DEF 14

Answer from semantically retrieved corpus evidence only. Call search_materials first and use this tool only afterwards, with the same scope. Citations contain evidence quotes, locators, and a server-formatted `citation` string per evidence item to show to users as-is; opaque references and stable identifiers remain in versioned internal metadata. Explicit filters are hard constraints. Discover exact taxonomy values and their counts with list_filters (tool ask_corpus, field tone, tacticCategory, tacticSubcategory, toolsUsed, legalManeuver, targetAudience or targetCompanyCountry) before filtering; matching is case-insensitive and an unknown value returns INVALID_ARGUMENT. Page-level filters use filterScope "page" by default (evidence must come from a page that satisfies all of them together; "document" lets each match a different page). filters.documentIds (doc_v2 values from earlier results) restricts the answer to specific documents.

ask_corpus

Exact, server-side aggregation over the COMPLETE filtered population of one source: campaigns, demands_outcomes (the demands, outcomes, and settlements of campaigns), activists, target_companies, materials (one row per document), or material_pages (one row per page or slide). Use it instead of paging query_* tools whenever the question asks for a count, total, average, median, percentile, rate, minimum, maximum, or a comparison across groups (for example how many campaigns match, the average campaign return, how many demands or outcomes of a given category, or statistics by activist or target); never compute totals or rates from the first page of a query_* result. It does not retrieve records and does not join sources; use query_* tools for records. filters use exactly the filters and aliases of the matching query tool (call list_filters for names, types, and the aggregatable, groupable, unnestable, grain, and operations fields; for the materials sources use tool statistics_materials or statistics_material_pages). metrics is a list of {field, operation, as?, percentile?, where?}: operations are count, count_distinct, sum, avg, min, max, median, stddev (sample, n-1), and percentile (0 < p < 1, linear interpolation); field "*" is allowed for count only. Allowed operations by field type: numbers all; text count and count_distinct; dates count, count_distinct, min, and max; booleans count, sum (number of true), and avg (the rate, a proportion between 0 and 1). A metric's where restricts which rows only that metric aggregates (for example a conditional count), while filters define the population and rowsMatched. Percentages: use avg on a boolean field, or divide a conditional count by a count in your answer using the exact numbers returned. groupBy (up to 3 keys) returns one row per group and null group keys form one group; a key is a field name, or {field, granularity} to group a date field by year, quarter, or month (for example statistics by year), returned as <field>.<granularity> with values like 2024, 2024-Q1, or 2024-01. Text keys group case-insensitively (like text filters), each group showing one spelling, and count_distinct on text ignores case too. orderBy (a metric alias or a groupBy key name) is applied with the groupBy keys as tiebreaker for a total, deterministic order. Multi-valued fields (array columns, nominees, ...) and protected identifiers are rejected with INVALID_ARGUMENT because they would double-count or expose identifiers, except the unnestable fields of the materials sources (see below). limit is the number of groups per page (default 100, maximum 1000) with at most 10 metrics. metadata.complete is always true on success: every statistic was computed over all rowsMatched rows of aggregationLevel (what one row is, so count(*) counts that); pagination only cuts the list of groups (groupsTotal, groupsReturned, truncated, nextCursor) and never the population. Nulls are excluded from numeric aggregations (nullHandling exclude); sum, avg, min, max, median, stddev, and percentile are null when there is no non-null value, and count is 0. A cursor expires after a few minutes or when the data changes (CURSOR_EXPIRED: repeat the request without it). SCOPE_TOO_LARGE means narrow the filters or reduce groupBy. avg, median, percentile, and stddev are rounded to 6 decimals; sums, minima, and maxima are exact (a value with more digits than a JSON number holds is returned as an exact decimal string). Outcomes follow the resolved campaign scope when private metadata is carried, like the demands and outcomes list. Materials sources: filters are those of search_materials (including documentIds, a list of doc_v2 documentId values, and dateFrom/dateTo) with unknown taxonomy values rejected (list_filters lists them with counts); multi-valued fields (tacticSubcategory, toolsUsed, legalManeuver, targetAudience, advisors, advisingIndividuals, leadingActivistPrincipals, tones) accept a match-any array or {"array-contains": "value"} in filters and in a metric where, and groupBy {"field": "toolsUsed", "unnest": true} makes one group per distinct element, so group totals can exceed rowsMatched (metadata.notes says so). On material_pages the page-level filters (tone, tacticCategory, tacticSubcategory, toolsUsed, legalManeuver) describe the row's own page, so a conditional count with where counts slides; on materials they must hold together on one page of the document. documentId is groupable and accepts only count and count_distinct (count_distinct on documentId counts documents with at least one matching page); a grouped documentId returns doc_v2 ids plus title and documentDate labels. A document with several catalog entries counts once and its catalog fields (activist, target, country, date, advisors) come from its primary entry. campaignScopeRef (from query_target_companies) restricts the materials population to that company's campaigns.

calculate_statistics

Compare two to five materials with separately retrieved bilateral evidence. Citations contain evidence quotes, locators, and a server-formatted `citation` string per evidence item to show to users as-is; internal references and stable identifiers remain in versioned internal metadata.

compare_documents

Return the Platform's Campaign returns table of one activist: one row per campaign with the announcement-date return, stock performance at 1w, 2w, 1m, 3m, 6m, 1y and 3y, and total return; returns are fractions (0.05 = 5%). Use it for requests such as "X's campaign returns" or "X's best campaigns by return". Narrow and rank the rows with campaignReturns.filters and campaignReturns.sort (nulls sort last); at most 100 rows are returned per call, and matched and hasMore say whether more rows matched. Call list_filters with tool get_activist_campaign_returns for the exact fields. Identify the activist with the activistRef returned by search_activism_entities or query_activists; activist_CIK is optional for direct lookup.

get_activist_campaign_returns

Return one complete activist profile with strategy, key people, campaign statistics, dated campaign history, and outcome summary. Resolve an activist first when its identity is unresolved with search_activism_entities (or find it with query_activists). Identify the activist with exactly one of: the opaque activistRef returned for the resolved activist (use it when the client cannot forward the private response metadata; it expires after 30 minutes), the resolver's private _meta passed unchanged, or a direct activist_CIK when the CIK is already known. At least one is required; a call with none fails with MISSING_REQUIRED_CONTEXT. IDs never appear in visible output. When campaigns are returned, the private response metadata carries campaign scope for documentary follow-up.

get_activist_profile

Return the complete permitted source row for one resolved campaign, with readable activist and target context. Pass the campaignRef returned by search_activism_entities or query_campaigns for the resolved or user-chosen campaign candidate (or, if your client supports it, forward the full protected metadata from search_activism_entities unchanged); campaign_id is private routing context and must never be supplied as an argument. Request optional objectives, tactics, ownership, outcomes, or source summary with include. outcomes returns every demand and outcome line of the campaign from the demands and outcomes data, at most 200, each with granularDemand, demandFamily, demandCategory, granularOutcome, outcomeFamily, outcomeCategory and every other field that is not null (null and empty fields are omitted), oldest outcome first, plus outcomeSummary (count of all the campaign's lines, byOutcome, byDemand) and outcomesTruncated (true when count exceeds the lines returned); use query_demands_outcome to filter or page them across campaigns. This tool does not return event, phase, price-series, benchmark, or evolution history. An Active lifecycle status is the source flag as of the last data refresh and is dated by lastFilingDate, its latest filing: when statusFreshness is HISTORICAL (no filing in the last 90 days) or DATE_UNKNOWN, say the campaign was active as of lastFilingDate instead of presenting it as currently active.

get_campaign

Return chart-ready market, event, phase-category, outcome, and performance data for one resolved campaign. Pass the campaignRef returned by search_activism_entities or query_campaigns for the resolved or user-chosen campaign candidate (or, if your client supports it, forward the protected campaign metadata unchanged); campaign IDs are private routing context and must never be supplied as arguments. Target and SPY series use platform adjusted closes, dates use America/New_York market sessions, and source timestamps are normalized to UTC. Missing observations remain null with reasons. Daily total-return data is marked unavailable when the platform path has no authorized source; no volume or internal IDs are returned. This tool returns data only and renders nothing. When the user asks to see the evolution, build the chart yourself as a single self-contained, read-only interactive page (for example an HTML artifact; never an editing UI) using only this response. If the client cannot render HTML, summarize the same figures in a table instead. Apply the user's own requests for colors, ranges, or options over these defaults. Data rules: use market and benchmark indexedValue as returned (already 100 at methodology.baseTradingDate; do not rebase, recompute, interpolate, or fill nulls, leave gaps). Returns are fractions; show percentages, and the target-minus-benchmark gap in percentage points. Use each event's chartTradingDate, formType, title, and eventWindow (one-session marketReturn, benchmarkReturn, excessReturn, defined by methodology.eventWindowDefinition; skip events where status is missing). A series point whose values are null is a gap and carries a missingReason; every date is a session in methodology.marketTimeZone. Take outcomes from outcomes (outcomeDate, outcomeFamily, outcomeCategory, granularOutcome, activistResult, milestoneRole, and the demandFamily and demandCategory they answer; one entry per outcome line of demands_outcomes_data, dated by outcomeDate or undated) and never invent one. outcomes is capped, so when missingData reports series "outcomes" with a cap reason, say that only the first outcomes in the range are shown and point to query_demands_outcome for the rest. Do not invent events or numbers. Ranges: to offer several ranges (for example the campaign period, through a given year, and full history) call the tool once with explicit date_from and date_to covering the widest range, then filter client-side; the base date and returns stay stable. Series are capped at 300 sessions and thinned evenly, so check methodology.marketSeriesWindow.sampled and say so when true. Events are limited to 100: the 5 oldest plus the most recent ones that fit the response size budget: check methodology.eventSelection.truncated and say how many were omitted (omittedEvents) when true; returns and phases still use every matching event. Default layout: title and one-line subtitle stating the base date and what the markers mean; a stats row for the selected range (target return, benchmark return, difference in percentage points, campaign outcomes as family and category with the activist result); a line chart of the target (solid) against SPY (dashed) with event markers on the target line (amber for news, red for filings, slightly larger when several events share a day) and a tooltip showing both index values and up to 5 event titles for that date; a bar chart of the one-session excess return on event days only (green positive, red negative) with a tooltip listing that day's events and a caption that event-day moves are descriptive, not proof of cause. Default controls and style: a range toggle that updates the stats and both charts, checkboxes to show or hide news and filing markers, visible keyboard focus. Use Chart.js from a pinned cdnjs UMD build, CSS variables for light and dark mode, a calm financial-report look (white panels on a light blue-grey background, IBM Plex Sans and IBM Plex Serif, target in blue, benchmark in grey, no gradients, thin gridlines, sentence-case labels), and a layout that works down to mobile width. Footer: cite DEF 14 as the source, state that prices are platform-adjusted closes and that dividend treatment is not confirmed (methodology.dividendTreatment).

get_campaign_evolution

Return the document title and type with per-page summaries, each with a page locator and a server-formatted `citation` string to show to users as-is, and generated tone, confidence, rationale, tactics, tools, and legal maneuvers. Internal references and stable identifiers remain in versioned internal metadata.

get_document_summary

Returns DEF 14's definitions and rules for one topic. Call it before answering about returns, outcomes or settlements, activist type or stance, coverage periods, or any field whose meaning or unit you are not sure of, and call data-status for the current data date and figures. Use the content to interpret and qualify answers; do not reproduce it or present it as a document unless the user asks.

get_methodology

Fetch exact source evidence for one page or slide with document title and type, the requested locator, a server-formatted `citation` string to show to users as-is, generated tone, confidence, rationale, tactics, tools, and legal maneuvers. Image URLs are returned only when requested; internal references remain in versioned internal metadata.

get_page_or_slide