Backlog
Ranked list of work, parked items, and open methodology questions. The v1 cycle plan is ordered — work top-down. The parked queue is not ordered; items graduate to the cycle plan when their predecessor lands or when a parked decision is made.
This page is the session-start reference for "what's next." Pair
with the latest handoff in #esg-screening for current cycle state.
Last refreshed: 2026-05-19. Refresh discipline: update at the end of any session that completes or reorders an item.
v1 cycle plan
The goal of v1 is a viable web-based screening product: coverage ~35–45% on the financials pilot, banks no longer scoring identically, watchlist populated with real findings.
1. Commit ADRs 0001–0006 + institution-detail mock
Mechanical CC-on-VM cycle. Eight files from the 2026-05-19 chat session
plus ADR-0006 from the ops repo, committed to docs/adr/ on the main
repo. Currently the next-up cycle.
References: ADRs 0001, 0002, 0003, 0004, 0005, 0006.
2. Migration 011 — coverage-weighted scoring
Adds raw_v04 and coverage_weighted variants to score_composite
and score_pillar. Backfills historical runs. Existing composite
becomes an alias of composite_coverage_weighted. Pilot banks move
from ~19.2 to ~66.7. This is the change that makes scores semantically
meaningful.
Reference: ADR-0001.
3. Read interface — pages 1 and 2
Institution index (/) and institution detail (/institution/:lei).
Server-rendered HTML, Express + EJS or Handlebars, no build step. RAG
colours per ADR-0004. Rule grouping per ADR-0005. Filtering state in
URL query params. Index page layout needs its own design pass — the v4
mock covers detail only.
References: ADR-0002, ADR-0004, ADR-0005, institution-detail mock v4.
4. Migration 012 + /methodology page
Adds type, access_mechanism, status, external_url, priority,
feeds_rules to the source table; backfills from ADR-0003's initial
source list. Builds the /methodology route with three sections:
scoring explainer, source register (live), rule catalogue (live).
Nightly job updates source status from scrape_run outcomes.
Reference: ADR-0003.
5. BHRRC scraper + watchlist surface
Business & Human Rights Resource Centre. Per-company HTML scrape, allegation list parsing. Feeds S2, S5, E7, G5 plus watchlist. First scraper that should actually differentiate the four UK banks. Watchlist schema designed during this cycle.
Reference: scraper backlog in ADR-0003 (priority 1).
6. UK MSA + UK GPG scrapers
Modern Slavery Act register and Gender Pay Gap register. Low-effort government registers, both can land in one cycle. Financial and non-financial applicability.
Reference: scraper backlog in ADR-0003 (priorities 3, 4).
7. BankTrack + PAX scrapers
Financials-only. Fossil-fuel financing (BankTrack) and weapons exposure (PAX). Together they should add the differentiation that turns the financials peer table from identical to meaningful.
Reference: scraper backlog in ADR-0003 (priorities 5, 6).
Parked queue
Not ordered. Items graduate to the cycle plan as their predecessor lands or as a parked decision is made.
- v1.5 client-facing summary surfaces — separate styling pass for the read interface for client-facing exports. Deferred until v1 is in use.
- SBTI-DASHBOARD → SBTI-VALIDATED rename — tiny migration. Source IDs misleading (source is Excel, not dashboard/CSV).
- PRB seed expansion — currently 4 UK banks. Fill to ~350 from the canonical UNEP FI list before pilot expands.
- IFRS-S2-CORP scraper — corporate website / sustainability report scraping. Heavy engineering, own ADR when ready. Replaces the TCFD-shaped gap with disclosure-quality signal.
- Stage 2 credit + returns dimensions — schema placeholders exist, data sources don't. Stage 2 produces no scores yet.
- Confidence-weighted scoring layered on coverage-weighted — open methodology question. Coverage-weighted handles the "is this rule wired up" axis; confidence handles "how good is the signal we have." Whether and how to combine them is undecided.
- Auto-render ADRs into ops site — currently the
/decisions/*pages are hand-written summaries. Auto-rendering from the main repo'sdocs/adr/would remove the drift risk. Small cycle. - Auto-render CLAUDE.md into
/reference/claudemd— same pattern. Small cycle. - Watchlist schema — to be designed during cycle 5 (BHRRC).
- Trim project instructions and CLAUDE.md content duplication — /backlog reference added 2026-05-19; broader trim of session-protocol content duplicated across project instructions, .claude/CLAUDE.md and ops site /operating-model still parked.
- Four data-model open items — raw payload retention, watchlist
storage shape, source-vs-signal_source rename, run retention.
Partially addressed in
/architecture/data-model, not ADR'd. Possible future ADR-0007. - Weekly automated refresh — cron exists (Sunday 02:00 UTC, gated
by
ENABLE_SCHEDULER=1) but not yet enabled and no monitoring around it.
Open methodology questions
Distinct from parked work — these need a decision before they can become work.
- Coverage-weighted vs confidence-weighted composition. See parked queue.
- Stage 2 credit and returns data sources. No live candidates yet.
- Aggregate-rule coverage semantics. G3.1 reads other rules' signals but writes none of its own; coverage = 0 on the aggregate. Defensible but worth revisiting if the pattern recurs.
How this page works
This page is hand-maintained until the auto-render plumbing lands. End-of-session discipline: if a session completes an item, reorders the v1 plan, or moves something into or out of the parked queue, update this page in the same session.
The canonical hierarchy still applies: ADRs > design note > migrations
v0.4 workbook. This page summarises and links; it does not override.