Merge pull request #248 from AdaWorldAPI/claude/plan-doc-surgery-from-3578405

plan-doc surgery: delete §18 Vsa10k BF16 hallucination + DU-3 precision note

Author: AdaWorldAPI

.claude/plans/callcenter-membrane-v1.md

@@ -1136,88 +1136,3 @@ Runtime lookup:
   PersonaId + KernelHandle → cognitive configuration for this turn
 ```
 
----
-
-## § 18 — VSA Precision Tiers + Generational Compression (Father-Grandfather)
-
-**Trigger:** 2026-04-23 — clarification that L4/L5 is the speed lane and must stay
-at fingerprint resolution; L3 is where precision lives; VSA is the wire format.
-
-### Three precision tiers
-
-| Tier | Format | Size | Properties | Where |
-|---|---|---|---|---|
-| **Fingerprint** | `[u64;256]` = `Fingerprint<256>` | 2 KB (16 Kbit) | Hash-quantized. One-way (no unbind). SIMD-fast Hamming. | L4/L5 substrate (30 ns/bind) |
-| **Vsa10k BF16** | `[bf16; 10000]` | 20 KB | Effectively lossless (Jirak noise floor not crossed). Supports `unbind(role)`. | L3 memory / cold storage |
-| **Vsa10k f32** | `[f32; 10000]` | 40 KB | Fully lossless. Full bind/unbind algebra. | Offline training / precision UDFs |
-
-**L4/L5 is the speed lane (motion, learning, fast dispatch).** The 16Kbit
-fingerprint is the right format there — shrinking it would leave the speed
-lane; inflating it to Vsa10k would blow the L3 memory budget at 30 ns/bind.
-This boundary is a hardware-budget invariant, not a design choice.
-
-**L3 is where precision is affordable.** The callcenter operates at human-turn
-rate (seconds between commits), not substrate rate. Full Vsa10k BF16 at L3
-costs 20 KB/row — trivial at that cadence. Alternatively: **RaBitQ**-quantized
-columns in Lance provide zero-copy ANN search at < Vsa10k RAM cost while
-preserving approximate unbind accuracy.
-
-### VSA as the wire format
-
-VSA IS the wire format — the medium through which cognitive content travels
-between layers:
-
-```
-Wire format:   bundle(SPO triples) → trajectory [Vsa10k or Fingerprint<256>]
-               Markov ±5 window  → bundle last 5 trajectories → episodic context
-               Markov ±500 window → bundle last 500 cycles → long-range context
-               SPO role bind     → bind(role_key, payload) → superposed into trajectory
-```
-
-`CognitiveEventRow` is NOT the wire format — it is the BBB-safe SCALAR
-projection of the wire state. The wire (VSA trajectory) stays inside. The
-projection (Arrow scalars) crosses the gate.
-
-### Father-Grandfather generational compression
-
-**Motivation:** Per-cycle fingerprints accumulate at 2 KB/cycle. 10K cycles
-= 20 MB (manageable in hot tier). 10M cycles = 20 GB (cold storage, still
-feasible). But for full-precision Vsa10k cold storage at 20 KB/cycle,
-generational compression prevents unbounded RAM growth during replay.
-
-**Hierarchy:**
-
-```
-"Son" (hot tier, L4/L5):
-  Per-cycle Fingerprint<256> rows — 2 KB each — fast, hash-quantized
-  Lance version-tagged; last M rows in hot dataset
-
-"Father" (~100-cycle bundle, L3):
-  MergeMode::Bundle over last ~100 per-cycle fingerprints
-  → single Vsa10k BF16 (20 KB) representing 100 cycles
-  → CK-safe (I-SUBSTRATE-MARKOV); saturating bundle preserves Markov property
-  → 100 × 2 KB → 20 KB: 10:1 compression; unbind survives at full L3 precision
-
-"Grandfather" (~1000-cycle bundle, cold tier):
-  MergeMode::Bundle over last ~1000 cycles (or over 10 Father vectors)
-  → single Vsa10k BF16 (20 KB) representing 1000 cycles
-  → 1000 × 2 KB → 20 KB: 100:1 compression; Markov property preserved
-  → unbind(role, grandfather) = approximate role trajectory over 1000 turns
-```
-
-**Relationship to existing Markov windows:**
-- ±5 window → per-cycle fingerprint braid (existing, L4/L5)
-- ±500 window → episodic memory (existing, L3)
-- ±1000 window → grandfather bundle (new, L3/cold storage)
-
-**CK safety proof (informal):** `MergeMode::Bundle` is commutative and
-associative in expectation (Johnson-Lindenstrauss + concentration-of-measure
-at rate ~e^(−d), I-SUBSTRATE-MARKOV). Bundling 100 fingerprints into one
-Vsa10k is a saturating bundle; the Chapman-Kolmogorov property holds for the
-bundle as a whole — the state transition from "Son" to "Father" is a valid
-Markov step.
-
-**Implementation note:** deferred. Current hot dataset stores fingerprints
-only. Father/Grandfather columns are Phase B additions when cold dataset
-accumulates > 10K cycles. No schema change to `CognitiveEventRow` required —
-generational bundles live in a separate Lance dataset (compression tier).

.claude/plans/unified-integration-v1.md

@@ -193,13 +193,6 @@ ORDER BY dispatch_score DESC
 LIMIT 10;
 ```
 
-**Precision note:** The UDF signatures above use `[u64;256]` (fingerprint, L4/L5 tier)
-for dispatch scoring — fast Hamming distance, approximate role overlap. This is correct
-for the dispatch path (30 ns/bind speed lane). For deep role recovery via `unbind()`
-at full precision (RoleDB Phase B), the UDF input should be Vsa10k BF16 from the L3
-cold dataset — deferred until the L3 Vsa10k or RaBitQ cold columns exist. See §18
-of `callcenter-membrane-v1.md` for the precision tier architecture.
-
 **BBB compliance:** UDFs operate on internal_dataset only. Results are scalar
 (f32 dispatch scores, u32 Hamming distances). VSA types never cross to external_dataset.