Skip to content

BACKLOG.md

Latest commit

 

History

History
62 lines (48 loc) · 5.72 KB

BACKLOG.md

File metadata and controls

62 lines (48 loc) · 5.72 KB
Raw

ERCF — Product Backlog

Priority 1 — High Impact, Feasible

  • Data freshness indicators — show in UI when ACAPS/INFORM and UCDP were last updated; flag stale data (>30 days for ACAPS, >1 year for UCDP).
  • Mass atrocity / genocide indicator — structured UI flag for OOS cases that are genocide or deliberate massacre (currently only in out_of_scope_reason text). Teresa Cantero suggestion.
  • Recalibration with new cases — pipeline to add new cases to the corpus and re-run full_calibration.py when the historical dataset expands.

Priority 2 — Medium Impact

  • Dynamic D-scores — model conflict escalation over time (D1/D6 drift); currently static snapshot.
  • Multi-corridor scenario — compare two evacuation routes side-by-side (different distance, terrain, D4).
  • Export to PDF/Word — generate funding appeal template with scenario summary, cost breakdown, and historical comparisons.
  • Mobile responsive improvements — current layout optimised for desktop; radar and charts need responsive sizing for tablets.
  • Conflict duration distribution chart — histogram showing duration distribution across the 31 historical cases corpus. Would give users context on typical conflict timelines to inform the planning window selection. Distinct from the mortality calibration model — purely a contextual visualization. Data available: all 31 cases have duration_days in historical_data.py.

Priority 3 — Research / Academic

  • Expert panel validation of D1–D7 weights — current weights are calibration-derived; no peer review yet.
  • L4 structural limitation resolution — L4 cannot simultaneously capture high-mortality urban sieges (Aleppo: 31,000 deaths) and large-enclave precision operations (Gaza Cast Lead: 965 deaths). Requires either sub-levels or a separate model branch.
  • Confidence intervals on cost estimates — current outputs are point estimates; uncertainty bands would improve funding appeal usefulness.
  • Integration with OCHA FTS — link evacuation cost estimates to the OCHA Financial Tracking Service for real-operation comparisons.
  • LOOCV within-2× computation — not recomputed for v7 (only R² LOOCV available); requires implementing leave-one-out within-2× rate.

✅ Recently Completed

  • Supply price volatility (was Priority 1) — implemented as the Market Price Adjustment feature: live EIA (Brent crude) + FRED (food basket) commodity prices, /api/commodity-prices/{iso3}, applied as a fuel/food cost-adjustment factor.
  • Walking evacuation integration (was Priority 2) — walking mode's cost now feeds directly into the headline cost and updateDecisionAnalysis() break-even (static/app.js), not just its own separate panel.
  • Transport mode integration — Ground/Air (fixed-wing)/Air (helicopter)/Walking all now drive the headline cost and Decision Analysis consistently (static/app.js _isAltMode branching), with a token-guarded fetch to avoid stale-response races between modes.
  • Vulnerable population retention model — vulnerable individuals modeled as ~2× less likely to evacuate (ref: AARP/FEMA Post-Katrina 2006; WHO Disability & Disasters 2005), adding a "vulnerable population support" ($2.50/person/day) cost line to the Assistance Cost breakdown.
  • Emergency extraction rate recalibration — reduced from a 10–95% range to 1–8% of the remaining population (L1–L4), aligned with documented UNHAS medevac/security-evac share (~0.25% of passengers) instead of conflating "mass extraction operation probability" with "individual medevac need".
  • Cumulative chart fix — the "Cumulative Assistance Cost" chart (formerly "Cumulative Financial Cost") now recomputes calcRemaining() day-by-day instead of showing only the flat $3.50/day survival baseline, so its final point always matches the Total Assistance Cost figure.
  • Decision Analysis methodology fix — "Evacuate now" uses the full mobilization cost (resources.totalCost) instead of a marginal per-evacuatedPop recompute, consistent with the headline card; added a cost-of-delay note under the break-even chart.

Known Technical Debt

Item Status Impact
LOOCV within-2× not recomputed for v7 Open Low — R² LOOCV is the primary metric
GeoNames population data is pre-conflict By design Medium — users should adjust for displacement manually
Access multiplier L4 (×4.0) unvalidated Open Low — conservative estimate, no published source >4×
ercf_auditoria.zip (75 MB) in git history Open Low — GitHub warns, doesn't block; .gitignore fix pending
context_ai.py Claude API — prompt not cached Open Low — single-call endpoint, not high-frequency
weather_data.py seasonal factors not validated Open Low — used as modifier, not primary output

Data Update Frequency

Dataset Update cycle Current version How to update
UCDP GED Annual (typically March/April) v26.1 (data through 2025) Replace data/ged261.csv; update endpoint version string
ACAPS/INFORM Monthly Live via API Automatic (API client with session cache)
GeoNames Continuous (crowdsourced) Live via API Automatic (/api/city-population)
Historical cases Manual 31 cases (1991–2024) Edit historical_data.py; re-run calibrate.py

Out of Scope (deliberately excluded)

  • Real-time conflict monitoring or event streaming
  • Individual-level civilian tracking or identification
  • Legal or operational advice of any kind
  • Mortality forecasting beyond the 0–90 day calibration window
  • Non-conflict natural disaster scenarios (different cost and mortality structures)
  • Conflict forecasting or escalation prediction