RSHL Whitepaper Audit Report

Date: 2026-05-20

Auditor: Automated Code vs. Theory Cross-Reference

Document: RSHL-Inventor-Disclosure-2026.md (1151 lines, 52K chars)

---

Executive Summary

Critical Gaps Found:

1. Sparsity σ=0.04 — Code uses σ=0.12 (default) or σ=0.10 (feature flag), not the claimed 4%.

2. Predictive scoring weights — Code uses 0.10/0.65/0.10/0.45; whitepaper specifies 0.20/0.55/0.15/0.20.

3. Epistemic immune system — Component 1 (Dynamic Calibration) missing; Component 3 (Three-Angle Protocol) incomplete.

---

Contribution-by-Contribution Audit

1. Sparse ternary {-1,0,+1} semantic encoding

Status: PARTIAL ⚠️

Evidence:

• src/core/sparse_vec.rs:12-13: pub const SPARSITY: f32 = 0.12; (default)
• src/core/sparse_vec.rs:10-11: #[cfg(feature = "sparsity_010")] pub const SPARSITY: f32 = 0.10;
• Whitepaper Section 4: "σ=0.04 (4% active)"
• Whitepaper Section 17: "Exactly 4% (~655 active dims) — ternary τ operator enforces σ=0.04 at encoding time"

Impact: Higher sparsity (12%) increases expected random dot-product overlap from ~25.6 std to ~44.3 std, reducing effective 3σ isolation capacity from ~43,000 to a lower bound. The Boid engine's apply_to_universe hardcodes 4% requantization (target_nnz = DIM * 0.04), but this only affects post-flocking positions, not the original encoded vectors.

---

2. Three-layer encoding with 6-tier entity weighting

Status: PARTIAL ⚠️

Evidence:

• src/core/sparse_vec.rs:223-398: Five layers implemented (trigrams, word-hash, word-bigrams, char-bigrams, char-4grams)
• Whitepaper Section 4.3: "Three layers of features"
• Whitepaper Section 17: "Three: surface trigrams + entity-boosted word hashing + bigrams"

Impact: Additional character-level layers improve surface-form matching but increase the accumulator's nnz before thresholding, potentially changing the statistical properties of the final ternary vectors.

---

3. Hybrid dual-channel retrieval scorer

Status: VERIFIED ✅

Evidence:

• src/core/universe.rs:1158-1159, 1207-1208, 1280-1285: All query paths implement the hybrid scorer
• Whitepaper Section 5: "0.6×cosine + 0.4×keyword_overlap"

---

4. Confidence step-function amplification

Status: VERIFIED ✅

Evidence:

• src/core/universe.rs:967, 1018, 1158, 1207, 1280: All retrieval functions implement the step-function
• Whitepaper Section 5.2: "strength_bonus jumps 0.50 → 0.85 at conf≥2.9"

---

5. Structured epistemic cell (Claim object)

Status: VERIFIED ✅

Evidence:

• src/core/claim.rs:4-46: Full Claim struct
• src/core/universe.rs:70-87: Cell struct with all metadata fields
• Whitepaper Section 11: Complete Cell specification

---

6. Fibonacci torsion / golden phase angle

Status: PARTIAL ⚠️

Evidence:

• src/core/sparse_vec.rs:803: const GOLDEN_ANGLE: f32 = 2.399_963_1_f32;
• src/core/sparse_vec.rs:800-833: project_vogel_spiral() uses golden angle for cell placement
• Whitepaper Section 6.3: "θ = (pos × α_g) mod 2π; phasor_coherence = cos×cos(Δθ)"
• Whitepaper Section 17: "Phase geometry in HD memory — first in any AI architecture"

Impact: The golden angle is used for spatial initialization but NOT for retrieval scoring. The whitepaper's claim that "phasor coherence" improves retrieval precision is not implemented in the query path.

---

7. SpiralState golden-ratio temporal oscillator

Status: VERIFIED ✅

Evidence:

• src/core/spiral.rs:17-24: All constants match whitepaper exactly
• src/core/spiral.rs:52-84: tick(), recompute(), tau_r() implementation
• Whitepaper Section 7: "b = ln(φ)/(π/2) = 0.306349, Δθ = 0.05/tick, τ_R ∈ [0.5, 1.0]"

---

8. Boid flocking in D=16,384

Status: VERIFIED ✅

Evidence:

• src/core/boid_engine.rs:15-38: All constants match
• src/core/boid_engine.rs:124-252: Full run_boid_iteration() with all safeguards
• Whitepaper Section 8.3: "sep=1.5, align=1.5, cohere=1.5; anchor immunity ≥3.5"

---

9. Continuation vector (dual-vec cell)

Status: VERIFIED ✅ (structure) / PARTIAL ⚠️ (scoring weights)

Evidence:

• src/core/predictive.rs:12-15: Comments claim 0.55 * predictive_match
• src/core/universe.rs:2382: Actual code: let score = 0.10 * sim + 0.65 * predict_match + 0.10 * mh - 0.45 * rec;
• Whitepaper Section 5.3: "0.55·continuation_match"
• Whitepaper Section 17: "0.55 weight in predictive score"

Impact: The continuation match is overweighted (0.65 vs 0.55), and similarity is underweighted (0.10 vs 0.20), making retrieval more trajectory-dependent than specified.

---

10. ConversationTrace HD working memory

Status: VERIFIED ✅

Evidence:

• src/core/predictive.rs:37-68: Full ConversationTrace implementation
• Whitepaper Section 9.3: "permute-bundle rolling accumulator; VSA residual stream"

---

11. Four-component epistemic immune system

Status: PARTIAL ⚠️

Evidence:

• src/core/universe.rs:1625-1664: FID scan fully implemented
• src/core/universe.rs:1746-1802: ingest_and_verify() exists but only calls should_accept_claim()
• src/core/universe.rs:1910-1947: should_accept_claim() checks confidence floor, evidence, and basic contradiction — no Three-Angle Protocol
• Whitepaper Section 10.2: FID algorithm matches exactly
• Whitepaper Section 10.3: Three-Angle Protocol with PHYSICS_RESONANCE_FLOOR=0.55, COHERENCE_FLOOR=0.40 — these constants are mentioned in seed data but NOT used in ingest_and_verify

Impact: The epistemic immune system is missing its most sophisticated component (dynamic calibration) and its primary gate (Three-Angle Protocol) is reduced to a simple contradiction check. This significantly weakens the system's ability to reject false claims.

---

12. Native multi-agent shared lattice

Status: VERIFIED ✅

Evidence:

• src/core/claim.rs:48-52: Layer constants including LAYER_CELLULAR = 2
• src/core/boid_engine.rs:191-194: Per-user isolation in Boid engine
• src/core/universe.rs:2314-2318: Per-user query filtering
• Whitepaper Section 12: "7 topological regions with distinct trust profiles"
• Whitepaper Section 14.2: "Discord channels and roles define which agents speak where"

---

13. Explicit SynapticLayer with Hebbian LTP/LTD

Status: VERIFIED ✅

Evidence:

• src/core/synapse.rs:44-63: All constants match whitepaper exactly
• src/core/synapse.rs:115-142: record_co_firing() with full neuromodulator gating
• src/core/synapse.rs:318-391: NeuralBus::query_associative() implements stages 1-3
• Whitepaper Section 17: "BASE_LTP=0.035, chi_gate=1−χ×0.8, dopamine×0.8, phi_g×0.5; LTD_IDLE=80 ticks; fan-out=32; 8192 synapse cap"

Note: The 12-step NeuralBus is documented as comments but not executable as a single pipeline. Engine::query() implements stages 1-4; run_homeostasis_cycle() implements pruning + Boid (stages 10-11). Stages 5-9 and 12 are distributed across the engine tick.

---

14. Five-layer Scale Manager

Status: VERIFIED ✅

Evidence:

• src/core/scale_manager.rs:6-62: All 5 layers with distinct parameters
• src/core/boid_engine.rs:163-170: Maturation/degradation logic
• Whitepaper Section 17: "Quantum/Syncytium/Cellular/Organ/Body; per-layer speed, decay, replenish, neighbor radius; automatic maturation/degradation transitions"

---

Additional Findings

A. Predictive Scoring Formula Discrepancy

The whitepaper and code comments in predictive.rs specify:

0.20 * sim + 0.55 * predict_match + 0.15 * mh - 0.20 * rec

The actual implementation in universe.rs:2382 uses:

let score = 0.10 * sim + 0.65 * predict_match + 0.10 * mh - 0.45 * rec;

This shifts emphasis from raw similarity (-0.10) and consensus (-0.05) toward trajectory prediction (+0.10) and anti-recency (-0.25). The comment in predictive.rs is misleading and should be updated to match the code, or the code should be restored to the documented weights.

B. Encoding Sparsity Discrepancy

The whitepaper claims σ=0.04 (4%, ~655 active dims) throughout. The code uses σ=0.12 (12%, ~1966 dims) by default. The Boid engine's apply_to_universe does requantize to 4% (target_nnz = DIM * 0.04), but this only affects cells AFTER flocking, not the original encoding.

Recommendation: Either update the whitepaper to reflect 12% sparsity, or adjust SPARSITY to 0.04 and re-validate retrieval performance.

C. Missing ingest_and_verify Three-Angle Protocol

The whitepaper describes a rigorous protocol:

1. Angle 1: Query lattice for positive evidence

2. Angle 2: Query lattice for contradicting evidence

3. Angle 3: Compute domain resonance score

4. Physics resonance floor (0.55) for established-physics claims

5. Coherence floor (0.40) for all claims

The actual ingest_and_verify():

1. Creates a temporary claim

2. Calls should_accept_claim() — which only checks: confidence ≥ 0.45, evidence non-empty, no direct contradiction

3. Does NOT query the lattice for corroboration or contradiction

4. Does NOT compute domain resonance

5. Does NOT enforce physics/coherence floors

Recommendation: Implement the full Three-Angle Protocol in ingest_and_verify().

D. Missing Dynamic Calibration

Component 1 of the epistemic immune system ("Dynamic Calibration") is entirely absent. There is no mechanism that tracks retrieval accuracy and adjusts confidence thresholds.

Recommendation: Implement a calibration tracker that monitors query→response accuracy and adjusts effective confidence thresholds per region.

E. Phasor Coherence Not Used in Retrieval

The whitepaper claims Fibonacci torsion "improves retrieval precision" via phasor coherence. The project_vogel_spiral() function uses the golden angle for spatial placement, but no retrieval function (query(), predictive_query(), cosine()) uses phase-angle-based coherence scoring.

Recommendation: Either remove the retrieval-precision claim from the whitepaper, or implement phasor_coherence = cos(θ_q) * cos(θ_cell) as an additive term in retrieval scoring.

---

Summary Table

---

*End of Audit Report*