Maya-Mimamsa P1: Apurva A Reward-Receipt-Gated Eligibility Trace for Delay-Invariant Credit Assignment in Spiking Neural Networks

Version 2 (2026-08-04). Revised to the NLL Universal Paper Format v6. The original text is retained in full; nothing has been deleted. Corrections appear as marked blocks placed at the section that carries the claim, and each one states what the paper said, what the data show, the corrected claim, why it happened, and what still stands. This version narrows the novelty claim; every result stands.An accumulating TD(λ) trace at γλ = 1 is a non-decaying trace, and replacing/reset traces are exactly consume-on-update, so Apurva is an accumulating trace at λ = 1 with a multiplicative reset on reward — the claim that no existing mechanism does this is withdrawn, and Sutton & Barto, Singh & Sutton (1996), Izhikevich (2007) and Gerstner et al. (2018) are now cited.The experimental design is the best in the corpus and is untouched: four falsifiers committed before results and run, an integrity ablation that could have killed the claim, a load-bearing κ sweep, and two honestly reported losses. The e-prop characterisation as τ-locked is correct. Everything below this line is the original description from version 1. It is retained unchanged for the record. Where it conflicts with the corrections above, the corrections stand. Two framings it repeats have since been withdrawn in full — the Bhaya Quiescence Law and the Buddhi S-Curve. Both are addressed in Maya-Meta P1, now superseded, and in the self-audit of Maya-Meta P2. This paper introduces Apurva, a new type of eligibility trace for spiking neural networks and reinforcement learning in which the credit assigned to an action does not fade with time but instead remains banked in full until a reward actually arrives, at which point it is consumed at a rate governed by a parameter called kappa — a mechanism directly inspired by the Pūrva-Mīmāṃsā philosophical concept of apurva, the unseen potency an enjoined act carries until it bears its fruit. The central motivation is to address the well-known failure of conventional time-decay traces (of the TD(λ)/e-prop variety) under delayed rewards, where the causal link between action and outcome is stretched across many timesteps. On a delayed-reward cry-selection task run across 10 random seeds and two reward regimes, Apurva maintained flat evaluation accuracy of 0.69 (deterministic) and 0.70 (probabilistic) across action-reward delays from 0 to 128 steps, while the time-decay baseline collapsed toward chance by delay 64, with Apurva exceeding the baseline at the longest delay in all 10 of 10 seeds. The work is presented with explicit reporting of limitations and is the first paper in the Nexus Learning Labs Maya-Mīmāṃsā series, a line of research examining learning mechanisms through the lens of classical Indian epistemic and ritual philosophy. Series: Links: GitHub Repository (private — to request access: email research@nexuslearninglabs.in with subject Code Access Request — Maya-Mimamsa-P1 and your research context) | Interactive Dashboard | FAQ | Full Series Index — venky2099.github.io Nexus Learning Labs, Bengaluru · UDYAM-KR-02-0122422 · BHASKAR IN-0526-9452JSORCID: 0000-0002-3315-7907 · VAIRAGYA_DECAY_RATE = 0.002315 — an ORCID-derived provenance mark, not an experimental parameter. Each paper's disclosure block states whether it reached a result in that paper.Canary: MayaNexusVS2026NLL_Bengaluru