The Forge

Feed it provider API keys. Get back a ranked catalog of the datasets actually available to you, plus a slate of strategy candidates built on top of them.

Overview

Most quants spend more time hunting for usable data than on alpha: provider APIs have inconsistent coverage, undocumented gotchas, and uneven quality across asset classes. The Forge encodes that exploration as a pipeline of cooperating LLM agents, each with a narrow job.

The pipeline runs once per scan and produces three artifacts:

1. A catalog of datasets discovered across your active providers, each scored 0–100 on six quality dimensions and assigned a verdict.
2. A list of research plans: hypotheses the Planner agent proposed by combining your useful datasets.
3. A set of strategy candidates, each one a runnable backtest with metrics.

None of this auto-deploys. The whole point is to surface options for you to review.

Agents

Four cooperating agents, each with a narrow job.

Catalog

The catalog is a per-install table of every dataset the Forge knows about. Each row is a (provider, capability, asset_class) triple with quality scores, a verdict, and links to provenance.

The six quality dimensions

Each dimension is scored 0–100 by the Probe agent.

The aggregate score

The six dimensions combine as a geometric mean, not arithmetic, so a zero in any single dimension clamps the aggregate to ~zero. A polygon endpoint that's 90/100 on five dimensions but 0/100 on coverage isn't "average". It's unusable, and the verdict reflects that.

# From auracle/forge/catalog.py
# Floor the zeroes at 1 so log() is defined.
floors = [max(d, 1) for d in dims]
score  = math.exp(sum(math.log(d) for d in floors) / len(floors))
# Score == 1 if any dimension is 0; ~90 if all are 90.

Verdict thresholds

The aggregate score maps to a verdict by these thresholds:

Research

A research plan is a structured hypothesis. The Planner emits them as JSON in forge_research_plans.plan_json; you can also write them by hand.

{
  "hypothesis": "OI divergence on Coinalyze precedes 4h mean reversion in BTC-USD",
  "inputs": [
    { "dataset_id": 12, "alias": "oi"   },
    { "dataset_id":  7, "alias": "bars" }
  ],
  "feature_spec": {
    "oi_change_24h": "oi.delta(24h) / oi.rolling(7d).mean()",
    "ret_4h":        "bars.close.pct_change(4h)"
  },
  "label_spec":   { "horizon": "4h", "kind": "log_return" },
  "evaluation": {
    "walk_forward": { "train": "12mo", "test": "1mo", "step": "1mo" },
    "costs":  { "bps": 5 },
    "metrics": ["sharpe", "max_dd", "hit_rate"]
  }
}

Plans can be queued, paused, retried, or rejected. The Planner won't re-emit a plan it already proposed in the past 30 days (content-hash dedupe).

Candidates

A strategy candidate is the output of running Synthesis on one plan. It lives in forge_strategy_candidates with:

• A pointer to the source plan and the inputs used.
• The generated Python file path under strategies/_forge/.
• The backtest result blob (equity curve, trades, per-period metrics).
• A score against the plan's metric targets, plus a Planner-summarized verdict.

In the web UI, candidates appear in a table sorted by Sharpe. Clicking one opens the candidate review surface: equity curve, trade list, generated code, and a Promote button. Promotion copies the file out of _forge/ into your active strategies directory; only then can it be deployed.

Limits & audit

The Forge spends money (LLM tokens) and writes code (synthesized strategies). Both are bounded.

Cost caps

• Per-scan: hard cap of $5 USD of LLM spend by default. Configurable in auracle forge config.
• Per-day: $25 USD aggregate across all agents.
• Per-dataset probe: $0.50 + 2 minutes of API time.

Hitting any cap pauses the scan; the partial catalog is preserved.

Audit trail

forge_agent_runs records every agent invocation: which agent, which dataset, input fingerprint, token usage, wall-clock time, output fingerprint, and the install_uuid. Replay a scan by re-running the same input fingerprints; the output should hash-match (modulo non-determinism flagged at the LLM call boundary).

Promotion gate

Generated strategies are sandboxed and never auto-deployed. You review the code, the backtest, and the metrics before Promote moves the file into your active strategies directory; only then can it be deployed live.

Costs

Default budgets per agent (configurable per install).

Cost is logged in forge_agent_runs.cost_usd. Aggregate with a direct SQL query (SELECT SUM(cost_usd) FROM forge_agent_runs WHERE created_at > now() - interval '30 days') until the auracle forge costs CLI ships in v1.1.