Official Indian Patent Specification — Patent Office, Government of India Statutory Form: FORM 2 — THE PATENTS ACT, 1970 (39 of 1970) & THE PATENTS RULES, 2003 (Provisional Specification). Title of Invention: SYSTEM AND METHOD FOR CONTINUOUS NEUROMORPHIC DELIBERATION GATING, LOCAL RECURRENT ATTRACTOR DISCORDANCE RESOLUTION, AND ASYMMETRIC HARM SHIELDING. Applicant & Inventor: Infant Nirmal Priyan J, Adrient Pvt Ltd, Bangalore, Karnataka, India. Official Filing Date: 18th day of September, 2026. 1. Abstract & Dual-Tier Cognitive Architecture Conventional sequence architectures think at uniform computational speed: every output token costs exactly one forward pass, regardless of whether the problem is trivial or profoundly complex. When auxiliary reasoning modules or chain-of-thought methods intervene, they suffer from the Deliberation Harm Problem — destroying more correct reflex predictions than they rescue. This invention presents a dual-tiered neuromorphic architecture: Tier-1 (Continuous Recurrent Reflex Substrate): 5,120 leaky continuous-time cells executing O(1) step updates. Tier-2 (Deliberation Cortex & Episodic Transition Memory): Activated conditionally by top-2 decision margin thresholding, governed by an Asymmetric Harm Shield that guarantees deliberation produces net accuracy improvements. 2. Core Mathematical Formulations Formula 1 — Continuous Margin Deliberation Gate Δ_t = z_t[(1)] - z_t[(2)] if Δ_t ≥ τ_margin → Bypass Deliberation Tier [Measured Bypass Rate: 81.26%] When the top-2 reflex logit margin is wide, reflex confidence is mathematically proven sufficient. Over 80% of steps execute without deliberation overhead. Formula 2 — World-State Fork-Score-Commit fork: z_k = z(t), k = 1..K rollout: z_k ← Pulse.step(z_k, hypothesized_event) score: V(z_k) = value_head(z_k) commit: answer = decode( argmax_k V(z_k) ) When margin Δ_t < τ_margin, the model forks its continuous latent world-state into K parallel rollouts across CPU cores, selecting the highest-value trajectory. Formula 3 — Asymmetric Harm Shielding Rules Switch to c* iff: 1. Shortlist: c* ∈ Top-4(z_t) 2. Discordance: c* ≠ y_reflex(t) 3. Dominance: D(c*) ≥ 0.50 && N[context] ≥ 5 The harm shield prevents false-positive interventions, delivering an empirical rescue-to-break ratio of 5.24 : 1. Formula 4 — Zero-Parameter Skill Loom Routing signature = substrate.probe(event) expert = argmax_e( confidence(e, signature) ) [Zero extra gating parameters] Detachable expert packs (Code, Indic 22 languages, Math, Vision) dock into substrate spare capacity, routed directly through the surprise signature. 3. Empirical Proof of Efficacy (5-Seed Confirmation Panel) In pre-registered confirmation across 5 independent random seeds (42, 123, 456, 789, 1000) with 10,000 scored decisions on complex second-order sequences: Experimental Parameter Measured Result Statutory Specification Target Statistical Significance Net Accuracy Gain +3.73 percentage points ≥ +2.00 pp threshold 95% CI: [+3.27%, +4.18%] Rescue-to-Break Ratio 5.24 : 1 (461 Rescues / 88 Breaks) ≥ 4.00 : 1 floor p = 5.468 × 10⁻⁶² (McNemar χ² = 252.07) Margin Compute Bypass Rate 81.26% of steps bypassed ≥ 70.00% bypass Low-power CPU preservation Parent Substrate Invariance Bit-exact identical phase space Zero trajectory perturbation Non-destructive side-channel inspection
