v0.1.1

• Package version is 0.1.1.
• Removed the remaining Spanish from the claim detector, the briefing text checker, the writer prompt, the writer module note, and the tests and eval that planted Spanish sample sentences. Those checks now use English sentences. Publish-claim detection and relative-time checks still cover the English wording.
• Restore CI workflow after history reset
• Dropped repository-visibility lines from the README, SECURITY.md, docs/ADAPT_TO_A_REAL_CLIENT.md, docs/PROVENANCE.md, docs/SCOPE.md, and the 0.1.0 note. SECURITY.md tells reporters to use GitHub private vulnerability reporting ("Report a vulnerability" on the Security tab).
• Reworded the README use-case row, the docs index entry for docs/HOW_IT_WORKS.md, and the discovery paragraph so they read as a walkthrough for engineers and buyers.
• Checked Markdown for Spanish. docs/HOW_IT_WORKS.md is already English, and no other Markdown file was still in Spanish. Numbers, commands, links, and the labels [EST], [SYN], MEASURED, SIMULATED, and UNVERIFIED are unchanged.
• Tests remeasured on this docs change, Python 3.12.3, pytest 9.1.1, package 0.1.0, APPROVAL_STATE_KEY unset: 56 passed. Statement coverage 91.83% (2068/2252). pytest-cov total with branch coverage on: 88.51% (floor 80%). ruff check, ruff format --check, and mypy reported no issues.
• Translated docs/HOW_IT_WORKS.md from Spanish to English. Commands, numbers, [EST], and links are unchanged. The README docs index no longer calls that walkthrough a regional variety of Spanish.
• README is English, with a three-line Start here, a business-value section, a Proof section, and "What this does NOT prove".
• Added docs/BUSINESS_VALUE.md. Money figures are the existing [EST] lines from docs/BUSINESS_CASE.md. No new volume was added.
• Tests remeasured 2026-10-22, Python 3.12.3, pytest 9.1.1, package 0.1.0, APPROVAL_STATE_KEY unset: 56 passed (55 on 2026-10-17, plus one docs-contract test). Statement coverage 91.83% (2068/2252). pytest-cov total with branch coverage on: 88.51%, the same total recorded on 2026-10-17.
• LICENSE is the standard MIT text, copyright Diego Lescano. The note that NYC Open Data snapshots are not relicensed moved to NOTICE.
• Briefing text now gets each incident date and an as-of date, and retries when counts, plurals, dates, or relative time do not match.
• The publish step is an approval gate (simulated actor): HMAC state, approver allowlist, tamper rejection. It is not an authenticated channel.
• Publish-claim check is a weak verb-plus-object heuristic with a negation window. Previous regex, 2026-10-17, n=20, small: 0/20 paraphrases and 3/20 negation false positives. External verification VERIFY_REPORT_9, same day, n=20, small, not conclusive: 7/20 recall and 3/20 false positives. This lab has not re-run that corpus.
• Sensitive categories are excluded from the public briefing and map.
• The offline map includes an OpenStreetMap tile attribution control. Tiles are still not requested.
• briefing --live and briefing --online are explicit. The default briefing reads snapshots only.
• Live cost uses cost_in_usd_ticks (1 tick = 1e-10 USD) when the provider returns it. The pre-call estimate ignores cached tokens.
• SODA has a per-fetch budget and a refresh budget. A partial snapshot fallback is labeled mixed, not online.
• Incident dates in the prompt and the checker use the America/New_York calendar day.
• Live cost_usd sums cost_in_usd_ticks across retries, including a rejected attempt.
• The approval gate (simulated actor) binds the HMAC to the program id and a one-time nonce. The public lab key is refused outside demo mode.
• Borough counts in the briefing text are checked. A category count scoped to a borough is only partially checked.
• The prompt includes borough counts computed by the code. The model is told to copy them and not to count. A category count scoped to a borough is only partially checked against that borough. False positives remain, and the form "Brooklyn saw 4…" is not covered.
• The text checker is not a reliable control. External verification VERIFY_REPORT_10, 2026-10-17, n=19 incorrect texts, small sample, not conclusive: 13/19 detected and 1/10 false positives. External verification VERIFY_REPORT_11, 2026-10-17, a different set (24 incorrect texts and 15 correct texts), small sample, not conclusive: 13/24 detected and 3/15 false positives. The two sets are not comparable. This lab did not re-run either set (not measured yet). The checker does not cover "Brooklyn saw 4…", invented casualties, or invented superlatives.
• briefing --live also requires LLM_BASE_URL and LLM_MODEL. If either is missing, the command exits 2 and names it. It does not call a model.
• External verification VERIFY_REPORT_11, 2026-10-17, model grok-4.20-0309-non-reasoning: 5 of 6 briefing --live runs accepted model text, and 0 were rejected for a borough count (n=6, small sample, not a rate). This lab did not re-run that.
• SOCRATA_RUN_BUDGET_S bounds only the SODA refresh. MAX_LIVE_SPEND_USD is the money cap for the model. There is no single wall-clock cap for briefing --online --live.
• The approval gate (simulated actor) binds the state HMAC to the nonce-ledger head. Deleting the ledger fails closed. Restoring the ledger together with an old APPROVED file still republishes. It is not an authenticated channel.