EdgeDepth Research

Use this when the user asks how common one exact condition was across eligible symbol-minute buckets in a time window. It returns true, false, eligible, and absent/ineligible counts plus prevalence. Do not use this for occurrence episodes, forward outcomes, multi-condition studies, or causal claims. Missing values are excluded rather than treated as false. This is a free deterministic computation.

Use this when the user supplied at least two exact historical moments and wants the deterministic feature intersection across them. Do not use this as similarity search or treat shared features as predictive evidence; test an agreed condition separately with base_rate or run_scan. This is a free deterministic computation.

Use this when reading your saved hypothesis or original observation. Do not use for general market scans or other accounts. Private Radar admin pilot. With id, read the account-owned saved definition, all attempted variants, counts, failures and supporting/contradictory examples. With event, read only the original Radar observation and an explicitly unrun example draft; never future outcomes. With neither, list your hypotheses. full=true restores every saved receipt. Opening never computes, retries or spends allowance. Other users cannot read these records.

Use this when the user wants the saved interpretation of a move by publication id. Returns the same explanation, supporting and challenging evidence, alternatives, next checks, exact market/window/as-of, source revisions and evidence gaps used by web pages and channel drafts. Optional edition pins the expected release; stop on 409 (changed) or 410 (corrected), never substitute newer evidence silently. This does not create an assessment, call an LLM, compute comparisons, consume study allowance or publish. Do not use this as causal proof, a forecast or a validated trading lesson. Use get_investigation_bundle for underlying recorded evidence; propose new studies separately with explicit approval.

Use this when the user wants an existing versioned investigation by its SHA256 id. Compact by default; full=true restores pinned input observations. Deterministic descriptive metrics, exact event/as-of bounds, contradictions and missing analyses come from the same API as the web. No model call, new investigation, scan, allowance debit or publication. Missing comparable populations must remain unavailable; propose an exact study through the existing approval flow before computing them. Do not use this to infer initiators, causality, spot lead/lag, hidden liquidations or forecasts.

Use this when the user wants to list citable public EdgeDepth reports or retrieve one by its 8-character canonical hash. Default is a compact 1h overview with counts and stored integrity status; full:true returns the exact definition and pinned result. Do not use this to present invalid or withdrawn reports as healthy, and do not use this for unpublished user research. This is a free read.

Use this when the host has read one or more attached chart screenshots and needs to verify their exact market and close boundaries against recorded candles and coverage for free. Read edgedepth://research/screenshots first. It separates uncertainty, missing dates, price conflicts and duplicate event views. Do not use this to guess dates, substitute venues or nearby moves, interpret hidden readings, or run a study. The host supplies structured facts; the server has no automatic attachment access.

Use this when raw-prose clients need fallback interpretation or the host cannot construct structured intent. Prefer prepare_study for capable AI hosts; this tool adds an external interpreter call. Pass the user question unchanged, without filling in unstated thresholds, dates, markets or outcomes. It returns a proposed research_query.v2 document, chip provenance, unsupported fragments, and clarification notices; it never runs the scan. Show one short plain-English proposal with condition, markets, dates, outcome/horizon and metering. Label inferred or unprovided values as assumptions needing approval. Keep JSON and diagnostics in tool details, available on request. Resolve unsupported fragments and ask only material questions. Call run_scan only after explicit human confirmation, submitting the same document. Do not use this for live quotes, trading advice, trade execution, or when the caller already supplied an exact document. This free proposal step uses the configured external language interpreter.

Use this when ground_screenshots resolved an event and the user wants its existing recorded lead-up or supported detector geometry, optionally preparing an exact editable setup. This rechecks the same coordinates and uses the web investigation owners; it never launches a scan and consumes no allowance. Historical snapshot entitlement and replay access still apply. Optional study preserves its exact market roster, UTC dates, reached versus finished, direction, magnitude and horizon. Do not use this on unresolved images, as an automatic drawing recognizer, for a forward rate, or as evidence that a winning screenshot sample validates a setup. Show a short concrete proposal and obtain explicit human confirmation before using the returned document with run_scan or outcome_first. Same-period results remain exploratory; check a separate period before validation claims.

Use this when you need the valid EdgeDepth query grammar, supported feature ids, operators, windows, limits, or machine-actionable error codes before constructing or repairing a query document. For a natural-language question, construct structured intent and use prepare_study; it already uses the registry-backed interpreter. Do not use this to answer a market question; it returns capabilities, not historical evidence. Do not invoke even this capability lookup for live prices, personalized buy/sell advice or trade execution. The whole registry is large: pass search or feature_ids to read one family, and compact to drop the per-feature prose. This is a free deterministic read.

Use this when you need to verify supported symbols, recorded coverage, data availability, or whether an instrument remains available before running or describing research. Use symbols for selected full records; use full only when canonical whole-universe bytes are required. Do not use this to support live-price, personalized buy/sell or trade-execution requests, even to check whether a market is supported. It does not return historical outcomes. This is a free deterministic read.

Use this when a prior run_scan returned an opaque page cursor and more occurrence rows are needed. Re-send the exact document and cursor; if its dataset revision changed, restart at page 1. Do not use this with a constructed or edited cursor, for a different document, or to infer rates from page rows. Continuations are free.

Use this when the user starts from an OUTCOME (a move of a stated size, in a stated direction, inside a stated horizon) and wants to know what the record was doing in the minutes before every move like it. It returns the outcome population with its denominators, a feasibility verdict, and one row per reading per lead-up offset carrying TWO counted shares: how often that reading sat outside its usual band before these moves, and how often it did so across every eligible minute in the same scope. Each row also counts the reading before the same move in the other direction (before_opposite_moves, null when that population is below the floor): a reading about as common there describes how far price moved, not which way, so never read it as direction. Every row is labelled selected on the outcome. A row is NOT a rule, a candidate, a finding or a predictor, and the row order is the gap between those two shares, which is display order and not a ranking: never present a row as something that works. The setup-first rerun each row carries lets run_scan count how often the original move followed the condition across all eligible minutes. You may describe both row shares immediately; do not claim a forward rate until that rerun returns. The same-period rerun remains exploratory, not independent validation. Keep its original target, using full_outcomes if the projection omits that rung. Do not use this to filter outcomes, to mine for a strategy, or to recommend a trade. A scope with too few realised moves is REFUSED with the counts and four honest adjustments rather than answered underpowered, and a refusal spends no allowance. Require explicit human approval of the exact study before calling this tool. Present a short plain-English confirmation with condition, markets, dates, outcome/horizon, all unprovided assumptions, and possible allowance consumption. A model-supplied flag is not human approval. After a material definition change, obtain fresh confirmation; equivalent cano…

Use this when proposing a bounded Radar hypothesis comparison. Do not use to execute or validate a strategy. Free deterministic preparation for the private Radar admin pilot. Supply the exact event and draft, plus existing id when revisiting. Returns an unrun proposal with exact four requests, revision, plan hash and allowance estimate; creates no account record. Draft fields are question, mechanism, challenge, from/through and reserveFrom/reserveThrough (UTC dates), population/a/b (three-item native clauses), decision (test/keep/reject), note. Reuse a draft from get_hypothesis; disclose every proposed assumption. One Binance asset; six existing numeric readings, A/B/A+B with common eligibility; population additionally allows native daily price/OI-ratio crossings. +5% MFE within 4h is fixed. Development <=31 days; seven full days before reserved evaluation; anchors end by September 11, 2026. Evaluation is never queried. Present exact scope, conditions, dates, outcome and allowance for human approval before run_hypothesis. A plan hash is not consent.

Use this when preparing a setup-first study. Interpret the user text/image in the host, then submit structured intent here. Deterministic validation, canonical definition, provenance and fresh allowance estimate; no LLM, scan or charge. Propose assumptions explicitly, preserve reached versus finished, and ask one material clarification when needed. For EdgeDepth screenshots with a confident symbol/time, use snapshot_at to retrieve actual readings first. Never invent timestamps or thresholds. Return one short proposal for human approval; call run_scan with its unchanged document only after approval. Existing approved exact documents may run directly. Do not use for live prices or trading advice.

Use this when a pointed move or symbol needs an explicit sector population. Do not use it to authorize computation or infer historical membership. Returns recorded sector choices, the exact eligible Binance crypto perpetual roster, coverage and membership limitations. Missing, ambiguous, thin or oversized sectors remain blocked, never broadened. Current tags are not point-in-time membership. Use the returned roster in prepare_study or investigate_move; this does not approve computation.

Use this when an exact where-only research_query.v2 document has been confirmed and the user explicitly wants the matched occurrence distribution beside every other eligible predicate-false bucket. Do not use this for sequences, as a default reference baseline, for covariate matching, significance, or causal claims. Read rates from each side's present denominator. Require explicit human approval of the exact study before calling this tool. Present a short plain-English confirmation with condition, markets, dates, outcome/horizon, all unprovided assumptions, and possible allowance consumption. A model-supplied flag is not human approval. After a material definition change, obtain fresh confirmation; equivalent canonicalization does not require another approval. Keep the exact document inspectable in tool details and provide it on request; include the reproducibility key with the result. Rates come from outcomes_summary over all occurrences; page rows are examples, never the denominator. Outcome fields cannot be filtered; expect OUTCOME_IN_PREDICATE if tried. A fresh cohort computation can consume research allowance units; cache hits, reruns, and 304 revalidations are free.

Use this when executing the next request of an explicitly approved hypothesis comparison. Do not use to search combinations, open evaluation or retry unknown work. Requires research:read and research:hypotheses plus private Radar admin access. Only after human approval of the exact prepare_hypothesis proposal, submit that unchanged proposal. This records all four planned requests before running the first. Fresh API requests can each consume allowance. Continue serially using the returned continuation instead of a proposal until it is absent (at most four native requests under the same approval). Never change the definition or infer approval from a model flag/hash. No continuation means stop. After timeout or ambiguous completion, get_hypothesis first; never retry a running/unknown step. Missing, failed and negative results remain saved. The account record and web link are shared with the web interface. Reserved evaluation is never queried; no automatic trading, monitoring or validation.

Use this when an exact research_query.v2 document has been confirmed and the user wants historical occurrences and what followed. It returns the exact definition, counts with denominators, forward outcomes over all matches, an unconditional same-scope reference baseline when available, a reproducibility key, and replay-linked representatives. Read rates from outcomes_summary and use each horizon's present count as its denominator; horizons run from 30m to 7d, and a horizon absent for an occurrence is a record-edge fact that is counted absent, never a small value. Do not use this for live quotes, personalized buy/sell advice, trade execution, prose interpretation, or outcome filtering. Require explicit human approval of the exact study before calling this tool. Present a short plain-English confirmation with condition, markets, dates, outcome/horizon, all unprovided assumptions, and possible allowance consumption. A model-supplied flag is not human approval. After a material definition change, obtain fresh confirmation; equivalent canonicalization does not require another approval. Keep the exact document inspectable in tool details and provide it on request; include the reproducibility key with the result. Rates come from outcomes_summary over all occurrences; page rows are examples, never the denominator. Outcome fields cannot be filtered; expect OUTCOME_IN_PREDICATE if tried. sequence.within accepts 15m/30m/1h/4h/12h/24h and the matching ISO aliases. sort accepts times.anchor_time, outcome.mfe_24h or outcome.mfe_7d, asc or desc: an outcome sort orders the returned page only (the biggest runs first, and the max_mfe_7d representative carries a replay handoff to the biggest one) and changes no count or rate; outcome.* in a predicate is still OUTCOME_IN_PREDICATE. identity.symbol must be an exact lowercase Binance USDT-M perpetual symbol. A fresh initial scan can consume research allowance units; cache hits, continuations, reruns, and 304 revalidations are free. Succ…

Use this when the user wants to test whether the outcome distribution of one confirmed population changes when its existing anchors are partitioned by one setup-time condition. It returns split_true, split_false, and split_absent denominators and outcomes without moving or creating anchors. Do not use this for sequences in the split, cursor populations, causal claims, or two independently selected scans. Require explicit human approval of the exact study before calling this tool. Present a short plain-English confirmation with condition, markets, dates, outcome/horizon, all unprovided assumptions, and possible allowance consumption. A model-supplied flag is not human approval. After a material definition change, obtain fresh confirmation; equivalent canonicalization does not require another approval. Keep the exact document inspectable in tool details and provide it on request; include the reproducibility key with the result. Rates come from outcomes_summary over all occurrences; page rows are examples, never the denominator. Outcome fields cannot be filtered; expect OUTCOME_IN_PREDICATE if tried. A fresh stratified computation can consume research allowance units; cache hits, reruns, and 304 revalidations are free.

Use this when the user asks whether an approved setup was profitable under explicit trade rules. This is separate from descriptive run_scan outcomes. Require explicit human approval of the exact setup, roster, dates, direction, entry timing, stop, target/trail, hold, fees, slippage and overlap before running. Proposed defaults: next-minute open entry, 1% initial stop, no fixed target, 2% trailing distance from completed closes, 240-minute hold, 6 bps fee and 10 bps slippage per side, skip same-market signals until exit. Direction must be chosen explicitly. Explain stop gaps use the worse open, stop wins same-bar ambiguity, and trailing updates start next bar. Missing historical opens mean unavailable trades. A fresh run can consume allowance. Rules or scope changes require a new proposal and approval; a model flag is not approval. Use complete-result summary for win/loss counts, average wins/losses and expectancy after fees/slippage. Funding is omitted, so never call returns fully net. These are event trades, not portfolio equity or leverage. Do not use MFE maxima as realized exits or promise profitability. Preserve original question, selected measurement and source investigation in source_measurement. Limits: explicit 1-100 markets, 31 days and 5000 signals. Alerts remain web-managed setup recurrences and never execute these rules.

Use this when the user asks to save a hypothesis or decision. Do not use to run studies, delete history or change alerts. Requires explicit user intent to save and optional research:hypotheses permission, plus Radar admin access. Save the draft or Test/Keep/Reject decision at the known revision (0 for the id returned by preparation). Existing attempts and original source are immutable. Refuses concurrent edits; read the current history before reconciling. A failed/ambiguous save must be checked with get_hypothesis, not blindly repeated. No study, alert, trade or publication. Keep means retain for more testing, not validated edge.

Use this when you need the recorded feature values, window aggregates, and fired rulebook ids at one exact symbol-time, often to turn an observed moment into candidate scan clauses. Do not use this to find similar moments, compute outcomes, or infer that a selected moment is typical. Current-minute reads are available on every key; past-minute reads require the corresponding account entitlement.