / spreadsheets / scans READ-ONLY ~90 files
Closer to top = closer to working memory. Outer tiers wait for a named trigger.
HIERARCHY 02

Compression ↔ expansion

The always-loaded CLAUDE.md carries a compressed mirror of the full pipeline. The expansion lives in meta/ and unfolds only when a structural recommendation is in scope.

COMPRESSED · TIER 0 CLAUDE.md Reasoning Discipline 6 rules 5 inline self-flag markers ~150 words always in working memory EXPANDS WHEN TRIGGER FIRES → EXPANDED · TIER 1 structural-recommendation.md 17 sections 7+1 pipeline stages 9-section output template 6-step critique protocol 6 self-flags as artifacts 5 confidence bands handoff protocol ~3,500 words loaded on structural trigger only
~150-word compressed always-on · ~3,500-word expansion behind a trigger
HIERARCHY 03

The two-file recommendation system

One trigger fires both files. HOW (the pipeline) + WHAT (the catalog) are always loaded together, never one without the other.

STRUCTURAL-RECOMMENDATION TRIGGER any rec affecting ownership / entities / succession / governance / compliance THE HOW structural-recommendation.md the pipeline + the discipline A B C D E F G Z 7+1 pipeline stages + 9-section output template + 6-step critique protocol + 6 self-flags as artifacts + 5 confidence bands + professional-handoff protocol THE WHAT structural-trap-catalog.md per-statute landmines with provenance 1 2 3 4 5 6 7 8 8 active entries · full provenance ENTRIES 9 – 22 14 stub anchors · not citable until drafted to full depth 95%+ safe-harbor 65-79% defensible <65% below threshold stub (not yet citable)
Same trigger fires both files · same Stage A scans them together
HIERARCHY 04

Pipeline cascade

Eight stages run in order. Each stage emits an artifact. A recommendation that lacks any artifact does not surface — the engine returns an open question instead.

IF Z FINDS A REAL FLAW · RETURN TO A A retrieval context manifest B time-state chronological table C reject-pattern match table D counter-arg strongest challenge E quantitative claim list w/ backings F conflation disambig. glossary G confidence + sign-off map Z adversarial inline challenge Each stage emits an artifact. Missing artifact ⇒ no surface.
Stages A–G filter the proposal · Stage Z is the final adversarial gate
HIERARCHY 05

Catalog entry · three layers

Each catalog entry is a stacked cross-section. Substantive content on top · provenance in the middle · downstream discipline at the base.

ENTRY 03 CA R&T §62(a)(2) — Proportional ownership preserved through entity CALIFORNIA · PROP 13 LAYER 1 · SUBSTANTIVE What fires when. The mechanic. Where this vault has hit it. Triggers when real property transferred to entity & proportional ownership preserved Consequence no change-in-ownership at contribution; Prop 13 base year carries LAYER 2 · PROVENANCE How strong is the source. When last checked. Per-fact-pattern confidence. Primary authority CA R&T §62(a)(2) · BOE Rule 462.180 Last verified 2026-05-16 Verification confidence 80% DOMESTIC 55% LOOK-THROUGH-BVI LAYER 3 · DOWNSTREAM DISCIPLINE What this entry constrains in everything that depends on it. Inheritance any rec depending on this cannot exceed 55% on the look-through claim Verification debt role: CA real-estate tax attorney · none currently engaged what how strong what depends
Same anatomy on all 8 active entries · stubs have header only until drafted
HIERARCHY 06

Confidence ceiling

A recommendation inherits the floor of the lowest-confidence catalog entry it depends on. No silent uplift.

100% 95% 80% 65% 55% 40% 0% SAFE-HARBOR WELL-GROUND DEFENSIBLE JUDGMENT SPECULATIVE CATALOG ENTRY #3 CA R&T §62(a)(2) look-through-BVI 55% per-entry verification confidence CEILING RECOMMENDATION one-step §351 BVI → LLC would justify 80% blocked by ceiling 65% aggregate of sub-claims bounded by entry floor A recommendation cannot present a higher confidence than the lowest-confidence catalog entry it depends on.
Each catalog dependency contributes a floor · the rec sits at the minimum
HIERARCHY 07

Query cascade

A structural query flows down through the tiers and stages, gathering constraints as it falls. The recommendation surfaces at the bottom — or returns to the top as an open question if any stage fails.

TIER 0 always TIER 1 trigger TIER 2 query RENDER surface 01 · INPUT “one-step or two-step?” 02 · DETECT CLAUDE.md fires on “structural recommendation” 03 · CO-LOAD structural-recommendation.md structural-trap-catalog.md 04 · STAGE A · RETRIEVAL catalog scan: 1 2 3 4 5 6 7 8 8 entries match · wiki retrieval: [[bvi-structure-one-vs-two]] manifest: 12 prior analyses · 4 rejected patterns · 3 open questions 05 · STAGES B–G · CONSTRAIN AGAINST CATALOG B C D E F G time-state · reject-pat · counter-arg · quant · conflate · conf+gap 06 · INHERITANCE COMPUTES CEILING rec conf ≤ min(entry confs) · floor from #3 (55%) + #5 (40-50%) · lands at 65% 07 · STAGE Z · ADVERSARIAL PASS find the strongest flaw · if real → return to Stage A 08 · RENDER · 9-SECTION OUTPUT TEMPLATE TL;DR · Recommendation · Why-not-alternatives · Reasoning chain Quantitative claims · Confidence · What-doesn’t-solve · Triggers · Sources 09 · SELF-FLAG PASS · 6 CHECKS cleared → moss ✓ chips · fired → gold ⚠ chips inline 10 · RECOMMENDATION SURFACES IF Z FINDS A REAL FLAW · RETURN TO A
Top-down cascade across tiers · failure routes back to Stage A · success surfaces at the bottom