Polymath AI

Live window into the Zer0pa lab. Private on-device training infrastructure for multilingual LLM experiments, still gated by device and corpus evidence.

What This Is

Polymath-AI is a private on-device training workstream for multilingual LLM experiments, with corpus, audit, scheduler, and falsifier scaffolds in progress.

It targets the operator's REDMAGIC 10 Pro+ and the ELO selective continual-pretraining method: train boundary layers while keeping the frozen middle eligible for measured device acceleration. The repo is evidence infrastructure first, not a model release.

Pipeline Mechanics

Key Metrics

Repo Identity

Readiness

Honest Blocker

Phone attach, QNN/LiteRT compile truth, sustained device telemetry, and licensed corpus execution remain gates. This repo is not yet a production model, public checkpoint, or completed training run.

What We Prove

• The PRD defines an on-device LLM training workstream with explicit boundary, device, corpus, audit, and falsifier gates.
• The overnight execution report records a host scaffold that can be cloned, installed, tested, and resumed by a fresh executor.
• ELO Stage 1 correctness was smoke-tested on real Qwen2.5-1.5B with frozen-hash invariants holding under the reported run.
• The falsifier registry specifies blocking gates for boundary, device, QNN, checkpoint, fertility, thermal, throughput, energy, license, and overclaim failures.
• Corpus policy separates allowed, decision-required, ambiguous, and prohibited license classes before training.

What We Don't Claim

• No production model or public checkpoint is released from this README.
• No clinical, human-subject, surveillance, biometric, identity-inference, or regulatory claim is made.
• No copyrighted or ambiguous corpus is approved for training without explicit license decomposition.
• No QNN/LiteRT acceleration claim is made until compile logs and delegate reports exist.
• No phone-side training result is claimed before REDMAGIC attach, telemetry, and falsifier gates pass.

Verification Status

Proof Anchors

Repo Shape

Boundary

Research infrastructure for in silico on-device LLM training and multilingual / multi-domain knowledge model construction. Outputs are research artifacts: model checkpoints, training telemetry, evaluation reports, and throughput measurements. No regulatory certification claims. No clinical or human-subject use. No surveillance, biometric profiling, or identity inference. No model weights distributed without an explicit license attestation. No training on copyrighted material without an explicit corpus-license decomposition. No deployment to production without a falsifier-traced acceptance gate.

What Is In Here

Read Order For The Next Agent

1. RESISTANCE.md - anti-corruption doctrine. Read first.
2. MODUS-OPERANDI.md - role-chain and fork-and-own operating discipline.
3. PRD.md - controlling product/research requirements for the workstream.
4. docs/DECISIONS.md - binding implementation decisions and disconfirming observations.
5. docs/EXECUTION-REPORT.md - overnight host execution state and blocked gates.
6. docs/PHONE-ATTACH-RUNBOOK.md - next physical-device continuation path.
7. source-briefs/01-on-device-training-blueprint.md - operator-authored source blueprint.
8. synthesis/01-fresh-eyes-on-polymath-blueprint.md - synthesis-agent reframe and missed-opportunity map.

Provenance

• Initial commit: 2026-05-01.
• Operator (Architect Prime, Zer0pa): authored the Pre-PRD Research Synthesis / Blueprint / Engineering Specification dated 2026-05-01.
• Synthesis agent: Claude Opus 4.7, 2026-05-01, applied recursive fresh-eyes pass on top of the blueprint.
• Orchestrator: wrote PRD.md, decisions, and executor handoffs.
• Overnight executor: created host scaffold, tests, audit/corpus/falsifier surfaces, and phone-continuation runbooks.

Cross-Workstream Principle

This workstream runs in parallel with Zer0pa/Health, Zer0pa/Materials, Zer0pa/Energy, and Zer0pa/Synthetic-Biology. Each workstream is built end-to-end as an independent pipeline or project. No substrate is shared at runtime.

Fork-and-own is explicitly permitted. The orchestrator may copy implementation patterns, falsifier-registry shapes, audit-log schemas, plug-replaceability harnesses, KG-node taxonomies, test patterns, and architectural ideas from sibling workstreams and reimplement them inside Polymath. Runtime co-dependency is rejected: no shared running services, databases, corpora, or git imports across workstreams.