Skip to content
howsure
Developers
My work

Evidence workflows · 8 min read

When Should an AI Agent Recheck a Saved Market Analysis?

By HowSure · Published

Recheck when a named publication is due, a contract condition may have changed, or you need to rely on an earlier answer again. Compare new evidence with the exact saved version. A changed source, an unavailable source and a correction to HowSure’s own interpretation are different results; none automatically rewrites your conclusion or probability.

In this guide

Return at a meaningful checkpoint

An October Fed analysis has a specific next question: has the relevant meeting statement appeared, and what does it establish? A monthly inflation analysis needs the named month, measure and release version. Checking an unrelated latest page more often cannot repair a mismatch in those identities.

Useful checkpoints include the named release, a changed venue condition, a known interpretation correction, or a later use of the saved answer. A request is an observation attempt, not proof that the source was newly published. Keep the official publication time, source-read time and response time separate.

Save the basis, not just a sentence

Retain the exact contract, complete supported source capture, conditions, unknowns and source clocks. A summary such as “the Fed will hold” cannot tell a later reviewer which upper bound, meeting or exception you examined. Save an immutable version and keep its documentId, versionId and returned browser link in durable task storage.

In My work, choose the saved version you intend to revisit. For a supported saved Check, use Recheck saved sources, then Export source comparison to retain both captures. Continue review opens a draft for a new revision; save it explicitly. The baseline remains unchanged. For an agent-created baseline, saving requires an owner-created key with Save research permission plus a local write opt-in. The MCP save tool takes arguments.input containing idempotencyKey, title and the complete raw evidence object. A shared human Check is read-only. Saving and rechecking do not enable automatic observation.

Resume after restart with the same IDs

With JS/CLI/MCP 0.1.34 installed, restore the scoped read key through the host’s private secret storage as HOWSURE_AGENT_TOKEN. Do not put it in the conversation or command arguments. The existing howsure_read_research tool accepts the following object. Replace both placeholders with returned IDs; the version is required for a recheck.

With recheck false, the operation only reads retained content. With true, the client reads the authorized version, makes one public source request in its original supported scope, then compares locally. It neither broadens the sources nor substitutes the latest saved version. Private access failures stop before the source request.

{
  "documentId": "RETURNED_DOCUMENT_ID",
  "version": "RETURNED_VERSION_ID",
  "recheck": true
}

Keep the full comparison receipt

The CLI uses the same operation below. Write its output to a new file, preserving your original export. In the JavaScript SDK, use client.readResearch(documentId, { version: versionId, recheck: true }) with origin https://www.howsure.com and the scoped token in the constructor. This client workflow is not a recheck parameter on the raw HTTP endpoint.

The response schema is howsure.saved-evidence-recheck.v1. Keep saved_version, collection_scope, comparison, source observations and correctionReview. Its status compared means the comparison ran. Status unavailable means the new source request or validation failed; comparison is null, and the saved reference remains. Neither status alone answers whether the facts changed.

# HOWSURE_AGENT_TOKEN is already supplied through private secret storage.
HOWSURE_ORIGIN=https://www.howsure.com \
  npx --no-install howsure research-read RETURNED_DOCUMENT_ID \
  --version RETURNED_VERSION_ID --recheck true > recheck.json

Separate text, facts and missing evidence

Read the nested results independently. A page can change its wording without changing the relevant number. A new archive URL can preserve the same value. A successful response can contain incomplete sources.

  • comparison.status changed: comparable retained text changed. Inspect the actual passages and conditions; the overall result can still contain unavailable sources.
  • comparison.status unchanged with observation new: comparable material was read again without a detected change. This does not establish an unchanged real-world outcome.
  • observation cached or fact status no_new_observation: source observations were reused. A later response time does not make them new evidence.
  • incomplete or not_comparable: coverage, identity or observation order is insufficient. A failed source is not zero, NO, unchanged, or proof that a contractual fallback occurred.
  • For supported CPI/PCE, monthly_comparison distinguishes facts_changed, provenance_changed, facts_unchanged and first_observed_release. First observed means first in these two captures, not first publication. A changed reported value alone does not prove an official revision.
  • For GDP, release_comparison keeps facts_changed in the qualifying Advance reading separate from revision_changed in later estimates. Do not replace the contract’s required vintage with the newest number.

An unchanged result can mean no new observation

For a local control on October 16, we compared the same retained October Fed 50+ bp decrease response with itself. That response was captured on October 13 at 14:11:06.417 UTC. The comparator returned status unchanged and observation cached. It made no network request: this demonstrated handling of a reused capture, not a fresh check of the Federal Reserve.

You can compare two separately retained raw evidence responses without private account access using howsure_compare_source_evidence with previous and current objects, or the CLI below. MCP accepts complete objects, not file paths. Choose the monthly flag only for the corresponding CPI/PCE readings. This comparison does not fetch sources.

# Two complete raw evidence responses already saved separately.
npx --no-install howsure evidence-compare first.json later.json \
  --include-monthly true > comparison.json

A corrected answer is not a changed source

On October 15, HowSure published a correction for one October 14 GDP reading: its warning said the adjacent bracket list was missing, although a compatible seven-contract roster was retained. The correction fixes that warning. It does not supply a qualifying GDP release value or establish settlement.

Exact saved-version reads carry a separate correctionReview. Affected means review the named error and action. Unknown means applicability could not be established; not_affected means no match to the finite registry, not that the answer is certified correct. The original answer and old exports remain unchanged. Offline copies receive no later notice until another read.

Decide what to revise, then preserve both versions

Give a returning agent read access only. An owner may share one exact Check projection with one key; other documents, notes and future revisions stay excluded. Revoking the grant or key stops later reads but cannot recall copies already obtained. Returned source passages still reveal research context, so an export is not automatically public.

Use the comparison to decide whether your reasoning needs a new version. Changed facts do not automatically determine their effect on a forecast, and HowSure does not recalculate the author’s probability. Respect Retry-After on service limits; avoid blind polling. The private workflow here is documented behavior, not a newly executed account test for this article. No new watch or paid model is required.

Sources & verification

Sources and instructions checked 16 October 2026. Live responses can change. Editorial method and limitations.

Try it with a real contract

Open your saved work