Issues

• Nightly full eval is red: infrastructure failure

  infra-failurenightly-eval-failure
  ChelseaKR/fare-policy-assistant#257 · Filter by author github-actions opened on Sep 30, 2026

• Mutation (advisory) has reported success over a crash for five weeks, beside a published 75% score nothing derives

  ChelseaKR/fare-policy-assistant#246 · Filter by author ChelseaKR opened on Sep 13, 2026

• evals.chelseakr.com has no robots.txt, sitemap, meta description, or canonical

  ChelseaKR/fare-policy-assistant#244 · Filter by author ChelseaKR opened on Sep 13, 2026
  1

• A price the corpus deliberately withheld is scored as a price the assistant invented: 53 cases, unchanged answers

  bug
  ChelseaKR/fare-policy-assistant#242 · Filter by author ChelseaKR opened on Sep 10, 2026

• Publish reduced-fare eligibility as cited facts, so a rider can compare eighteen agencies without being told they qualify

  enhancementfeature
  ChelseaKR/fare-policy-assistant#232 · Filter by author ChelseaKR opened on Sep 7, 2026

• evals.chelseakr.com has been frozen at the 2026-07-12 build for eight weeks because the publisher concludes skipped, not failure

  ChelseaKR/fare-policy-assistant#230 · Filter by author ChelseaKR opened on Sep 7, 2026
  2

• Add a plain-language readability check with EN and ES floors to the eval harness

  enhancementfeature
  ChelseaKR/fare-policy-assistant#218 · Filter by author ChelseaKR opened on Sep 6, 2026
  2

• Redact incidental PII from a question and answer the redacted form, not refuse

  enhancementfeature
  ChelseaKR/fare-policy-assistant#217 · Filter by author ChelseaKR opened on Sep 6, 2026
  2

• Compute a cited pass-versus-single-ride break-even from parsed fare facts

  enhancementfeature
  ChelseaKR/fare-policy-assistant#215 · Filter by author ChelseaKR opened on Sep 6, 2026
  3

• Scope a deployment to one agency so an embed answers only for its host agency

  enhancementfeature
  ChelseaKR/fare-policy-assistant#214 · Filter by author ChelseaKR opened on Sep 6, 2026
  2

• Add a local browser worksheet for judge relabeling and native-Spanish rating

  enhancementfeature
  ChelseaKR/fare-policy-assistant#213 · Filter by author ChelseaKR opened on Sep 6, 2026
  3

• P3-5: the harness-generalization remainder is tracked in two archived repos

  roadmap
  ChelseaKR/fare-policy-assistant#204 · Filter by author ChelseaKR opened on Sep 5, 2026
  3

• P2-5: the manual screen-reader and keyboard walkthrough is still unfilled

  roadmap
  ChelseaKR/fare-policy-assistant#201 · Filter by author ChelseaKR opened on Sep 5, 2026
  2

• Nightly full eval is red: quality gate failure

  nightly-eval-failurequality-regression
  ChelseaKR/fare-policy-assistant#198 · Filter by author github-actions opened on Sep 5, 2026
  28

• The nightly full eval was red for fourteen straight nights (2026-08-22 to 2026-09-04) and nothing surfaced it

  bug
  ChelseaKR/fare-policy-assistant#184 · Filter by author ChelseaKR opened on Sep 4, 2026
  3

• adversarial suite is red by design: floor 0.90 is honest, the 0.0 measurement is a harness vocabulary gap (plumbline#54)

  ChelseaKR/fare-policy-assistant#181 · Filter by author ChelseaKR opened on Sep 4, 2026
  3

• Spanish parity gap of 7.5 points on the 2026-08-22 full live run, and the gate has no waiver for it

  ChelseaKR/fare-policy-assistant#165 · Filter by author ChelseaKR opened on Aug 22, 2026
  4

• Judge calibration sits below three AIEV floors, and the worksheet that would fix it has 0 of 37 rows filled

  ChelseaKR/fare-policy-assistant#143 · Filter by author ChelseaKR opened on Aug 15, 2026
  4

• The public evidence hub still serves the 2026-07-12 report, 201 cases and twelve agencies ago

  ChelseaKR/fare-policy-assistant#140 · Filter by author ChelseaKR opened on Aug 15, 2026
  7

• cross_agency fell from 100% to 57.1% across the eighteen-agency expansion, with nine named failures

  ChelseaKR/fare-policy-assistant#138 · Filter by author ChelseaKR opened on Aug 15, 2026
  6