Noah

Fresh standalone review_v4: privately seal your full independent current-world research, estimate/distribution, sources, contextual mechanisms, counter-case, assumptions and uncertainty before Noah's number is exposed. Noah independently researches, then internally consults your full contribution and calculates one final forecast. Take the research seriously; never submit a mandatory revised number after a comparison. Follow the exact returned continuation. Present Noah's authenticated final faithfully with substantive evidence and uncertainty; omit model debate, intermediate estimates, gaps and credit for changing Noah. One optional final concern may flag a specific missed fact, mechanism, contradiction, settlement error or unsupported confidence, with sources and reasoning but no replacement number. Submit it once through standalone_forecast_final_concern_v1; Noah owns recalculation. Checks assess retained/supplied material without fresh retrieval. Skipping the check never blocks delivery. Explicit final audits use standalone_forecast_audit_v1 with audit_requested:true only when the user requests the audit. Historical v1/v2/v3 follow their retained contract. Retrieve one durable async dataset result. The safe default is a compact digest; request result_view:'full' only for a decisive packet. If pending, respect retry_after_seconds and poll the same job without resubmitting. For a nacfo_ collaborative forecast, follow collaborative_forecast_orchestrator.next_host_action: authenticate the host opening, run the one bounded challenge and finalise from the stored challenged receipt. Never call legacy noah_reconcile_forecast or transcribe a replacement number inside that workflow. A read-only lookup; creates no Predict state.

Retrieve a compact research board for 1-12 durable jobs in ONE host tool call. It waits against one shared deadline, never resubmits work and never returns the full packets. If host_action_required is poll_now, transparently report that the same stored jobs remain in progress and repeat this tool with the same IDs after retry_after_seconds. If compose_now:true and do_not_poll_again:true, follow collaborative_forecast_orchestrator.next_host_action for any nacfo_ forecast, or compose the ordinary report. Never call legacy noah_reconcile_forecast inside a nacfo_ workflow and never invent a native or final number.

TARGET OPTION ROWS: forecast_target_contract.options contains 2-64 objects, each with exactly three required string fields: option_id, label, definition (additionalProperties:false). Use option_id, never id. option_id is 1-80 ASCII letters, digits, underscores or hyphens; label is 1-200 characters; definition is 1-800 characters. Example target row: {"option_id":"zero","label":"0","definition":"The resolved count is exactly 0."}. Preserve supplied option identity and order. Exact discrete choices remain categorical even when labels are numeric strings. These are target definitions, not probability-distribution rows; do not put probabilities or extra fields here. This row contract still applies when a connector displays options as Array<unknown>. OPTIONAL PLAN PREVIEW. Hand Noah a sanitised brief and get back the crisp sheet plus one bounded, server-authoritative research programme without starting retrieval. Fresh Deep execution does not depend on this separately permissioned tool: noah_request_dataset accepts the same lead/intents/prior fields and internally plans and starts every required component. Use this preview only when the user explicitly wants to inspect or revise the plan before execution. Host priors never count as evidence. Every trend report uses trend_report plus consolidated discovery and includes dated Noah probabilities through its existing lightweight forecasts; use automatic trend_forecasts and bounded forecast-lane recovery rather than a separate collaborative forecast. Add compare_periods only for explicit matched-temporal semantics or exact period_a/period_b windows. When a report brief asks for a forecast, the host constructs the most natural scoreable target and supplies forecast_target_contract with the report lead, marked host_constructed_target:true: use a named official series level as numeric_continuous with its unit, otherwise the obvious binary or categorical target. If no period was given, omit resolution_date and prepare the meaningful endpoint before assessment; never invent a six-month deadline. Attempt the forecast child; if it cannot run after bounded recovery, deliver the report with one method note. If no scoreable target can be constructed, return the report with its dated outlook and scenario distribution, not a reader-facing estimator error. Preserve the exact server roster and stop on structural mismatch instead of adding or merging local rows. PORTABLE BINARY RULE: for an ordinary natural two-outcome probability question, send the user's exact wording as client_brief and OMIT forecast_target_contract, resolution_question, resolution_date, yes_pays_on and no_pays_on; Noah binds binary_event automatically; neutral preparation must establish a meaningful omitted horizon before assessment. The backend owns relative scheduled-event identity/date resolution; never invent that date or replace a client-supplied date. Do not manufacture or retry alternate predicates. Supply forecast_target_contract only when the user provides special settlement semantics or requests categorical, ordinal, continuous, integer, count or comparative geometry. If a supplied special contract fails validation, preserve all of its fields and stop; never delete them and retry the same question as an ordinary literal forecast without an explicit user choice to change products. Add Other/tie only when the real resolution space needs it. There is no fixed job or lane count; the plan is never the final answer.

POST-ESTIMATE RECONCILIATION RECEIPT — LEGACY COMPATIBILITY ONLY. This preserves historical non-nacfo_ reconciliation semantics for runs explicitly begun on the old path. Do not use it for a fresh forecast or for any workflow with a nacfo_ collaboration ID. Current collaborative forecasts seal the host first, bind Noah's stored native result, run one evidence-receipted challenge and finalise through noah_follow_up_dataset actions without a fixed blend, movement cap, host override or crowd weight.

UNIFIED RESEARCH COMMISSION. IMPORTANT: a normal Deep response is an action-only durable admission receipt, not an answer for the user. When it returns host_action_required:'poll_now', execute its exact noah_get_results next_tool_call in the same response and continue those same read-only polls until the shared board is terminal; never narrate the pending receipt or promise future completion. Ask Noah's field network for a shaped evidence file. For a fresh Deep report, supply the best lead shape plus every distinct intent: this single approved call plans and dispatches every authoritative required component. A plain trend report does not imply a forecast or comparison. For any probability/distribution ask, bind forecast_target_contract: binary_event for two settlement events; categorical for exhaustive named outcomes; ordinal_categorical for ordered bands; numeric_continuous for a measured value; numeric_integer or count for whole numbers; comparative_numeric for two quantities. Other/tie is optional and belongs only where the real resolution space needs it. Untyped distribution asks stop before retrieval rather than becoming YES/NO or qualitative trajectory. If the user requests a historical evidence cutoff, bind that same instant in structured evidence_cutoff for a general report or as_of for an exact or typed forecast; a cached client may place the identical value in time_window for promotion. A market-price snapshot timestamp is not an evidence cutoff. market_edge is accepted only as an explicitly named legacy compatibility shape.

OPTIONAL DRY RUN. Use this once when coverage health, measurable silence, scope tightening or an unfamiliar specialist argument materially needs a diagnostic or the selected shape's deep manual. It is not a prerequisite for a unified Deep dataset commission and must not create a separate approval pause for familiar work. Defaults to predict/live backend diagnostics through the isolated adapter; use explicit mode=mock only to test the tool surface.

TARGET OPTION ROWS: forecast_target_contract.options contains 2-64 objects, each with exactly three required string fields: option_id, label, definition (additionalProperties:false). Use option_id, never id. option_id is 1-80 ASCII letters, digits, underscores or hyphens; label is 1-200 characters; definition is 1-800 characters. Example target row: {"option_id":"zero","label":"0","definition":"The resolved count is exactly 0."}. Preserve supplied option identity and order. Exact discrete choices remain categorical even when labels are numeric strings. These are target definitions, not probability-distribution rows; do not put probabilities or extra fields here. This row contract still applies when a connector displays options as Array<unknown>. MANDATORY FINAL MECHANICAL CHECK after a completed forecast_this board. Call this once after fixing every proposed answer and before answering the user. Legacy binary rows retain the exact YES/NO/KNIFE-EDGE check. Typed numeric rows validate all seven quantiles, their order, unit, precision, bounds and integer semantics. Typed categorical rows validate the complete ordered option vector, labels, 100% total and unique or tied leader. Unknown geometries are refused with the supported list, and a non-binary row never receives a binary-label correction. This tool does not inspect evidence, create or change a forecast, access external sources or write state.

Search Noah's Archive MCP DOCUMENTATION ONLY: the capability catalogue, shape manuals and instruction texts. It never searches the news or the web. For current news, call noah_request_dataset with shape_id:'editorial_discovery' (News Wire / News Desk) or shape_id:'news_search' (News Search) instead.