name	Atlas README
description	Autonomous scientific-method research loop that builds a causal graph of validated market claims via pre-registration, walk-forward OOS, and a code-enforced promotion gate
type	reference
updated	2026-07-26

Atlas

What is this? Atlas is an autonomous scientific-method research loop. It generates falsifiable hypotheses, pre-registers a falsification criterion and significance threshold, runs walk-forward out-of-sample experiments, records typed evidence, and makes an explicit promote / kill / continue / pivot decision — building a causal graph of validated claims (not free-text associations). The core is domain-agnostic; the current application is crypto markets (Bitstamp hourly BTC/ETH/SOL). It runs unattended as a systemd service; the CLI exists for development and debugging. It contains no in-product LLM code.

Lifecycle. active (deployed). This is the one canonical implementation; there is no successor. An earlier, unrelated differentiable-programming exploration, crypto-agent-dp-lab, is an archived predecessor, not a successor—its thesis was never validated.

What is verified today — and what is not claimed.

• Verified: 199-test hermetic suite green; the promotion gate, walk-forward math, statistical tests, the forward-prediction ledger/scorer, and the S3-P2 escalation gate all have coverage; the loop runs autonomously and scores its forward predictions unattended.
• Not claimed: any validated market edge. The promotion gate has never fired — 0 promoted primitives, 0 causal edges, 69 refuted nodes. The only scientific output to date is negative knowledge (falsifications and confirmed-null forward windows). Backtests model a 26 bps fee but ignore slippage/funding/liquidity: backtest ≠ live. See docs/CASE_STUDY.md.

Fastest meaningful check.

make setup    # create/refresh .venv (Python 3.11+)
make check    # lint + hermetic tests + repo/clean/prompt validation

make help lists all targets. Tests need no network or credentials.

Where to go deeper.

• docs/architecture.md — composition roots, dependency direction, artifact roles, runtime paths, deployment, and agentic-safety posture.
• docs/CASE_STUDY.md — outsider-legible, truthful account of what has and hasn't been demonstrated.
• docs/lineage/crypto-agent-dp-lab.md — why the earlier differentiable- programming prototype remains a separate, unvalidated archive.
• AGENTS.md — concise agent/contributor instructions; CLAUDE.md — full governance and promotion-gate rules.
• repo.toml — shape/lifecycle/risk declaration (ADR-0050).

Repository layout

• src/atlas/ — library + CLI: models/ (domain), storage/, analysis/, data/, generation/, adapters/, plus runner.py (the loop) and cli.py.
• tests/ — hermetic pytest suite.
• deploy/ — systemd unit (atlas-runner.service).
• findings/, .reviews/ — historical research log and adversarial reviews.
• scripts/ — operator tooling (check_repo.py, deploy-check.sh, migrate_claim_hash.py) and older one-off research scripts.
• Runtime state (.atlas/, sessions/, reports/, data/, root *.jsonl) is gitignored and created on use. graph/causal_graph.json is tracked and rewritten by the live runner — see docs/architecture.md.

Promotion gate

A claim becomes a promoted reasoning primitive only with ≥2 strong evidence records from distinct experiments, at least one out-of-sample or live, meeting the pre-registered threshold, with no unaddressed strong contradiction. Enforced in code. Pre-registered fields are immutable.

Suggested GitHub metadata

• Description: Autonomous scientific-method research loop building a causal graph of validated market claims (pre-registration, walk-forward OOS, code-enforced promotion gate).
• Topics: causal-reasoning, scientific-method, hypothesis-testing, pre-registration, walk-forward, falsification, autonomous-agent, quant-research, python.