Maya-Sutra-P5: The Floor and the Cost of Forgetting — Quantisation Limits of On-Chip Weight Decay, an Upper Bound on the Energy of Erasure, and the Absence of Habituation on BrainScaleS-2

Maya-Sutra P5: The Floor and the Cost of Forgetting stops at the number Paper 2 did not explain: on-chip integer weight decay lets go earlier than its floating-point model. The construct is Nirveda (निर्वेद; the fading of response to the unchanging). Three questions are asked of the same fixed six-neuron BrainScaleS-2 rig. The floor (B1): the decay kernel (w·F)>>8 discards its fractional remainder; as F approaches 256 the ideal per-event decrement falls below one LSB while truncation continues to subtract a full one — at F = 255, exactly 1 LSB per event, so a weight of 60 is silent by k = 35 on every one of five independent nightly calibrations (exact sign test, one-sided p = 0.031; worst day-to-day spread one spike on a 3-rep mean across all 40 conditions). Carrying the discarded remainder stochastically (on-PPU xorshift32) restores the ideal exponential 60·(F/256)^k: still 3.1 pops at k = 60. The arithmetic cross-check closes the loop — weight 40 reached by three different routes (deterministic k = 20; stochastic k = 100; F = 253 stochastic k = 35) yields the same 2.0-pop response. The cost (B2): an interleaved decay-vs-no-op differential on the board’s INA219 rail (216 samples per arm per day, six days) does not clear the pre-registered significance bar — the running t reached 6.39, and the next run pulled it to 2.19 against a required 4.77. The pre-committed stopping rule, declared blind, is why that reversal is in the paper instead of a false measurement. The result is an upper bound: energy per bit erased < 1.8×10⁻¹¹ J ≈ 6.4×10⁹ × kT ln 2 — an engineering overhead factor; the Landauer limit itself sits 19 orders of magnitude below this instrument. The unbidden fading (Part A): the substrate does not habituate. Across a 20+-condition battery, two candidate decrements (5.8%, 7.4%) failed same-day replication and are withdrawn; spike-frequency adaptation engages tonically — present from the first presentation, flat thereafter — which is precisely not habituation’s hallmark H1 (Rankin et al. 2009). Also in this paper, kept deliberately: a “decay-event ceiling” recorded on 3 August was the experiment runtime silently dropping plasticity events — identified by the platform team on the support ticket, fixed, empirically confirmed (every ceiling cell went to exactly 0.0), and retracted in full with the pre-fix data shown. Pre-registered predictions (29 July, before first silicon) carry their verdicts in the repository, including one falsified by simulation before any chip time. Scope-lock per Maya-Meta P2: reproducible signatures within THIS architecture under THIS protocol — Property, not Law; one chip, one kernel, five calibrations for B1, six for B2, one for Part A’s full battery. Series: Part of the Maya-Sutra Series — Nexus Learning Labs’ hardware-native track, characterizing neuromorphic substrates directly (one substrate, one honest characterization at a time). The arc: let go (P1) → forget (P2) → remember (P3) → listen (P4) → the floor and the price of it all (P5). Research characterization only. Links: GitHub Repository (private — to request access: email research@nexuslearninglabs.in with subject "Code Access Request — Maya-Sutra-P5" and your research context) | Full Series Index — nexuslearninglabs.in Nexus Learning Labs, Bengaluru · UDYAM-KR-02-0122422 · BHASKAR IN-0526-9452JS · ORCID: 0000-0002-3315-7907 · VAIRAGYA_DECAY_RATE = 0.002315 (embedded in all canonical hyperparameters) · Canary: MayaNexusVS2026NLL_Bengaluru