Introducing a new way to review your PRs.

Read the announcement

koala73 / worldmonitor

Real-time global intelligence dashboard. AI-powered news aggregation, geopolitical monitoring, and infrastructure tracking in a unified situational awareness interface

View on GitHub

29 documents · Updated Nov 20, 2026

docs / forecasting-scenarios-and-resilience

Forecasting, scenarios, and resilience

This decision-support area turns current signals into probabilistic forecasts, tests those forecasts against later outcomes, explores defined disruption scenarios, and measures national capacity to absorb and recover from shocks. Forecast calibration and family-based evaluation help keep published skill claims measurable. It connects live intelligence to products that show both possible futures and the limits of those futures. For the wider product model, see Intelligence products; for brief generation, see Intelligence briefs and analysis; for company-level context, see Company monitoring intelligence; for source collection, see Data ingestion.

Decision-support model

The system uses three related views of uncertainty. Forecasts describe what may happen and how likely it is. Scenarios show the relative impact of a defined shock. Resilience scores describe the capacity that can shape the result.

These views answer different questions and should not be treated as one score:

  • Forecasts: What event or change is likely within a stated horizon?
  • Scenarios: If a defined disruption occurs, which countries or sectors are most exposed?
  • Resilience: How well is a country positioned to absorb and recover from shocks?

Forecast data reaches the dashboard through a generated service client. The application service uses a shared public path for the common unfiltered feed and a credentialed path for filtered feeds and scorecards. This separates cacheable shared data from caller-specific or session-gated data.

Forecast lifecycle

The forecast lifecycle turns monitored inputs into published claims, adds context for analysis, and keeps a record of whether past claims proved correct. The forecast service exposes the current feed, the evaluation scorecard, and the simulation package and outcome paths.

Forecast publication

The forecast producer reads current signal feeds and builds a bounded set of claims by domain, region, probability, and time horizon. Each published claim can carry supporting signals, cascade effects, market context, alternative perspectives, and a case file that explains the base, escalatory, and contrarian paths. The dashboard first loads the forecast list and fetches the larger case file when an engineer or analyst opens a forecast.

The public feed distinguishes a healthy empty result from a backend failure. Domain and region filters narrow the returned claims without changing the canonical forecast set.

References: scripts/seed-forecasts.mjs• L1112-1140 scripts/seed-forecasts.mjs• L16427-16446 proto/…/forecast.proto• L117-170 server/…/get-forecasts.ts• L12-43 src/…/forecast.ts• L30-59

Forecast evidence archive

The evidence archive preserves the news records needed to judge forecasts after the shorter-lived digest data expires. It stores self-contained records for the eligible evidence stream and publishes a coverage marker only when the requested lookback is proven. This lets the resolver distinguish a quiet news period from a missing or incomplete evidence window.

The backfill tool is a bounded repair path for the archive cutover. It reads the still-available source window, reuses existing archive records when possible, falls back to readable story records, and runs in dry-run mode unless explicitly applied. It writes records in batches, reads each record and index score back, and writes the coverage marker last. A partial, malformed, truncated, or unverified copy therefore cannot claim complete coverage.

References: scripts/backfill-forecast-evidence-archive.mjs• L63-75 scripts/backfill-forecast-evidence-archive.mjs• L143-223 scripts/backfill-forecast-evidence-archive.mjs• L226-293 scripts/seed-forecast-resolutions.mjs• L2802-2837 scripts/seed-forecast-resolutions.mjs• L2839-3029 scripts/_forecast-evidence-archive.mjs• L10-33 scripts/_forecast-evidence-archive.mjs• L133-225

Forecast evaluation

Every enriched forecast can carry a resolution spec, which is the rule that defines how the claim will be judged. A hard spec uses a permitted feed and a measurable condition. A judged spec uses a question for later judges when no stable numeric measure matches the claim. The forecast horizon supplies the resolution deadline, so the claim and its evaluation window remain linked.

The daily resolution process ingests forecast history into a working ledger, samples live feeds, resolves due hard claims, and sends eligible judged claims to an evidence-backed judging lane. The judged lane requires a complete evidence window before it asks independent judges for a result. It records attempts and bounded retry state, keeps transient failures pending, and seals terminal outcomes with a reason when the evidence cannot be recovered or the judges cannot agree. Terminal receipts are retained as the audit trail.

The scorecard summarizes the rolling record with resolved, pending, judged, scored, and void entries. It reports Brier and log scores, calibration buckets, domain and generation-origin groups, publication coverage, uncertainty intervals, and comparison with market skill. It treats a forecast family as one question with related windows, so confidence intervals and skill minimums do not count repeated windows as independent evidence. Synthetic, shadow, and unattributed origins remain separate from headline skill so backfill or trial data cannot look like proven forecast performance.

The scorecard also makes the evaluation process visible. It reports whether judged forecasts meet their service level, why forecasts became void, and whether the forward void share has enough resolved observations for comparison with its target. Projection horizons are graded separately from the main forecast score because their fixed curves and point-in-time meaning are different.

The forecast record shown in the product has explicit states for loading, unavailable, insufficient evidence, and ready. A ready record shows the Brier score and sample size. It can also show the cohort base-rate comparison and void rate. Domain reliability badges use only published-origin rows and remain unmeasured below the minimum family sample, so small samples do not look like proven skill. A stale marker stays visible when the scorecard is old. During an accuracy audit, the product shows the audit state instead of presenting the record as verified.

The scorecard endpoint reads the published record through a declared public shape. A missing record is not treated as a healthy zero score. A missing seed clock or an old seed makes the result stale. A failed backend remains degraded, and an unavailable or degraded response remains unusable to the product rather than being rendered as a valid track record. The generated client and server contracts carry these states with the scorecard, family outcomes, audit marker, and related score fields.

References: scripts/seed-forecast-resolutions.mjs• L277-299 scripts/seed-forecast-resolutions.mjs• L3200-3262 scripts/seed-forecast-resolutions.mjs• L3289-3328 scripts/seed-forecast-resolutions.mjs• L3376-3422 scripts/_forecast-scorecard.mjs• L124-199 scripts/_forecast-scorecard.mjs• L1082-1108 scripts/_forecast-scorecard.mjs• L1212-1262 proto/…/forecast.proto• L32-60 proto/…/get_forecast_scorecard.proto• L9-18 src/…/service_client.ts• L146-170 src/…/service_server.ts• L149-171 server/…/get-forecast-scorecard.ts• L22-89 server/…/scorecard-fields.ts• L9-32 src/…/forecast.ts• L61-65 src/…/forecast-record.ts• L14-35 src/…/forecast-record.ts• L42-117 src/…/forecast-record.ts• L294-333

Forecast calibration

Forecast calibration adjusts published-origin probabilities only when recent resolved evidence shows that the adjustment is safe to use. The system fits a separate monotone curve for each domain from a rolling cohort of resolved YES and NO outcomes. It weights each forecast family as one unit, which prevents repeated windows for the same question from overstating the evidence.

Calibration remains a shadow result until a forward cohort validates it outside the data used to fit the curve. The activation gate requires enough total and per-domain families, then checks that the calibrated Brier score is not worse than the raw score beyond a fixed margin. If the map, gate, or supporting read is missing or stale, publication falls back to raw probabilities. The decision and map are archived with the forecast run so later review can reproduce the published scale.

The following diagram shows how resolved forecast evidence supports a guarded calibration decision.

flowchart LR
    A[Resolved published forecasts] --> B[Rolling family cohort]
    B --> C[Domain calibration map]
    C --> D[Forward shadow evaluation]
    D --> E{Activation gate}
    E -->|Eligible| F[Publish calibrated probabilities]
    E -->|Not eligible| G[Publish raw probabilities]

References: scripts/_forecast-calibration.mjs• L172-196 scripts/_forecast-calibration.mjs• L287-343 scripts/_forecast-calibration.mjs• L515-529 scripts/_forecast-calibration.mjs• L601-671 scripts/_forecast-scorecard.mjs• L1318-1356 scripts/_forecast-scorecard.mjs• L1436-1461 scripts/_forecast-scorecard.mjs• L1541-1628

Forecast simulation

Deep forecast simulation tests selected high-value state units through theater-based paths. The producer creates a simulation package from the current deep forecast snapshot and publishes a pointer to it. A simulation request uses that current package and receives a server-created run identifier. Triggered tasks can also carry an opaque package fingerprint so the worker can detect package rotation without exposing storage paths.

The worker processes eligible theaters through two rounds. It returns path summaries, actor roles, reactions, stabilizers, invalidators, timing markers, and confidence. The outcome also records how many theaters were eligible, completed, or failed. Completion is reported as completed, partial, no eligible theaters, or all theaters failed. After the outcome is written, the forecast pipeline applies simulation decorations and awaits a rescore of affected forecasts.

This flow is asynchronous. A caller can read the package, trigger a run, and poll the outcome. A run-specific lookup returns the stored result when available. If no run-specific result is available, the service can return the latest available outcome with a fallback note. The dashboard service also exposes the latest theater summaries as a compact string for forecast presentation.

The following diagram shows how a forecast run becomes a simulation result and then feeds back into forecast presentation.

flowchart LR
    A[Current signals] --> B[Forecast run]
    B --> C[Deep state units]
    C --> D[Simulation package]
    D --> E[Simulation task queue]
    E --> F[Theater rounds]
    F --> G[Stored outcome]
    G --> H[Forecast decorations and rescore]
    H --> I[Published forecast feed]

References: scripts/seed-forecasts.mjs• L13445-13475 server/…/trigger-simulation.ts• L24-46 scripts/process-simulation-tasks.mjs• L1-45 scripts/_simulation-queue-constants.mjs• L27-70 scripts/seed-forecasts.mjs• L19022-19129 scripts/seed-forecasts.mjs• L19248-19419 server/…/get-simulation-outcome.ts• L90-173 src/…/service_client.ts• L321-364 src/…/service_server.ts• L410-416 src/…/forecast.ts• L67-70

Scenario processing

The scenario service answers a narrower question than deep forecast simulation: what is the relative exposure to a defined physical or policy shock? It uses a public catalog of scenario templates and a protected asynchronous run path so the same assumptions can be applied across countries and sectors.

Scenario catalog

Each template defines the shock boundary. Physical templates identify affected chokepoints, disruption size, duration, affected HS2 sectors, and a freight-cost multiplier. Tariff-shock templates have no physical chokepoint and instead apply to trade in selected sectors. An empty sector list means that the scenario applies to all sectors.

The catalog is the contract shared by the public template listing, the request validator, and the worker. A worker must use the same template meaning as the API that accepted the request.

References: server/…/scenario-templates.ts• L1-120 server/…/list-scenario-templates.ts• L7-25 proto/…/list_scenario_templates.proto• L5-28 src/…/scenario-templates.ts• L1-14

Scenario job lifecycle

The scenario flow accepts a template identifier and an optional country scope. The service validates both values, places a job on the scenario queue, and returns a job identifier for polling. An always-on worker atomically moves each job from pending to processing, validates its identity, combines the template with current chokepoint and exposure data, and stores the processing, completed, or failed state for 24 hours.

For a physical shock, the result ranks countries by exposed trade multiplied by the assumed disruption and cost shock. For a tariff shock, it uses sector vulnerability instead of route exposure. The output is a relative ranking of the top affected countries. Each ranked country also reports how much requested evidence was evaluated, so a partial subtotal is not mistaken for low exposure. The result is not a currency forecast and does not estimate total economic loss.

The worker reads exposure data in bounded batches within a fixed compute budget. It preserves missing, malformed, incomplete-route, and not-seeded records in the coverage report. This makes an incomplete run visible to the dashboard instead of presenting a smaller result as a complete view.

The worker protects queue progress across restarts. It requeues jobs left in processing after a crash, skips a job whose owner-scoped result already exists, and removes each claimed job from processing after completion or failure. The status contract makes work state visible as pending, processing, done, or failed.

References: proto/…/service.proto• L10-35 server/…/run-scenario.ts• L28-89 scripts/scenario-worker.mjs• L228-359 scripts/scenario-worker.mjs• L373-397 scripts/scenario-worker.mjs• L403-487 server/…/get-scenario-status.ts• L62-98 proto/…/get_scenario_status.proto• L39-73

Resilience scoring

The Country Resilience Index is a 0–100 view of national shock absorption and recovery capacity. It combines long-run capacity with current pressure and recovery resources. It is a complement to short-horizon instability signals rather than a replacement for them; the methodology describes the full index contract in Country Resilience Index.

Score model

The score uses six domains: economic, infrastructure, energy, social-governance, health-food, and recovery. These domains are grouped into three pillars: structural readiness, live shock exposure, and recovery capacity. Each dimension carries its score, data coverage, observed and imputed weight, freshness, and imputation class so a user can see how much confidence to place in the result.

The active overall score uses the pillar-combined form. This form reduces the ability of a strong pillar to hide a weak one. The response also keeps baseline score, stress score, stress factor, trend, 30-day change, score interval, and headline eligibility. These fields support both a country view and a ranked view.

The shared resilience scoring path treats score generation as one coherent unit. It creates the public score and its indicator trace together, gives both a generation identity and construct identity, and stores the trace beside the score. Cache readers reject entries from a different formula, schema, or active construct state. This prevents a score from being paired with an explanation or ranking built from a different scoring generation.

Confidence and headline status use data quality as well as the numeric score. Freshness can reduce effective coverage for confidence reporting. A country is headline-eligible only when coverage is sufficient, the population or coverage rule is met, and the score is not low-confidence. The ranking path keeps countries that fail this gate visible as lower-confidence or greyed-out results instead of treating them as equal-quality headline entries.

References: docs/…/country-resilience-index.mdx• L50-77 server/…/_shared.ts• L82-145 server/…/_shared.ts• L409-485 server/…/_shared.ts• L882-1015 server/…/_shared.ts• L1230-1264 server/…/_dimension-scorers.ts• L560-705 server/…/_dimension-scorers.ts• L982-1094 server/…/_dimension-scorers.ts• L2030-2065 server/…/_dimension-scorers.ts• L3043-3148 server/…/_dimension-scorers.ts• L3613-3693 server/…/_dimension-scorers.ts• L4089-4116 proto/…/resilience.proto• L5-87 proto/…/get_resilience_score.proto• L18-53

Indicator trace

The resilience indicator trace explains how a country score is built at the indicator level. It shows whether each input was observed, imputed, missing, replaced by a fallback, not applicable, inactive, retired, or affected by a source failure. It also connects normalized values and score contributions to weights, observation age, retrieval time, and source details.

The indicator response preserves the formula, data version, schema version, and construct versions beside the trace. This lets a reader relate a displayed contribution to the score generation context instead of treating raw values or weights as independent facts. Raw values are disclosed only when their source policy and retrieval evidence allow that disclosure.

Source policy is separate from a broad license label. For an observed value, the reporting path checks the exact providers and source paths that contributed to that value. It can expose raw data only for an audited source. Missing provider evidence, restricted terms, incomplete review, or a non-observed state keeps the raw value hidden while still allowing the normalized score and its attribution to be shown.

References: server/…/_indicator-trace.ts• L11-19 server/…/_indicator-trace.ts• L46-86 server/…/_indicator-trace.ts• L305-391 server/…/_indicator-source-policy.ts• L1-24 server/…/_indicator-source-policy.ts• L322-356 server/…/_indicator-source-policy.ts• L451-603 server/…/get-resilience-indicators.ts• L164-221 server/…/get-resilience-indicators.ts• L224-261 server/…/get-resilience-indicators.ts• L279-337

Ranking and reporting

The per-country score endpoint returns the full score when the country code is valid. The ranking endpoint applies a separate headline gate based on coverage, population or high data quality, and confidence. Countries that do not pass remain in greyedOut rather than appearing as equal-quality headline results.

Reporting surfaces include the resilience widget, country detail view, map layer, and service responses. The runtime manifest exposes the active formula, data version, cache state, construct versions, and interval availability. This gives operators and analysts a way to tell whether a reported score is current and whether its uncertainty data matches the active scoring method.

The score assembly treats deliberate construct states differently from data loss. Retired dimensions do not lower confidence. A feature-flagged dimension also stays out of the confidence mean while it is dark. A real source failure remains visible, and a fully imputed dimension can be reclassified as source-failure when seed health shows that its upstream data failed. Bulk ranking warm-up computes scores in memory, persists trace sidecars before score references, writes in bounded batches, and reports only scores whose writes succeed.

References: server/…/get-resilience-ranking.ts• L139-187 server/…/get-resilience-ranking.ts• L369-404 server/…/_shared.ts• L1090-1113 server/…/_shared.ts• L1141-1195 server/…/_shared.ts• L1401-1477 server/…/_shared.ts• L1569-1731 server/…/_dimension-scorers.ts• L3981-4060 server/…/_dimension-scorers.ts• L4118-4227 proto/…/get_resilience_ranking.proto• L9-17 proto/…/get_resilience_runtime_manifest.proto• L27-60 src/…/ResilienceWidget.ts• L47-128

Validation workflows

Validation checks whether these outputs are useful and honest under known data and formula limits. Forecast validation checks the integrity of a run and the quality of its resolution record. Resilience validation checks agreement with outside measures, behavior after known shocks, and sensitivity to changes in weights and inputs.

Forecast run validation

Forecast run evaluation reads the trace artifacts for one run and checks that the run state and deep snapshot exist, selected states resolve, canonical labels remain unique, reportable interactions stay within the simulation ledger, high-value candidates produce mapped signals, and convergence meets its quality target. The evaluation is stored with the run so a later review can connect a published result to the checks that supported it.

These checks protect the forecast funnel from silent collapse. A run can still produce forecasts while losing the state or interaction data needed to explain and evaluate them, so the evaluation treats those contract breaks as distinct from a simple empty feed.

References: scripts/evaluate-forecast-run.mjs• L42-113 scripts/evaluate-forecast-run.mjs• L116-133 scripts/seed-forecasts.mjs• L11411-11464

Resilience validation

The resilience validation bundle runs three complementary checks on a weekly schedule. External benchmarking compares the index with INFORM, HDI, and the WorldRiskIndex vulnerability component. Outcome backtesting tests whether higher baseline capacity aligns with faster recovery after known shocks. Sensitivity analysis perturbs domain weights, pillar weights, goalposts, and the pillar penalty to detect unstable rankings.

The benchmark is directional rather than a claim of identical methodology. The historical backtest has a known limit: it uses current live scores against historical outcomes, not scores captured before each shock. Sensitivity results provide a release gate so a change that moves too many dimension ranks cannot pass as a routine update.

References: scripts/resilience-validation-bundle.mjs• L1-29 scripts/benchmark-resilience-external.mjs• L433-530 scripts/validate-resilience-backtest.mjs• L1-22 scripts/validate-resilience-sensitivity.mjs• L150-229

Change checklist

When these files or sections are changed, remember to consider these:

  • Keep the forecast producer, resolution-spec builder, resolver inputs, and scorecard fields aligned. A new forecast horizon or source must remain scoreable through its deadline.
  • Keep the forecast evidence writer, archive reader, coverage marker, and backfill verification rules aligned. A marker must never claim a window that the archive did not write and read back successfully.
  • Keep archive retention, hash limits, truncation behavior, and judged-window coverage aligned. A bounded read must narrow or reject the usable window rather than turn missing evidence into a negative judgment.
  • Keep the judged attempt states, retry policy, horizon alerts, terminal receipts, and lane health fields aligned. A transient archive or judge failure must remain pending and observable.
  • Keep the forecast scorecard schema, public field allowlist, family-based sample minimums, published-origin rows, stale state, audit marker, and product record states aligned. A new scorecard field must not make an unavailable, audited, or small record look measured.
  • Keep the calibration fit cohort, family definition, map version, forward shadow cohort, activation gate, raw fallback, and archived publication decision aligned. A calibration change must not evaluate on fit data or silently compare different probability scales.
  • Keep the public unfiltered feed, filtered feed, scorecard access path, and generated client/server service contract aligned. A route or access change can send shared data through the wrong cache or hide a session-gated result.
  • Keep the simulation package pointer, trigger response, worker reader, and outcome lookup on the same run identity and package fingerprint contract.
  • Keep the scenario template registry, public template response, worker copy, and result formula synchronized. A template change can silently alter both validation and ranked impact.
  • Keep the scenario job identity, owner binding, result key, status lookup, and worker validation aligned. A job without the expected owner-scoped identity is discarded rather than processed.
  • Keep exposure batch limits, the compute budget, per-record coverage states, and partial-evidence fields aligned. A bounded run must remain visibly incomplete rather than looking like a complete ranking.
  • When a resilience formula, schema, pillar shape, construct, or eligibility rule changes, rotate the relevant score, ranking, interval, and history caches together. Also keep the in-payload generation and construct tags aligned so a flag change cannot serve mixed score generations.
  • Preserve coverage, freshness, imputation, and interval metadata when adding or retiring a resilience input. These fields control confidence and headline eligibility.
  • Keep indicator trace states, contribution values, source metadata, and formula, schema, or construct version tags aligned with the score response. A trace that describes a different score generation can mislead review without changing the headline number.
  • Keep raw redistribution policy and exact observed-source hints aligned. A source change can leave scores correct while exposing a value that lacks the required legal or provenance record.
  • Keep required seed health checks and feature-flag behavior aligned for energy and financial exposure. Turning on a construct before its seeds are healthy must produce a visible source failure rather than a midpoint score.
  • Keep retired and dark dimension rules separate from genuine zero-coverage or source-failure rules. The confidence mean and headline gate depend on that distinction.
  • Preserve the ordering of trace-sidecar writes before score writes and keep warm-up batch limits within the Redis transport budget. Otherwise rankings can point to missing explanations or silently omit freshly computed countries.

Important failure modes

TriggerConsequenceGuard
Forecast cache read failsThe feed is empty and marked degraded instead of looking like a valid empty forecast set.The forecast response carries explicit degraded and error state.
Forecast scorecard data is missing or its backend failsThe product can show no usable track record instead of a false zero score.The scorecard response and record projection preserve unavailable and stale states.
A scorecard has too few forecast families or too few YES/NO familiesSkill and domain badges can imply performance that the data cannot support.The scorecard reports family counts and marks results below the family and outcome minimums as insufficient.
A scorecard seed is under auditThe product can present old outcomes as verified skill.The service carries the audit marker and the record suppresses verified score display.
A forecast has no usable resolution feed or contractThe claim cannot produce a trustworthy automatic result and may remain judged, pending, or void.Resolution specs use an allowlisted source feed and an explicit deadline.
Forecast evidence coverage is missing, stale, truncated, or not proven after backfillJudging cannot establish that the full lookback window is present and can mistake a storage gap for a quiet evidence period.The archive uses a verified coverage marker. The backfill writes it only after bounded writes and read-back verification.
A judged archive read fails or covers only part of an entry's windowThe entry can be judged on incomplete evidence or can stall without a clear cause.The resolver fails closed, records archive coverage and attempt state, and keeps the entry pending until its recovery horizon is crossed.
Judged retries exhaust their age or attempt limitsA claim cannot be resolved safely and becomes a visible void instead of consuming the run forever.Bounded backoff, terminal reasons, attempt logs, and archive-horizon alerts make the transition explicit.
A calibration map or activation verdict is stale, mismatched, invalid, or ineligibleThe producer may apply an adjustment that has not been validated for the current map and evidence.The publication decision checks map version, gate age, gate eligibility, and a raw-probability fallback.
Calibration evaluation includes entries from the fit windowReported improvement can be in-sample and look stronger than future performance.The evaluation rejects overlapping emissions and uses a forward cohort.
Simulation package rotates after a task is queuedThe worker may run a package different from the one that triggered the task.The task carries an opaque package fingerprint and the outcome records rotation for diagnostics.
Scenario queue or worker is unavailableThe caller receives an enqueue or failed status instead of a completed impact ranking.Explicit job states, orphan recovery, and result retention expose the failure.
Exposure records are missing, malformed, incomplete, or exceed the compute budgetThe ranking contains only part of the requested evidence and can be misread as a complete exposure view.Per-record states, requested and evaluated counts, partialEvidence, and coverage.status expose the limit.
Resilience inputs are sparse or staleA score can have low confidence or fail headline ranking even when a numeric score exists.Coverage, freshness, imputation class, score intervals, and the greyedOut ranking path remain visible.
A resilience score and trace come from different generationsA user can see an explanation that does not support the displayed score.Immutable generation IDs and cache identities must match before the pair is returned.
An indicator source fails or raw-value provenance is incompleteThe score may retain a fallback or imputed result while the trace cannot present it as a normal observed input.Indicator states, source-failure reasons, retrieval timestamps, and conditional raw-value status remain explicit.
An observed source is not audited for the exact valueThe normalized score can remain available but the raw observation must not be redistributed.Exact provider and URL matching is required before raw exposure.
A trace sidecar or score write fails during ranking warm-upThe ranking can publish incomplete data or a score without its explanation.Trace writes happen first. Bounded pipeline writes retry and only successful score writes enter the warmed map.

Key files

FileWhy to read it
src/…/forecast.tsThe product-facing forecast access path and the split between shared feed, filtered feed, and scorecard requests.
src/…/forecast-record.tsThe user-facing states for forecast track record, audit status, stale data, sample limits, and domain reliability.
src/…/service_client.tsThe forecast service boundary for feeds, scorecards, simulation packages, outcomes, and triggers.
src/…/service_server.tsThe server route boundary and request/response contract for the forecast service.
server/…/get-forecast-scorecard.tsThe public scorecard read path, seed freshness rules, audit marker, and degraded response states.
server/…/scorecard-fields.tsThe public scorecard field boundary and the separation between contract fields and producer-only data.
proto/…/forecast.protoThe high-level shape of a forecast and its explanation and resolution data.
scripts/seed-forecasts.mjsThe forecast lifecycle from signal inputs through publication, deep state, simulation, and rescore.
scripts/seed-forecast-resolutions.mjsThe daily resolution ledger, dedicated evidence read, judged lane, scorecard publication, health state, and durable receipts.
scripts/backfill-forecast-evidence-archive.mjsThe bounded archive repair and cutover path that proves evidence coverage before it publishes the marker.
scripts/_forecast-resolution.mjsThe contract that makes published forecasts evaluable.
scripts/_forecast-scorecard.mjsThe rolling evaluation record, family-based skill measures, uncertainty, judged-lane health, and calibration shadow gate.
scripts/_forecast-calibration.mjsThe domain calibration map, forward validation, publication decision, and raw fallback.
scripts/_forecast-evidence-archive.mjsThe retention, record shape, coverage, and continuity contract for evidence used to judge forecasts.
scripts/evaluate-forecast-run.mjsThe checks that protect the quality of a forecast run and its trace.
proto/…/service.protoThe asynchronous scenario service contract.
server/…/run-scenario.tsThe request validation and scenario job acceptance boundary.
scripts/scenario-worker.mjsThe scenario job claim, exposure computation, coverage reporting, result lifecycle, and orphan recovery.
server/…/get-simulation-outcome.tsThe run-specific polling behavior and latest-result fallback.
scripts/_simulation-queue-constants.mjsThe shared simulation queue, outcome retention, run-ID, and package fingerprint contract.
server/…/_shared.tsThe shared score generation, cache identity, trace pairing, confidence, trend, ranking warm-up, and eligibility path.
server/…/_dimension-scorers.tsThe domain and dimension model. It also shows how observed, imputed, regional, retired, and failed inputs affect the score.
server/…/_indicator-source-policy.tsThe raw-value disclosure rules and exact source audit contract for resilience indicators.
server/…/get-resilience-indicators.tsThe country-level reporting boundary for score generation context and indicator evidence.
server/…/_indicator-trace.tsThe shared model that records indicator states, contributions, and dimension trace data.
server/…/get-resilience-ranking.tsThe ranking, coverage, and headline eligibility behavior.
docs/…/country-resilience-index.mdxThe user-facing methodology and rationale for the resilience index.
scripts/resilience-validation-bundle.mjsThe scheduled set of external, historical, and sensitivity validation workflows.