Drafted 2026-07-01. Status: ideas for evaluation, not commitments.
A third planning layer for the fare-policy assistant, produced from a fresh read of the whole repository (source, evals, CI, corpus, docs, git history). The two existing layers are:
docs/ROADMAP.md— the original build spec and its productionalization phases (P0–P3), largely executed.- The 2026-06 research pass —
docs/research/synthetic-personas-feedback.md(R0–R3 backlog, mostly executed) and, on the still-unmergedresearch-panel-and-roadmapbranch,docs/RESEARCH-ROADMAP.md/docs/USER-RESEARCH.md(RR1–RR10, RE1–RE7).
This folder deliberately does not restate any item from those documents. Where an idea builds on an existing item it references the item's ID (P1-4, R3-1, RR6, RE5, …) and goes beyond it. Everything here is net-new: structural fixes and expansion bets the existing roadmaps do not contain.
| File | What it holds |
|---|---|
01-deep-dive.md |
Current-state assessment from this read: architecture, strengths, observed structural debt, portfolio position |
02-large-scale-fixes.md |
FIX-01 … FIX-12 — deep structural fixes with effort, risk, and a measurable bar for each |
03-expansions.md |
EXP-01 … EXP-14 — expansion ideas in three horizons (deepen core / adjacent / transformative) |
04-impact-and-sequencing.md |
Impact×effort matrix, dependencies, a Now/Next/Later sequence, and the honest list of human/legal/SME/credential gates |
The repo's ethos binds every idea here: honesty as a feature, liability
candor, reproducibility, accessibility, multilingual equity, and the hard
rules in CLAUDE.md (no eligibility determinations, no
PII, every answer cited, corpus dated). Several ideas are explicitly gated on
things this environment does not have — live Bedrock credentials, a human
with a screen reader, legal counsel, agency partners — and
04-impact-and-sequencing.md separates those out rather than pretending they
can be closed by code alone. Where a claim below rests on something not
verified in this pass, the text says so.