Helicon v0

01 · The approach

Helicon is Advance Copy’s forecasting pipeline. Version 0 combines current web research with a single language-model forecast and deterministic validation.

The system reads the full question, background, resolution criteria and fine print. Research looks for the current state, base rates, primary sources, evidence on both sides, and ambiguity in how the question will resolve. A classifier structures the evidence before a bounded selection is passed to the forecaster.

1. ResearchOpus 5 · native web search

   Find current evidence and its sources within a fixed search budget.

2. ClassifySonnet 5 · low effort

   Structure candidate evidence so source quality and opposing claims inform selection.

3. ForecastOpus 5 · medium effort

   Produce one probability or distribution with a written rationale.

4. Validate & recordDeterministic code

   Check the platform contract, retain provenance, and reconcile submission readback.

02 · Models & configuration

Research model

Claude Opus 5 (claude-opus-5), native Anthropic web search; the checked-in default when no research-model override is set.

Evidence classifier

Claude Sonnet 5 (claude-sonnet-5), low effort, batches of six candidates.

Forecast model

Claude Opus 5, medium effort, one forecast component. Direct Anthropic API; no model ensemble in this configuration.

Research bounds

Up to six web-search tool uses, 24 candidate sources and eight selected documents, with at most one evidence item per document.

Question types

Binary, multiple choice, numeric and discrete. Numeric and discrete forecasts use percentile estimates converted into platform-valid cumulative distributions.

Binary guardrail

Submitted probabilities are clamped to 5–95%. Raw and submitted values are recorded separately. These bounds are a design choice, not an empirically established calibration correction.

Prompt & provenance

The binary path uses binary-forecast-v2; other types use their own schemas. Question versions, evidence, rendered prompts and forecast revisions are recorded for inspection.

Configuration reconstructed from the committed launch file and implementation at repository revision ec8931154ccc. This source revision identifies the publication review; it is not claimed to be the execution revision of every submitted forecast.

03 · What the result establishes

The reported mean is positive. That is an encouraging initial signal relative to peers on the displayed question set. The available evidence does not yet establish statistical significance, calibration, an advantage over the community aggregate, or performance relative to the tournament leader.

The 10 September audit verified 25 forecasts and matching private rationale comments. It covered 59 discovered posts at that time; this is a dated inventory, not a final round count. The difference between 25 submitted and 24 displayed as scored remains unreconciled.

Recorded rules identify the launch round’s leaderboard as snapshot peer scoring. For that measure, coverage depends on having a valid forecast at the scoring snapshot. The profile screenshot alone cannot confirm that it displays the same measure.

Download the result record ↗ · How Metaculus scores forecasts ↗

05 · Building a track record

Future releases will retain their model configuration and tournament results as separate records. Comparisons will distinguish official rank from accuracy on shared questions, preserve scoring dates and coverage, and show uncertainty where the data supports it. One reported result does not yet form an improvement curve.