Two reviewer-driven hardening pushes since the last update. On fillable forms, the pipeline now infers each field's type from the source instead of defaulting to a plain text box: dates, multi-line answer boxes, and single-select yes/no…

Phase-1 review flow shipped. Primary Data table shows ⚑ needs-attention when any constituent invoice was Haiku-flagged. Mark Reviewed checkbox persists per row; auto-clears when a contributing invoice is edited. Drawer shows per-invoice…

Drawer polish: Advanced-view toggle hides dormant Queue Info and Phase-4 enrichment. Auto-capture button promotes Haiku's new-fields-seen into extraction_data._haiku_additions (audit-trail, out of Josh's Primary schema). Accept-all batch…

Aggregator dedup: prod Firestore was silently ingesting duplicate invoices (same PDF, casing-drift vendor names). All row totals were 2×. Dedup by (vendor_canonical, account, invoice_date, total_amount) — caught 180 duplicates on Cloud…

UI aligned to Multco BRANDING.md palette (#316094 steel blue). New Pill primitive with 6 WCAG-AA severities. Dashboard rebuilt around Phase-1 KPIs (intake health, coverage, deduped, ⚑ flags) instead of dormant AP-workflow metrics.

The agent-written test now drives the real application — end-to-end grounding confirmed live against the QA environment. Earlier work taught the generator to write the right form fields; the gap that remained was navigation — the…

First working end-to-end Primary Data flow shipped: aggregator turns raw extractions into Josh's Utility_Primary Data row shape (42-column schema, ADR-002). WM lookup expanded 45.5% → 93% coverage. Review-grid + provenance drawer +…

Alignment complete — VISION, ARCHITECTURE, DATA_MODEL, ROADMAP, CHANGELOG, and 8 ADRs authored. Direction pivoted to Josh's Utility_Primary Data as the target (ADR-001) instead of the Smartsheet intake tab.

Closed the biggest structural gap in the converter: fillable forms had no output check at all. Form validation only ever inspected the extracted field list — never the finished web page — so accessibility scanning could not run on a form…

A six-week blind spot in the knowledge base — found, closed, and guarded. Fifteen hand-written scenario documents describing how UCR actually behaves had lived in the repository since late May but were never loaded into the database the…

Auth-drift root-caused as the mirage behind a 93-test 'code regression'. Every test using the external agency-staff fixture (8 top-level probes, 57 tests) flipped from 100% pass to 100% fail overnight with identical waitForURL:45s…

Five reliability fixes closed the gaps a large hands-on review exposed — each shipped with automated tests and deployed to production. Scanned, image-only PDFs (which give the converter no readable text to work from) can no longer slip…

First end-to-end MEDIC pipeline validation on live baseline data — cluster → LLM proposals → deterministic apply → net +18 passes in the target cohort. Post-baseline orchestrator kicked automatically after v18 landed (ADR 0032 · task…

A hands-on review of a 41-document batch — spanning new document types the pipeline hadn't seen before — hardened it against real-world variety. Fourteen documents were approved (several after small, precisely-worded editor fixes) and…

Measurement you can trust, plus a reviewer loop that actually learns. A new source-grounded completeness gate measures every converted document against the original PDF itself — not against the converter's own output — so silently dropped…

URL/tab root-cause fix — every editreferral test 404-ed since inception. +298 net non-permmx passing tests across two days. Chrome walk of one live QAT referral (2026-07-10) revealed the corpus's URL patterns don't exist. Every…

Renamed to Civic Sentinel; the presentation now carries a live nightly metrics strip (files classified, connectivity, cleanup candidates, content-overlap) fed automatically from the pipeline's own outputs

The self-improving loop now closes on formatting — the pipeline's single biggest remaining source of hand-corrections. A full grounded review (the Advisor's own data + the actual reviewer corrections + a read of the deployed code) found…

Vision extractor + sweep-CLI + Tier 1 production hardening — three shippable wins in one evening. (1) Claude vision on Vertex now converts image-only Jama Explorer sessions into workflow narrations. Pilot on Case Management Business…

Q-SEED-1 closed on the role axis: 23 of 23 authoritative UCR roles pinned in the seed catalog. The permission-matrix and state-machine specs that were skip-guarded on missing seeded state now execute. The seed catalog lives in…

Three grounded forms fixes shipped from a full pass over the Advisor's hand-fix backlog. The two-model panel proposed eight prompt rewrites across the recurring forms problems; on review, every one restated instructions the 273-line forms…

The Pipeline Advisor's Insights board was redesigned around one honest number. It previously reported a headline 'success' figure that counted every saved reviewer keystroke as a separate document — inflating the denominator so the number…

Fleets #1-5 shipped same evening: 5 pieces of plumbing so agents actually compose cross-corpus at generation time. Post-rollback Loren surfaced the real diagnosis: the corpus exists (311 permission_matrix.cases, 647 razor_form_fields, 264…

L4 rollback + "keep pushing" turned 0/5 Search into 2/5 same-day. The v3.4 full-batch reconvergence (37 spec files across the 4 hardest UCR areas) landed at 0% pass rate and got rolled back — the R7 rule taught "target the tab anchor" but…

The self-improving loop now reaches the pipeline's prompts, not just its code. Many recurring problems on hard documents come from the AI's reading of the page (a scanned form invents a placeholder label; a set of meeting minutes runs…

The system now diagnoses and fixes its own recurring problems. The Pipeline Advisor was rebuilt around the real signal — the hand-fix burden: every recurring reviewer correction (e.g. 'duplicate table', 'missing section heading') is…

Self-serve admin-perms diagnostic landed: 4 role codes unblock 32 specs, verified end-to-end. The platform's own data answered the long-outstanding test-user-provisioning ask. WebUI/MVC/SiteMapProvider.cs + Web.sitemap map every UCR menu…

The platform automatically surfaced 5 real legacy bugs in UCR QAT — including 3 that share a root cause. Mon-AM GUARDIAN baseline triage classified the 77 failures by error pattern + page snapshot: 5 hit a real "Server Error 404 —…

Honest Monday baseline: 66.2% pass excl. skips on the cleaned 210-spec / 324-test suite (was 72.2% on the 06-25 cohort). GUARDIAN run 1c72e8f0 landed at 151 ✓ · 77 ✗ · 96 ↷ in 47.4 min. Pass count is identical to the 06-25 baseline (151)…

The strict PageObject allowlist now actually constrains the live generator. The v3.2 system prompt (shipped Friday) instructed the model: "if a method isn't in pageobject_catalog.md, it doesn't exist — emit a TODO instead of guessing."…

Two reliability improvements. (1) Image sizing: images inside converted documents — embedded graphics as well as auto-described slide/map/diagram images — were rendering at full resolution and overflowing the text column; they now scale…

Retired the Xingwu PHP-audit cross-check — it was the initial one-time orphan-detection validation; ongoing Validation now tracks the Siteimprove Content Quality snapshot

Eight of nine holistic test categories now populated on prod (was six Friday). Sunday plowed through Monday-prep work that doesn't need VPN: a structural T-SQL parser extracted 264 stored-procedure contracts from UCR_app/**/Stored…

Library redesigned as a compact, sortable list (was large cards) — far more documents visible at a glance, sortable by name/quality/date, with search and a one-click 'Copy link' to share any document with a colleague. Guide and Help…

Editor-experience + audit-depth sweep: shareable deep-links + Copy-link across Analytics/Accessibility/Queue, bulk approve-reject + paste-many dictionary, a personal "My queue", skip-link/dialog a11y, PII name+DOB/address co-occurrence…

Semantic page intelligence: nightly embeddings (text-embedding-005) power a Content Overlap report (754 clusters of near-duplicate content — incl. real Drupal duplicate nodes), meaning-based search, similar-pages, and an Ask Agent…

"What changed since yesterday" — nightly snapshot diff surfaces new/removed files, status transitions, and per-bucket deltas on the Home dashboard and via a new /summary/changes API

Reliability hardening: output-validation guard turns the nightly build red + alerts when any artifact is missing/corrupt or the Vertex narrative falls back — ending the class of silent failures that masked stale data

Fixed three latent bugs from a full code review: link-checker crash (broken-links report was silently stale), a duplicate method returning empty group stats, and a dormant deploy config that could unset the dashboard's reload token

Form fixes from a deep per-form review: the Opt-In/Out form now shows its OPT IN explanatory text (a brittle keyword filter had been silently dropping it every run, which is why prior editor fixes never stuck); and a systemic…

Six of nine holistic test categories populated on prod (was one). A morning self-serve unlock (postgres root password was in our own Secret Manager all along) applied two schema migrations (008 + 009) without needing the Cloud-SQL-admin…

Major fix: fillable forms were silently losing most of their fields. A cleanup step was deduplicating form sections by their heading text and, when several checkbox groups shared a heading (e.g. a permit application with 85 checkboxes…

Fillable-form field ordering: many county forms embed a tab-order that doesn't match the visual top-to-bottom layout, so converted fields appeared out of reading order. Forms with significantly scrambled order are now re-sorted to read…

Cohort cleanup + streaming reporter shipped same-day. ADR 0023 Phase 1 streams test results live to the dashboard. 207 legacy drafts retired, 111 stale spec files deleted, full GUARDIAN re-fire on the cleaned 170-spec suite landed 151…

MEDIC gets a dedicated inbox page. Pending · Approved · Applied · Rejected lanes, expandable diff preview, one-click approve dispatches the apply workflow. Was previously buried inside the OPERATOR briefing card.

Platform Home rebuilt around tenants, not metrics. The 4 KPI tiles came down (they were misleading on a multi-tenant view). New 'Today's Focus' surfaces 1–3 action bullets; tenant cards show pass rate, fail count, MEDIC queue, draft…

Rikki Thunstrom's final knowledge-transfer touchbase captured. Two PDFs and a 60-minute Gemini transcript ingested through the file-upload UI; auto-fire produced 22 new drafts. The platform now carries her procedural knowledge in…

Workflows page wired to live data (500 workflows). Retired the hardcoded 17-stub list. One tap now connects an SME interview to the spec file on disk.

Forms quality pass: reviewed every fillable form against the cleanest one (the Metro/County Opt-In form) and lifted the shared form renderer so all forms benefit. Section headings no longer appear twice (once as a header and again as a…

Document-quality polish from a full-batch review (47 documents): real document titles now replace filename slugs and generic section headings; bullet lists crammed inside a single table cell render as proper lists; scanned engineering…

Review-screen trust improvements: the document Info panel now shows only independently-measured quality signals — accessibility (WCAG) and real content coverage measured against the source PDF — replacing earlier decorative bars that…

Phase 23 list-reconstruction fixes shipped: numbered lists that previously restarted at “1.” for every entry now render in correct sequence (1, 2, 3, …), and bullet blocks that collapsed into a single paragraph of literal “●” glyphs are…

MEDIC v0 first auto-applied selector patch shipped. Read 1.3K characters of Razor source, identified the wrong locator pattern, proposed and applied the fix as a draft PR.

Global Assembly layer (Phase 22) online: a deterministic reconciliation pass that re-derives clean document structure from noisy extraction — cross-representation de-duplication, numeric section ordering, document-title resolution, and…

ADR 0020 Phases 1+2+3 shipped in one afternoon. UCR_app legacy source spine moved into Cloud SQL so MEDIC can read it without a checkout. Azure DevOps source-change webhook wired.

Batch-2 output-quality pass (13 fixes) shipped after a 6-agent vision review: learned reviewer edits now persist (a heading fix that quoted a year was being silently discarded every run); accessible-PDF links no longer duplicate; form…

Output consistency pass shipped to production: every converted document now draws from one standardized Multnomah County brand palette (tables, headings, callouts and footers) instead of each renderer drifting into off-brand colors…

knowledge_v2 baseline + MEDIC build gate opens. First end-to-end run of the SME-document → executable-test pipeline. MEDIC architecture green-lit.

Outputs Hub launched: a permanent, all-staff document library that turns A11yReady from a one-shot conversion tool into a system of record. Any county staffer can browse every converted document and download its accessible HTML or PDF…

Form fidelity (H219/H219b): section headers now merge shaded bands with bold text headers (a table-column guard keeps med-table column names from masquerading as sections), and repeating cell-per-widget grids collapse into a single…

95.5% baseline pass rate vs UCR QAT. First green full-suite run: 168/176 passing across 71 specs.

OPERATOR v0 shipped — the 6th agent. Deterministic rule engine reads cross-agent signals and produces a daily morning briefing with concrete proposals.

Forms structural overhaul (H207–H211b): source section headers, professional field naming, page-primary multi-page ordering, fillable tables, and character-comb field merge — driven by real reviewer rejections

Vertex narrative parser now tolerates markdown code fences; dictionary extended 120 → 181 entries — both effective on tomorrow's nightly

Feedback → directives loop (H208): reviewer edits now change outputs via safe, per-file deterministic fixes instead of being write-only

Analytics page redesign: Errors → Diagnostics, Languages rebuilt on page-side data, Visualizations retired (unique charts merged into Analytics, redundant ones moved to Departments)

AcroForm anchoring + deterministic reading order (H203–H206): semantic field roles, role-driven sizing, and the “Other: ___” companion attachment

Cloud Run min-instances bumped to 0→1; Workbench crash fix + RouteErrorBoundary shipped

Section-overview UI restructure: collapsible sidebar, six top-level sections, every section has an overview page

Pipeline Advisor + intelligent per-cluster model routing

Editor attribution: page editor enrichment + Owner column on every report + Mine-only toggle

Per-department 30-day Health-score sparklines added to Departments page

Validation page polish: Siteimprove category reconciliation with month-over-month deltas

Initial one-time validation against the Drupal team's independent PHP audit: 92.3% orphan-detection agreement on 33,738 UUIDs (cross-check since retired)

Production launch — AI-native political news platform with 7 autonomous agents, 17 cron jobs, multi-source synthesis with bias auditing, and fully automated cross-platform social posting.

Production launch — 9-agent LangGraph SaaS platform producing complete go-to-market playbooks in under 10 minutes. Quick scan at ~3 min, full report at ~10 min.

Production deployment serving Multnomah County agencies