Get started
Work

Maya-Sutra P8 — Maryādā The Open Door

Year: 2026
Type: preprint
Abstract: Maya-Sutra P8: The Open Door — with the escape from an aversive boundary no longer punished, no instance of a host-in-the-loop BrainScaleS-2 agent was captured at the boundary, while its learned set-point and the direction of its substrate decomposition held. Preprint, version 1.0; not peer reviewed. Background. In a previous study on BrainScaleS-2 (Maya-Sutra P7, DOI 10.5281/zenodo.23551942), a host-in-the-loop agent learned to avoid the two aversive tiles at each end of a ten-tile strip, and 6 of 30 learning instances were instead captured at a boundary. The proposed cause was that the boundary is two tiles deep while the rule credits one step, so the only move out of the outermost tile is punished. That explanation had not been tested. Methods. On the same chip (hxcube7fpga0chip57_1), neuron placement and calibration file as P7, the P7 agent was run unchanged except for one function: in the open-door world a move towards the centre is never punished, which changes the reward of exactly two of the twenty possible moves (1→2 and 10→9, from −1 to +0.1). A full two-by-two design crossed the world with P7's three substrate manipulations (output populations mirrored, weights transplanted into the mirrored configuration, afferents exchanged), each seed's arms in one session. Thirteen predictions were registered in a pre-registration locked on 2 November 2026 before any learning instance ran; 48 learning and 12 transplanted instances followed in five sessions on 3–7 November 2026 (15,120 hardware steps). Results. In the open door 0 of 12 instances were captured, against 3 of 12 in the P7 world on the same days (the registered bars; Fisher p = 0.22, post hoc); pooled over three learning arms per world, 0 of 24 against 7 of 24 (p = 0.0094, post hoc; the arms share seeds). Every open-door instance completed memory with zero aversive steps. The set-point was 6.50 in both worlds, as in P7. In both worlds mirroring moved the median set-point by 0.10 tile, transplants rested at least 1.0 tile colder than their source in every eligible pair, and the afferent exchange moved the set-point 1.825 and 1.875 tiles colder. A step-exact decomposition (post hoc) shows that in both worlds the comfort bonus, not the penalty, built most of the learned centre-seeking weight contrast. All 13 pre-registered criteria were met; two at their limits. Conclusions. On this substrate, replacing the escape penalty with the bonus was sufficient to prevent capture in 24 instances; the registered reading is that the punished escape is necessary for capture here. P7's account of capture is supported; its account of how compensation arises is not supported as written. Scope per Maya-Meta P2: one chip, one placement, one calibration, five days — a property of this substrate, not a law of the hardware. Research characterisation only; nothing here is a medical device or medical advice. Series. Maya-Sutra, Paper 8: the hardware-native track characterising neuromorphic substrates directly, one substrate at a time. Let go (P1) → forget (P2) → remember (P3) → listen (P4) → the floor and its cost (P5) → carry (P6) → settle (P7) → keep the limit (P8). Data and code. The five session ledgers and archives (CC0 1.0), the frozen agent, surrogate, notebook and scoring script, the analysis and figure scripts, a twelve-check reproduction test and an 80-check verification of the paper's numbers (MIT), and the pre-registration with its adversary record (CC BY 4.0) are held in the repository github.com/venky2099/Maya-Sutra-P8 (private; access on request to [email protected] with the subject "Code Access Request — Maya-Sutra-P8"). Every number and figure regenerates from the ledgers. Venkatesh Swaminathan · ORCID 0000-0002-3315-7907. This is independent research conducted at Immanis Praeses Niralamba Learning Labs, Bengaluru, alongside M.Sc. study at BITS Pilani. It continues the work published as Nexus Learning Labs, 2025–2026. © 2026 Immanis Praeses Niralamba Learning Labs Private Limited · CIN U72100KA2026PTC227157 · Paper and figures CC BY 4.0 · Public canary: MayaNexusVS2026NLL_Bengaluru
DOI: 10.5281/zenodo.24070052
Language: English
Open Access status: green
OpenAlex ID: W7225661119
Cites:
Cited by:
Related to: