Vaanam / APM
Ancient Prior Model — Tamil Grammar Priors for Stellar Variability
What It Is
A transformer pipeline that classifies TESS stellar variability using Tamil-grammar-derived attention biases as a structural prior. The Ancient Prior Model (APM) tests whether canonical grammar tokens, cyclic positional encodings, and co-attestation-based attention bias improve few-shot variability classification on real astrophysical light curves. Targeted at EAS 2026.
Most "domain prior" research in ML hand-waves about inductive bias. APM tests, on real public TESS data, whether a non-numerical canonical prior — Tamil grammar — actually moves few-shot classification metrics versus matched controls. Pre-registered evaluation, fixed seeds {42, 137, 256}, deterministic reproducibility, prereg-hash-stamped JSON eval outputs.
What I Built
- TESS light-curve ingestion pipeline (MAST) with ecliptic-longitude / Nakshatra-cell / source-hash metadata derived once per record and reused deterministically
- Three matched track implementations — Track A (patch tokens + RoPE baseline), Track B (Track A + grammar token + cyclic PE), Track C (Track B + additive attention bias from grammar co-attestation matrix)
- Grammar bias builder: deterministic co-attestation matrix from `data/grammar/cells/` with schema validation; corpus files never edited in-place
- Pre-registered evaluation: hash-stamped eval JSON, schema validation, prereg-freeze check via `vanam` CLI
- Negative-control track with seed 9999 permutation to refute the "any prior helps" alternative
- Public-data Sprint 1 pilot — 5k-light-curve target on the Wang2025 catalog, all artifacts script-generated from committed JSON
Tech Stack
Sprint 1 in progress (Track A pilot extended past 224 stars on Wang2025). Local-hardware-only by design. Honest reporting of positive, null, or partial outcomes. Apache 2.0.