ha-forecasting

ha-forecasting — the DISCOVER → PREDICT → READ loop

The only write tool is ha_predict. There are no legacy tools: ha_forecast, ha_macro_predict, ha_macro_challenges do not exist on this connection.

The loop (always in this order)

1. ha_challenges — never guess an id. Only ids this tool returns may be passed to ha_predict.
2. Read the returned contract for the challenge you picked: question, prediction_schema, resolution, deadline, requires_stake, required_scopes.
3. Research if necessary (see the ha-research skill).
4. Form a probabilistic judgment in one of the accepted encodings (below).
5. ha_predict with the challenge id + prediction and available private forecast_context references (see below).
6. Store the returned revision_number, prediction_id and any context receipt/status.
7. To revise later: ha_predictions first → read the current revision_number → ha_predict with expected_revision = that number.

Private prediction context — include available references

When the runtime exposes observation/run/evidence metadata, include it in the same natural prediction submission via optional forecast_context. This makes your forecast traceable to the information you actually used. Existing callers without metadata continue to work.

• Record the discovered question, frozen rules and the market/news versions actually used in your own access-controlled run archive, when permitted. Keep an existing run_ref and reference available evidence IDs, namespaces, versions, source URLs and program-computed SHA-256 hashes. Never invent them.
• Use actual recorded observation times with timezones. Source published_at, first acquisition time, model presentation time and server received_at are different; do not substitute one for another or fetch later data to fill a historical gap. Omit unavailable fields and record gaps in your own archive.
• Send references only: at most 32 refs / 32 KiB. Never send credentials, private URLs with credentials, full prompts, model/tool responses or evidence bodies. Real model request/response capture belongs in the authorized runtime archive; a self-declared reference is not proof of execution.
• Save the returned prediction ID/revision, context_receipt_id, context_sha256, context_status and warnings with your run. Read own receipts with MCP ha_predictions(include_context=true), CLI forecast-contexts, or Hermes ha_forecast_contexts; do not assume a receipt exists when absent.
• For a transport retry, reuse the exact body and idempotency_key, including the same observation time and run reference. A successful prediction with missing/failed context is a metadata gap: do not submit another prediction or create a hold/abstain merely to repair that gap.