plan: BindSpace Columns E/F/G/H — 4→8 SoA integration (scientific cross-check: 7 SOUND / 7 CAUTION / 0 WRONG)

[AdaWorldAPI]

Summary

Integration plan for extending BindSpace SoA from 4 → 8 column families.

New file: .claude/plans/bindspace-columns-v1.md (457 LOC, 10 sections)
Updated: .claude/board/INTEGRATION_PLANS.md (registered)

The 4 new columns

Column	Type	Size/row	What
H	EntityTypeId (u16)	2 B	Foundry Object Type equivalent
E	OntologyDelta (32B repr(C))	32 B	Per-cycle structural learning
F	AwarenessColumn ([u8; 256])	256 B	BF16-mantissa-inline epistemic annotation
G	ModelRef (u32)	4 B	ONNX style_oracle per-row binding

Total overhead: +366 B/row (+5.9%), 26.2 MB total — still fits L3 cache.

Scientific cross-check (Opus agent, 5 frameworks)

Framework	Claims checked	SOUND	CAUTION	WRONG
Jirak 2016 (Berry-Esseen)	3	2	1	0
Pearl do-calculus	3	3	0	0
NARS (Wang)	3	0	3	0
Kleyko VSA	3	2	1	0
Shaw Kan extension	2	1	1	0
Total	14	7	7	0

All 7 CAUTIONs have explicit mitigations in the plan.

24 deliverables across 4 phases

Phase 1 (H): 4 deliverables, all S — pure DTO, no SIMD impact
Phase 2 (E): 6 deliverables, mixed S/M — ontology delta + pearl_rung gating
Phase 3 (F): 9 deliverables, includes L — touches ndarray SIMD ops + cascade
Phase 4 (G): 5 deliverables, mixed S/M — after LF-50/52 traits ship

Brutal honest review

What's good:

• The 4→8 column extension is well-scoped: +5.9% overhead for 3 novel capabilities beyond Foundry
• Scientific cross-check caught real issues (unbind awareness recomputation, product-functor-not-nat-trans, per-word-not-per-statement NARS) and all are addressed
• Pearl rung gating (B2) is the single most important constraint and it's embedded in the struct layout
• Build order (H→E→F→G) has genuine dependency logic, not arbitrary sequencing

What's aspirational / risky:

• Column F (256 B/row awareness) is the heaviest addition and the least proven. The "BF16 mantissa" analogy is compelling but the per-word awareness derivations (bit-purity, margin, uniformity) are engineering heuristics with no formal theory behind them. NARS review flagged this: per-word truth has no grounding in Wang's framework. This could turn out to be noise, not signal.
• The awareness composition through the cascade (D-F7) is marked L effort but could easily be XL — every SIMD op in ndarray needs a parallel awareness path, and the cascade has 5+ steps. Testing that the composed awareness is meaningful (not just accumulated noise) requires empirical validation that isn't planned beyond one test (D-F9).
• The OntologyDelta "novel pattern detection" (D-E3) is hand-waved. What counts as "novel"? The plan says "new entity type, new relation, contradiction" but doesn't specify the detection logic. This is where the real complexity hides.
• Column G depends on LF-50/52 which are queued with no timeline. Phase 4 may never ship.
• The Pearl 2³ × semiring mapping (§8) is suggestive but unproven. Each Pearl rung doesn't inherently require a different semiring — the mapping is a design choice, not a mathematical consequence. A simpler design would use one semiring (HammingMin) for all rungs and let the causal mask handle the separation.

What's missing:

• No performance benchmark targets. "Still fits L3 cache" is necessary but not sufficient — what's the per-cycle latency budget? How many ns does Column F awareness derivation add per row?
• No migration path for existing BindSpace consumers. Adding 4 columns changes the SoA layout; every existing BindSpace::new() call needs updating. The plan doesn't mention backward compatibility.
• No discussion of Arrow projection. The callcenter's ExternalMembrane projects BindSpace → CognitiveEventRow (Arrow scalars). Columns E/F/G/H need projection rules (which are exported, which are internal-only).
• The Annotated<T> product functor is elegant but may be over-engineering. Most consumers just want (value, u8) — the generic type machinery adds complexity that simple tuples avoid.

Bottom line: Phase 1 (Column H) and Phase 2 (Column E) are solid and should ship. Phase 3 (Column F) needs a proof-of-concept measuring whether inline awareness actually improves meta_confidence vs the current 1 - F.total before committing to the full 9-deliverable plan. Phase 4 (Column G) is blocked and speculative.

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh

---

Generated by Claude Code