Methodology and Sources

1.1 The deterministic cost formula

All reconstruction cost estimates are computed deterministically by formula. The AI orchestrator used on this platform (which classifies damage records and generates narrative summaries) never produces numeric output. Every number in every cost field derives from lookups against published benchmark tables, computed in scripts/compute-costs.js.

The {low, central, high} triple is generated by walking the low and high ends of each multiplier range simultaneously — not by applying the formula at the central value of each input individually:

• low = formula(all low-end inputs simultaneously)
• high = formula(all high-end inputs simultaneously)
• central = arithmetic mean of low and high

This is consistent with the RDNA3 uncertainty-range methodology and avoids artificially narrowing the range by mixing low-end and high-end assumptions across different inputs. [RDNA3]

1.2 Heritage and cultural asset premium

Assets classified under heritage.* carry a separate conservation-premium multiplier applied after the standard formula. This multiplier reflects the cost of traditional materials, specialist craftsmanship, archaeological monitoring, and UNESCO and Council of Europe conservation standards that are not captured in standard unit costs.

The heritage premium is held in a separate table in data/unit_cost_table.json and is not folded into the general regional multiplier. Any asset receiving the premium displays it explicitly in the "show cost formula" panel on the asset detail page.

1.3 Three reconstruction paths

Every asset is costed across three paths. The path multiplier adjusts the unit cost and unlocks different technology overlays and financing structures.

The build-back-better path costs more in absolute terms but typically attracts a higher grant percentage, compressing the concessional and equity burden on Ukraine's own balance sheet. All three paths are presented side-by-side so users can evaluate the trade-off directly. [EU_FACILITY] [RDNA3]

1.4 Anti-hallucination architecture

A hallucinated reconstruction cost cited in a publication by EBRD, EIB, or a policy team would represent a reputational failure. The architecture eliminates that risk structurally rather than through prompt instruction alone.

1. Stage 1 — Classification. The LLM extracts structured fields from OSINT input: asset_type, physical_specs.*, damage.destruction_level, wartime_status.*. All numeric specs are flagged with source provenance. Uncertain values default to pending_data. Temperature: 0.2.
2. Stage 2 — Retrieval and narration. The server performs all deterministic lookups and computes the cost formula. The LLM receives the complete structured payload and narrates it. It may rephrase; it may not introduce numbers absent from the payload. Temperature: 0.4.
3. Validation gate. Every numeric token in the narration is extracted via regex and matched against the payload (with formatting tolerance). Any unmatched numeric token aborts publication and routes the entry to human review. Dates are whitelisted; financial and physical-quantity tokens are not. Implemented in functions/lib/validation-gate.js.

System prompts are version-controlled in functions/_shared/prompts/ as Markdown files. No prompts are hardcoded in JS string literals.

3.1 Tranche taxonomy

Ukraine's reconstruction capital stack is more granular than a simple grant/debt split. The platform uses a 12-layer taxonomy ordered from most concessional to most commercial. Every reconstruction deal can be expressed as a combination of these tranches; percentage allocations within a deal must sum to 100. Political-risk insurance (PRI) is modelled as a credit enhancement — a boolean flag with named providers — not as a capital tranche percentage.

Public-sector and grant tranches

DFI and IFI tranches

Sovereign and private-sector tranches

3.2 Template assignment — sector × path matrix

Capital-stack templates in data/financing_templates.json are keyed by sector × path. When an asset record specifies both, the platform selects the matching template as the standard stack, with per-asset overrides permitted where named comparable projects justify deviation. Central-tendency grant percentages across the matrix:

Full templates including all 12 tranche allocations, PRI wrap flags, and structure patterns are in data/financing_templates.json. Template rationale for each sector is documented in docs/financing_tranches.md.

3.3 Wartime adjustment rules

In pre-armistice conditions, commercial capital tranches are materially compressed relative to a peacetime template. Rules from data/wartime_adjustment_rules.json:

PRI wrap is required on all retained commercial bank debt, institutional debt, and private equity in the pre-armistice state. The rationale is risk pricing: in active conflict, commercial lenders face loss scenarios that are not insurable at standard premiums without IFI first-loss and guarantee layering. Adjustments are informed by EBRD Ukraine portfolio deal structures and bilateral ECA utilisation rates. [EBRD_CASE] [MIGA]

In post-armistice fragile conditions, compression eases to 60% retention for commercial debt and private equity. Full drawback to peacetime templates requires durable peace and a sovereign-risk band of low or elevated.

3.4 Confidence levels

Low-confidence tranches carry the widest uncertainty bands and are most sensitive to EU accession trajectory and peace-state scenario. The Finance It tool exposes scenario toggles for these variables.

4.1 Historical precedents

4.2 The sovereign wealth fund parallel — Norway's Government Pension Fund Global

Norway's GPFG is the closest structural analogue to a Ukraine Reconstruction Trust. The Fund holds approximately USD 1.6 trillion in assets (2024) derived from North Sea oil revenues, governed under a rule that limits annual drawdowns to the expected real return — initially 4%, revised to approximately 3% of fund market value. Annual parliamentary appropriations may not exceed this limit except in extraordinary circumstances. The rule creates a fund that can deliver a predictable annual disbursement without depleting the corpus.

The structural parallel: a corpus derived from a windfall asset (frozen Russian reserves in place of oil revenues), governed under a drawdown rule (2.5%–5.0% depending on political will), converting the corpus into an annual availability payment that is known, predictable, and schedulable. This is precisely the feature that Marshall Plan experience shows is most valuable to reconstruction planners: not the size of the transfer, but its predictability.

The GPFG parallel also illustrates the governance requirement. Norway's rule works because it is legislated, administered by an independent manager (Norges Bank Investment Management), and subject to annual parliamentary scrutiny. A Ukraine Trust would require analogous governance — multilateral trusteeship, independent management, and regular public reporting — to provide the credibility on which institutional investors would rely when financing reconstruction projects against anticipated Trust disbursements.

4.3 The Trust mechanism as modelled on uVidNova

The annuity arithmetic illustrates why the Trust structure matters. A 4% drawdown on USD 286bn equals approximately USD 11.4bn per year — a contracted annual stream, unlike the uncertain and politically fraught prospect of a lump-sum settlement. At a 5× financing multiplier, that stream supports approximately USD 57bn in reconstruction investment per year without requiring any political agreement on the total reparations amount.

In the current platform default (proceeds only, pre-armistice), the ERA envelope is approximately USD 50bn total — a one-time financing against accrued interest, not an annual stream. The principal-seizure scenario (USD 280–300bn corpus) is what unlocks the annuity logic and its financing multiplier. Users can toggle between these scenarios in the Finance It tool.

4.4 Disclaimer — Trust mechanism