PACTA + TRISK Vietnam — Bank Climate Alignment & Transition-Risk Platform

An end-to-end demonstration platform that shows how a Vietnamese bank can measure loanbook climate alignment (PACTA — Paris Agreement Capital Transition Assessment) and stress-test transition risk at borrower level (TRISK), then turn the results into engagement and disclosure deliverables aligned with Decision 263/QĐ-TTg, TCFD, and ISSB.

All data is synthetic. Every number comes from the fictional "Mekong Commercial Bank (MCB)" case — synthetic loanbook, synthetic asset-based company data, and adapted PDP8 / NDC / IEA NZE scenario pathways. Nothing here is a real assessment of any institution.

Live demo: https://pactavn.streamlit.app

Architecture

scripts/run_engagement.R is the single orchestrator for every engagement, including the public MCB demo. scripts/pipeline_refresh.R is a thin compatibility wrapper that delegates to it with engagements/mcb-demo/engagement_config.json (run weekly by .github/workflows/refresh.yml); every flag it receives (--full, --dry-run) passes straight through.

scripts/generate_vietnam_data.R       synthetic inputs (loanbook, ABCD, scenarios)  [--full only]
        │
        ▼
scripts/pacta_vietnam_scenario.R      PACTA alignment vs PDP8 / NDC / NZE
        │                             → synthesis_output/vietnam/
        ▼
scripts/trisk_prepare_inputs.R        TRISK input packages per sector
scripts/trisk_sector_demo.R           TRISK stress runs (power / cement / steel)
scripts/trisk_scenario_grid.R         243-scenario sensitivity grid per sector
        │                             → synthesis_output/trisk/
        ▼
scripts/sector_prioritization.R       sector ranking (alignment × stress × exposure)
scripts/refresh_dashboard_data.R      publish frozen snapshot → dashboard/data/
        │
        ▼
dashboard/                            Streamlit app (7 pages; 3 operator-gated)
        │
        ▼
scripts/engagement_scoring.R          borrower engagement priority
scripts/generate_engagement_letters.R per-borrower letters
scripts/generate_disclosure_pack.R    TCFD / ISSB / Decision 263 pack
        │
        ▼
scripts/generate_refresh_audit.R      refresh audit report                          [mcb-demo only]

The whole chain, in order, is orchestrated by R/step_runner.R's run_steps(), invoked from scripts/run_engagement.R's config-driven step list (see "Running a client engagement" below).

Quick start

Prerequisites: R 4.5+ (Windows path assumed below) and Python 3.11+.

# Python deps (dashboard)
python -m pip install -r dashboard/requirements.txt

# R deps (pipeline)
& "C:\Program Files\R\R-4.5.2\bin\Rscript.exe" scripts/ci/install_deps.R

# Rebuild the TRISK chain + dashboard snapshot (writes pipeline_manifest.json)
$env:Path += ";C:\Program Files\R\R-4.5.2\bin"
Rscript scripts/pipeline_refresh.R

# Tests
python -m pytest dashboard/tests

# Run the app
python -m streamlit run dashboard/app.py

On Linux/macOS use plain Rscript on PATH. Always run from the repo root — all R scripts resolve paths relative to the working directory.

Refactor acceptance check

Any change touching scripts/ or R/ must leave every default-mode CSV output byte-identical to its pre-change content (PNGs and generated-timestamp text are exempt). Verify this with one command instead of hand-diffing:

$env:Path += ";C:\Program Files\R\R-4.5.2\bin"
Rscript tools/verify_refactor.R

This reruns scripts/pipeline_refresh.R and classifies every tracked file git diff reports as changed into expected churn (timestamps/manifests), known-volatile noise (a named list of TRISK CSVs with a regenerated run-id column — currently empty, since PHASE-04 made every run-id deterministic), or genuine drift. It prints BYTE-IDENTITY PASS and exits 0 only when no genuine drift is found. Pass --full to check against a pipeline_refresh.R --full run, or --skip-refresh to classify the current working tree without rerunning the pipeline.

Byte-identity answers "does run N+1 match run N?" — it is structurally blind to a cache that never regenerates, or two committed artifacts that silently disagree with each other. A second, independent mode answers that question instead:

$env:Path += ";C:\Program Files\R\R-4.5.2\bin"
Rscript tools/verify_refactor.R --invariants

This checks five cross-artifact consistency rules against the current tree without running anything: the TRISK scenario grid's base-parameter cell must equal the base (non-grid) TRISK run; no scenario vintage may exist at two paths; every engagement's data_source column must equal its own bank_slug; the supported-sector literal must agree across every file that hardcodes it; every published TRISK manifest sector must be a known registry sector. It prints INVARIANTS PASS and exits 0 only when every invariant holds. Both .github/workflows/ci.yml and .github/workflows/refresh.yml run this on every push and every weekly refresh.

Where each check runs. ci.yml runs both modes on every push and pull request: a byte-identity job runs the default (non---full) byte-identity check, and the r-tests job runs --invariants. refresh.yml runs --invariants after every weekly refresh, then runs byte-identity in --skip-refresh mode as a gate immediately before the auto-commit step — any genuine drift stops the workflow before dashboard/data and synthesis_output are pushed. To land an intentional, reviewed refreeze, re-run refresh.yml manually via workflow_dispatch with allow_drift: true, which skips only that gate step.

Running a client engagement

scripts/run_engagement.R runs the full delivery flow — intake, validation report, PACTA, TRISK, prioritization, snapshot, scoring, letters, disclosure — for any engagement config in one command:

$env:Path += ";C:\Program Files\R\R-4.5.2\bin"
Rscript scripts/run_engagement.R --config engagements/<slug>/engagement_config.json --raw-loanbook <path/to/loanbook.csv>

system2("Rscript", ...) needs Rscript on PATH even when the outer call used a full path — set $env:Path first as shown above (same requirement as scripts/pipeline_refresh.R). Add --dry-run to print the resolved step list without executing or writing anything — useful for checking which steps a config will run (e.g. whether the scenario grid is included) before committing to a full run. Omit --raw-loanbook to run against the loanbook already named in the engagement config, or add --skip-intake to skip intake and re-run only the downstream stages. mcb-demo is the only engagement allowed to publish into the public dashboard/data snapshot; any other engagement's snapshot_dir must point elsewhere, or the run refuses to start.

Repository map

Path	Contents
scripts/	R pipeline stages (incl. generate_vietnam_data.R), report generators, pipeline_refresh.R orchestrator
R/	Shared R modules sourced by scripts/*.R and loadable as the pactatrisk package: engagement config loader, sector registry, report toolkit, matching helpers, PACTA core functions
engagements/	Per-engagement config (engagement_config.json); mcb-demo/ documents the built-in defaults
data/	Synthetic Vietnam input CSVs (generated by scripts/generate_vietnam_data.R)
synthesis_output/	Pipeline outputs (PACTA, TRISK, scenario grid)
dashboard/	Streamlit app: app.py, pages/, lib/, tests/, frozen snapshot data/
intake/	"Bring Your Own Loanbook" schema contract + templates
pilot/	Bank-agnostic pilot conversion pack ({{BANK_NAME}} tailoring slots)
docs/	Methodology guides, deployment notes, assumption registers
reports/	Rendered self-contained HTML reports
plans/, research/	Implementation plans and research briefs
attic/	Superseded methodology-reference scripts, kept for historical context only

Operator features (env-gated, off by default)

Flag	Feature
BYOL_INTAKE=1	Intake Wizard (page 6): validate + map a client loanbook
OUTPUTS_LAYER=1	Outputs (page 7): engagement letters + disclosure pack
TRISK_LIVE_RERUN=1	Live TRISK rerun in the Scenario Builder (never on Cloud — no R runtime)
PILOT_ANALYTICS_ENDPOINT	Anonymous page-view counter (see docs/streamlit-deploy.md)
R_RSCRIPT	Path to Rscript for the R-calling features above

Methodology & caveats

• PACTA: r2dii.data / r2dii.match / r2dii.analysis (RMI); TRISK: trisk.model (Theia Finance Labs), after Baer et al. (2022).
• TRISK NPV/PD figures are illustrative stress indicators, not production credit-model outputs.
• Cement and steel use sector-level SDA context (not borrower-specific alignment); steel match coverage in the synthetic book is low.
• See docs/TRISK_Demo_Assumptions.md, docs/trisk_multisector_contract.md, and dashboard/data/README.md for the full assumption registers.

License

Proprietary — see NOTICE.md. Upstream R packages retain their own open-source licenses.