Would you have opened the health exchange tomorrow?

What the agent made

Ran it for real with Claude Code 2.1.300 and OMP 18.11.0 on 2026-10-15. The first session started with no saved memory; each later session was a new one that kept only the files in the folder. Your result will differ.

Open what Claude Code made, full screen

Then a brand-new session used what it saved on new input: see the proof.

The same prompts in OMP made their own versions, every session recorded and checked: session 1 · Open the dated launch record.

Make your own

1. Get the files

   Download the practice folder (ZIP)

   Unzip it: you get a folder called cli-practice with 5 practice files. Only dated prelaunch GAO, CMS, CGI, and House majority-staff excerpts and their source ledger. Later audit findings and enrollment notes arrive in the separate reveal session. It also has a for-later folder with the new files for the fresh sessions.

   SOURCES.csvcgi-2011.txtcms-2013.txtgao-2013.txthouse-2013.txt

2. Open your agent in that folder

   Open a terminal in cli-practice and type claude or omp --approval-mode always-ask. First time? Set up in about five minutes, with steps for Mac, Windows, and Linux.

3. Paste this prompt

   You are advising a responsible official at the close of September 30, 2013, before the federal health insurance marketplace opens for applications. Work only inside this practice folder with gao-2013.txt, cms-2013.txt, cgi-2011.txt, house-2013.txt, and SOURCES.csv. Do not change source inputs, use the network, install anything, read outside this folder, or ask for replies. Python 3 standard library only. Do not use em dashes or en dashes. Do not use later knowledge or pretend a plan was completed simply because its planned date is past.
   
   Answer one decision: is consumer enrollment demonstrably ready on the public record, and what exact evidence or narrower operating plan should an accountable official demand before opening? Build a dated requirements and readiness table that links the public promise to eligibility, data exchange, plans, consumer assistance, capacity and contractor scope. Separate a contemporaneous claim, a planned deadline, observed completion, independently verified pass, and unknown. Identify the strongest case for opening as well as what cannot be proved from public excerpts. Calculate exact day gaps only for exact dates, not vague months; distinguish CGI's award estimates from obligated dollars and distinguish the majority staff's concerns from verified tests. Do not use a modern failure as a prediction. Name an official only where you quote the exact dated primary record and its URL; avoid imputing intent or blame to individuals.
   
   Write only audit.py, analysis.json, check.py, and analysis.html. audit.py parses the source files, dates and ledger to build a traceable JSON evidence table, timeline gaps, the contract-value distinction, a defensible call and explicit unknowns. Make check.py independently parse the original documents without importing or invoking audit.py; check the decision-critical JSON values and tagged displayed HTML values against source text, ensure every source ID links to the original URL, reject a planned milestone shown as proven passed or a contractor ceiling shown as federal spending, and detect any later outcome in the frozen analysis. Include --self-test that demonstrates at least three distinct in-memory corruptions are rejected without modifying input or output files.
   
   Make analysis.html a complete, self-contained, offline evidence room, not a generic dashboard. Lead with the answer and an annotated launch-eve requirements table; include an inline source-linked SVG dated timeline with a true date axis, dated labels and source IDs rather than an invented completion curve. Give the reader a meaningful source-inspection control that opens actual quoted evidence behind a particular readiness question, while the complete baseline table and all decision-critical dates, numbers, citations and unknowns remain ordinary semantic HTML without JavaScript. Use the source URLs as working links. Art direction: a public-service launch checklist with an editorial heading and one restrained accent; a visible difference between promised, planned and proven, not traffic-light reassurance. At 360 pixels wide and wider, the page fits the screen and nothing scrolls sideways, not even inside a table, chart, timeline or code box; charts and timelines scale to the width; no word or amount ever breaks mid-word; narrow tables become stacked cards; each source shows as a short named link with its URL in the link, never printed as page text. Ensure keyboard, touch, focus and reduced-motion access. No external assets, fonts, requests or CDN. Every chart value must have units, source ID and a corresponding plain HTML row. Keep numerical inputs and any observed conclusions derived from source files rather than remembered answers.
   
   Run python3 audit.py, python3 check.py, and python3 check.py --self-test and state their exit codes.

   The prompt asks the agent to stay in this folder; it cannot enforce that. Read each file change and command before you approve it.

4. Check it yourself

   Open cli-practice/analysis.html in your web browser (double-click it), then:

   Human check: HC1 is dated June 19, 2013, not a September completion certification. Its planned July 31 certification, August 15 Navigator awards and September 15 plan-information load precede October 1 by 62, 47 and 16 calendar days respectively; the month-only data-hub testing target does not support an exact gap. CGI's HC3 announcement distinguishes $55.7 million base award from $93.7 million estimated value; HC1's $87,997,938 is CMS-reported obligations for CGI through March 31, 2013, a different measure. A credible call requires current evidence of end-to-end testing, capacity, plan data, security and contingency readiness, not a statement that GAO proved they failed as of September 30. Run check.py and --self-test and examine citations, both sides of the decision and source inspection.

   When we checked the recording above: 2026-10-14 re-record (360px prompt): exit 0, no changed inputs; in a clean rebuilt folder the lesson checks and self-tests all exit 0; headline call and key figures match the lesson human check and the other tool's run; headless sweep at 360/375/390 with all disclosures open and after every control, then resized 400-1300 in 4px steps: nothing past the edge, no sideways scroller, no printed URLs; controls change visible text; no external requests

Then prove it: fresh sessions

Each run below is a brand-new session, recorded for real in Claude Code. It remembers nothing from the run before except the files saved in the folder, so the short prompt is the whole instruction.

Work only inside this folder. Read your original frozen sources and prior outputs, then the newly added outcomes.txt and OUTCOME_SOURCES.csv. Leave all source and prior output files unchanged. No network, installs, outside reads, APIs or replies. Python 3 standard library only. Do not use em dashes or en dashes. The prior analysis remains a September 30, 2013 decision, not a prophecy.

Write only reveal.json, reveal.html, and reveal-check.py. Recompute the old dates from the prelaunch documents, then compare the frozen official requirements and unknowns with later GAO contract-management and technical findings, the later HHS inspector-general management findings, and the dated preliminary enrollment meeting notes. Identify each source's actual publication date rather than trusting a report number. Keep contract award value, FFM obligations and data-hub obligations separate. Label an early war-room tally by its meeting time and scope; do not present it as a certified final first-day total or proof of causation. Compare what a prudent official could have demanded then with what the later records actually establish; never write that hindsight made a specific person dishonest. Name officials only by quoting a primary document with its date and URL.

Build an offline, self-contained reveal.html with two clearly dated columns, knowable before opening and established later, and a plain HTML evidence table with every decisive date and number visible without JavaScript. Include a source-linked inline SVG comparison of the like-for-like FFM obligation snapshots, with dollars as units, zero baseline, source IDs and the greater-than bound preserved; do not graph unlike contractor award figures together. Give the reader a touch- and keyboard-friendly disclosure of a later official finding beside the earlier quoted requirement that it tests. The finding must visibly change when activated, and the no-script evidence must remain complete. At 360 pixels wide and wider, the page fits the screen and nothing scrolls sideways, not even inside a table, chart, timeline or code box; charts scale to the width; no word or amount ever breaks mid-word; narrow tables become stacked cards; each source shows as a short named link with its URL in the link, never printed as page text. No external assets, fonts, requests or CDN; support reduced motion.

reveal.json records the old evidence boundary, each dated later finding, source URLs, comparable arithmetic and unknowns. reveal-check.py independently reads both packets without importing or invoking audit.py, checks the key JSON figures and tagged HTML values, verifies which side of the date wall each citation inhabits, rejects false final-day enrollment, fabricated causality or a February publication date for a later GAO report, and uses --self-test to catch three distinct in-memory corruptions. Run python3 reveal-check.py and python3 reveal-check.py --self-test and state exit codes.

Do it for real

For a live public-service launch, obtain signed acceptance results across integrations, security, accessibility, realistic load, fallback channels and incident ownership. Publish uncertainty rather than a made-up green status. Protect real applicants' health, income and identity information; require authorized officials to make the opening decision.

Why an agent, not a chat window

The agent extracts dated commitments across independent records, calculates how much time the schedule leaves for integration, and tests whether evidence of completion is present rather than mistaking an expected date for a passed test. A separate session confronts the frozen call with later audits and preliminary enrollment meeting notes.