Watchtower Assistant & AI Guide
Contextual help, current public evidence and specialist research tools using your existing AI connection.
Documentation revision watchtower-ai-guide-20261019-v1. Source records, observation dates and provider availability are checked per request.Watchtower Assistant: Guide and Analyst
Open Assistant from the header or Ask AI beside a workspace. Guide explains Watchtower using the versioned handbook and methods. Analyst reads approved public datasets and source records for your question. Specialist tools include risk explanations, semantic queries, multilingual claims, country roles, report linking, contradictions, cited briefs, alert drafts, document analysis, map queries and forecast review. Choose a tool, review the context and sources, then Run AI. The signed-in account's existing OpenAI connection and Oracle request/concurrency limits apply. Static guide search and authored instructions remain available without a model connection. Requests are not automatically sent when you open a panel. Provider usage charges apply; Public Oracle prompt credits are not consumed by this assistant.
Global Risk Overview and the Systemic workbench
Risk Overview and Systemic Risk are consolidated into Global Risk Overview. Choose Risk map or All-domain matrix, select a country or region and a domain, then read Drivers & source evidence. Twenty domains are available; missing or expired evidence is unknown, not zero. Grey means insufficient eligible evidence. Composite evidence-pressure indices require at least two of five evidence families and are not calibrated crisis probabilities. The 1, 7 and 30 day controls select evidence lookback, not a future forecast horizon. Systemic outlook & track record retains Current, Outlook, Impact & dependencies, Evidence, Track record and Methods. Its original forecasts remain shared with Forecast Lab. Legacy workflows preserve the older 312-cell regional matrix and its saved watchlists, playbooks and brief archive with their original model meaning.
Ask AI to explain an existing score
Choose a country/domain and select Explain risk & change. The assistant receives the existing score components, source clocks, coverage and eligible drivers; official arithmetic remains deterministic. It can explain the current score from the supplied packet. A compatible prior snapshot is required to explain a change; a changed number alone does not identify its cause. Score deltas may reflect source revisions, changing coverage or reporting age. AI interpretation never replaces the displayed official index or creates a missing country score.
How assistant evidence and citations work
Each answer identifies its evidence snapshot and the exact source or documentation sections used. Expand a citation to inspect the supporting text, original source URL, source date/period and capture time. Source content is untrusted data, not instructions. Citation IDs and quoted spans are checked; unsupported numerical statements and a separate support-review pass can remove claims. These checks do not prove a publisher's claim is true. The assistant labels interpretations and gaps. It cannot manufacture missing observations, locations, legal effects, losses, historical revisions or outcome labels. Selected excerpts and page limits are not an exhaustive review of a dataset or document.
Semantic search and Data Q&A
Semantic search & Data Q&A interprets the question, expands search terms, proposes typed reads from the reviewed source catalogue and reranks retrieved candidates. It uses the existing exact/lexical index with model-assisted expansion and reranking; it is not a precomputed vector database. Inspect the query plan and source records. Country, period, unit and source coverage remain visible. Calculations use deterministic registered operations; generated text does not silently convert units or combine incompatible periods. Use the original Data Q&A for its deterministic source values and exports.
Multilingual claims and country roles
Use Multilingual claim extraction to distinguish source-reported occurrences, allegations, plans, denials and historical material. Use Country-role review to identify affected, actor, mentioned or unclear country roles. Every proposed classification needs supporting original text. Country roles are subject associations, not verified incident coordinates; publisher headquarters does not establish where an incident occurred. These are advisory annotations and do not change source records, incident verification, scores or official geographic assignments.
Semantic links, contradictions and revisions
Semantic report linking proposes same-document, same-claim, related-update or possibly-same-event relationships within the selected evidence. Contradictions & revisions compares cited claims, their time windows and scope. An update over time or a different denominator is not automatically a contradiction. Raw source records remain separate; candidate links do not create independent corroboration or change event counts. Retained versions can be inspected where the source supports them. If only a previous hash exists, earlier wording is unavailable and must not be invented.
Draft a cited briefing
Choose Cited briefing, describe the topic, geography and period, and select the public evidence to review. The assistant drafts developments, changes, context, gaps and next checks with citations. The result is a bounded draft, not a complete world situation report. Download the answer and evidence JSON or continue in the existing research workflow. Existing saved briefing editions are immutable and are not rewritten. Private watchlists, cases, portfolios and notes are not automatically sent to the model.
Review alert priorities and create an alert draft
Alert triage & rule drafts can suggest a review order for selected public evidence and propose a threshold rule using an available typed source field. State the desired country, measure, operator and threshold. The rule is handed to the existing Alert Center builder for review; choose Save paused or Save & enable yourself. AI does not enable rules, change thresholds, suppress alerts or send notifications automatically. Source identity, units, unknown values, initial baselines, cooldowns, budgets and existing permissions remain authoritative. Alert Center monitoring runs while a verified Watchtower browser tab is visible; it is not 24/7 server push.
Analyse attached or official-source documents
Document analysis accepts text and bounded PDF excerpts, or an approved official document selected from supported source records. The attachment is processed for this request; the extracted text selected for analysis is sent to your connected OpenAI provider only when you run AI. Text PDFs retain page references. Image-only PDFs require explicitly labelled model transcription of a bounded page excerpt; transcription can be wrong and should be checked against the original. Encrypted, malformed or over-limit files are rejected. Do not infer enactment, commencement or current legal force from passage or document publication. Original source instructions and exceptions take precedence over a summary.
Ask for a geographic selection
Map query assistant proposes supported country/region/source/date filters and can apply deterministic bounding-box or distance filtering to suitable source-provided point geometries. Review the interpreted area and the bounded dataset selection. A country anchor or inferred named place is not an incident coordinate; unsupported geometries and unresolved locations remain excluded from distance results. Open or apply a proposed native view only when you choose its action button. Private GIS layers are not read or uploaded automatically. Drawings and scenarios remain authored context, not validated impact predictions.
Review forecasts and their evidence
Forecast evidence review works with the exact registered target, model, issue time, horizon, source evidence and published performance. Financial outcomes remain deterministically resolved under their frozen definitions. Regional warning negatives require adequate sampling coverage. The assistant can summarize supporting and contrary evidence and propose an inconclusive/review disposition; it cannot adjudicate or rewrite the forecast ledger, backdate predictions, infer missing negatives, or declare live calibration from historical fit. Current-vintage backtests and genuinely issued prospective forecasts remain different.
Account context, replay and assistant memory
The assistant uses approved public workspace context and the evidence explicitly selected for a request. It does not dump the DOM, browser storage, portfolios, private case files or credentials into prompts. Conversation context is bounded and temporary; Clear removes it from this page. Sign-out, account verification changes and replay transitions cancel requests and clear account-scoped assistant state. Navigation/filter proposals are checked against the current context and allowed actions. Existing native Save actions govern persistence. Official scores, permissions, source collectors, payments and forecast outcomes are not controlled by generated text.
Private cases and map notes
War Room cases use the authenticated owner-scoped workbench and its native Save/Open/share controls. Local map notes, private browser GIS layers and scene recipes have their own storage semantics. Read the current workspace methods and storage labels before assuming that a case, layer or scene is shared or browser-local. The assistant does not automatically read private cases or layers. A user-attached document is an explicit per-request input, not an automatic import of private workspaces.