Social evidence and signal extraction

From messages to attributable features

Search input

social.search accepts a query, result bound and pagination cursor. Source-specific operators and historical coverage follow the connected source. Results retain the query scope and retrieval time.

{
  "query": "$AAPL lang:en -filter:retweets",
  "type": "Latest",
  "count": 20,
  "since": "2026-10-01T00:00:00Z",
  "until": "2026-10-02T00:00:00Z"
}

Identity, observations and pagination

A social observation retains the source item identifier, author identifier, publication time, retrieval time, language and query scope. Display names and profile images are mutable attributes, not identity keys. Edited or deleted posts retain their observed revision and availability state according to source rights.

Pagination extends the declared search rather than proving an exhaustive population sample. Search ranking, access limits and source outages affect coverage. A repost preserves its relationship to the original item so repeated distribution is distinguishable from independent corroboration.

Extracted claim

{
  "claim_id": "claim_example_1",
  "evidence_refs": ["ev_post_example:1"],
  "attribution": "reported_by_author",
  "entity_candidates": ["equity:listing:apple-common-xnas"],
  "instrument_resolution": "requires_confirmation",
  "corroboration_refs": [],
  "interpretation": "Research lead, not independently verified."
}

Signal construction

Attention features separate message volume, unique authors, source diversity and time concentration. The comparison window uses a historical baseline with compatible time-of-day and coverage conditions. Missing baseline history produces an unavailable statistic rather than an assumed zero.

Relevance and sentiment are inferred annotations attached to evidence, not verified facts or investment probabilities. The evaluation layer measures whether such features add information beyond price and news baselines. An attention spike is a candidate event whose meaning still depends on identity resolution, source quality and the instrument’s current state.

Automation boundary

Scheduled monitors persist a query, source scope, cadence and agent revision independently of an open conversation. Each run produces timestamped observations, deduplicates previously seen records and records coverage gaps. Media interpretation retains a reference to the source asset and the model revision used.

A trigger creates a research result or an action proposal according to its configured policy. Financial actions pass through the same authorization boundary as other execution requests. Source entitlements govern retrieval and redistribution; tool-returned content remains untrusted input to the agent.

Observation model and duplicate clusters

A social observation is a tuple o_i=(id_i, author_i, published_i, available_i, content_i, entity_i). Entity resolution associates a claim with an instrument or issuer and preserves ambiguity when several candidates fit. An exact post identity establishes source continuity; it does not establish that the post’s claim is true.

Duplicate clusters distinguish repeated distribution from independent observation. Exact repost links and content similarity can associate messages with one underlying claim. Feature aggregation records both raw message count and cluster count. A burst dominated by one syndicated claim therefore remains distinguishable from several independently sourced reports.

Coverage-matched attention normalization

N_t is the deduplicated cluster count in a fixed window. μ and s are the mean and sample standard deviation of comparable earlier windows in baseline bucket b(t), such as the same session hour. p_a is author a’s share of attributed observations; effective author count is a concentration diagnostic, not a count of statistically independent sources. A missing or zero-variance baseline leaves z undefined.

Worked attention example

Suppose a 30-minute window contains 30 deduplicated clusters and comparable earlier windows have mean 10 and standard deviation 5. The normalized attention value is (30−10)/5=4. If four authors contribute equally, the effective author count is 4; if one dominates, it approaches 1.

These diagnostics describe unusual activity and concentration, not a 4-sigma investment opportunity or an event probability. Search coverage can be selected and nonstationary. Evaluation compares features across matched retrieval conditions and tests their incremental association with subsequent outcomes, including no-event and no-trade cases.