Would you send these goggles into production?

What the agent made

Ran it for real with OMP 18.10.4 on 2026-10-14. 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 OMP made, full screen

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

Make your own

1. Get the files

   Download the practice folder (ZIP)

   Unzip it: you get a folder called cli-practice with 3 practice files. Only the Army's dated field account and budget exhibit, Microsoft's announcement published on the decision date, and the previously published FY2020 operational-test report. Later oversight, soldier-test results, and appropriations arrive in a new session. It also has a for-later folder with the new files for the fresh sessions.

   SOURCES.csvproduction.txttesting.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

   Work only inside this practice folder with testing.txt, production.txt and SOURCES.csv. Do not change inputs, read outside this folder, access the network, install packages, or ask for a reply. Use Python 3 standard library only. Do not use em dashes or en dashes. Your historical decision is frozen at the end of March 31, 2021. In public copy call the department the Department of War (DoW); preserve historical source titles exactly when quoting them.
   
   You advise an independent acquisition reviewer: does this public record justify an unconditional full-production commitment for the Army's Microsoft IVAS headset? Distinguish the announced production agreement from authorization of unlimited purchases, previous prototype testing from production-representative operational testing, an expected delivery from a verified delivery, the R&D budget line from the acquisition agreement, and what cannot be measured from the packet. Do not smuggle in later news or an unverified contract ceiling. Give a reasoned call and the specific test evidence and decision thresholds you would demand before full production. A favorable soldier response to some functions and observed defects both deserve a fair hearing. Do not guess physiological injury rates, test results for an untested configuration, cost paid, or the number of units purchased.
   When attributing a named person's position, quote the dated original and link the source; do not infer motive. Distinguish a reported field exercise from a scored, production-representative comparative operational test.
   
   Write only audit.py, analysis.json, check.py and analysis.html. audit.py must extract evidence and funding quantities from the original source text, construct the dated test/decision ledger and arithmetic, and produce analysis.json with source IDs and URLs, an explicit public-record call, requests for evidence, and unknowns. Separate claimed, measured, expected, and not yet shown in structured records. check.py must independently read inputs without importing or executing audit.py, recompute the budget comparison and delivery chronology using a different method, validate every decision-critical JSON value and tagged HTML representation with source citations and hrefs, reject treating expected devices as delivered, mixing development dollars with the agreement ceiling, a fabricated contemporaneous agreement value, and a claim that prototype tests establish full-rate effectiveness. Implement check.py --self-test with at least three distinct in-memory corruptions, each of which must fail its usual validation path.
   
   analysis.html must be a designed, self-contained offline evidence room, not a generic dashboard. Lead with the actual call and a strong visual contrast between the announcement and what had been tested. Include a compact source-linked inline SVG chart of the budget's comparable development-year dollar values, with dollars as units, zero baseline, readable labels, IDs and underlying full numbers in ordinary semantic HTML. Every reader-facing monetary amount must use complete dollars with a dollar sign and comma separators, including tables, chart labels, SVG <title> and <desc>, and accessible text; never describe a converted amount as "thousand dollars" or print a source-table number as though it were dollars. Original source units may appear only in a clearly labeled verbatim source quotation. Put source-linked proof next to each critical finding. Give the reader a meaningful touch- and keyboard-usable control to switch between the announcement's claim and the matching test evidence; activation must change visible consequential text, not just styling, and without JavaScript the complete finding and source list remain visible. Mark uncertainty without implying later findings were knowable.
   
   Make the result look designed for its subject: the visual motif is an acquisition flight-readiness checklist that a reader can challenge. Lead with the consequential finding, then show exact evidence and an achievable next test gate. Use restraint, readable type, strong focus states and no decorative animation. Use only source files in this folder. Build one decision-focused, source-linked visual from values actually parsed by your program, not decorative motion. Show units, dates, source IDs, and all denominator assumptions; never mix development dollars and agreement values in a single series. The complete baseline and source links must be present as normal HTML when JavaScript is disabled. No remote font, script, stylesheet, API, or image. At 390 pixels wide, the page fits the screen. No word or amount ever breaks mid-word; narrow tables become stacked cards. At desktop width retain a readable measure. Avoid horizontal scroll at both 390 and 1280 pixels.
   Check CSS box sizing: make every card, chart and side-by-side frame border-box, set grid children min-width: 0 and max-width: 100%, and account for padding inside the viewport. A 390px page with 401px-wide panels fails even if its body has no intentionally wide table. Do not solve overflow by clipping text or breaking words.
   
   Run python3 audit.py, python3 check.py and python3 check.py --self-test and report the 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: ST1 records 50 CS1 systems received, 300 CS2 systems received, 600 CS3 systems received, and 1,600 CS4 systems expected in April 2021: expected is not tested or delivered on March 31. ST3 is Army development funding $63.200 million FY2020 and $8.991 million FY2021, down $54.209 million or 85.7737 percent: neither number is the value or obligation on Microsoft's award. ST1 explicitly says the prototype STP 2 plan was not submitted for test-office approval, the test strategy was draft, and full-rate evidence was still to be collected. ST2 announces production without a dollar ceiling. Reject a blind-page $21.88 billion claim because this packet cannot verify a contemporaneous primary figure. Run independent check and three corruption tests; inspect source links, stage distinctions, and no hindsight.

   When we checked the recording above: Genuine clean OMP openai-codex/gpt-6-sol run exit 0, inputs unchanged. Independent check.py and four-corruption self-test pass. THOUSANDS_AMOUNT finds zero matches in recorded analysis.html. Browser widths 390/1280 show zero overflow; JavaScript-off 390 exposes finding and linked sources; SVG accessible title/description use full dollar amounts.

Then prove it: fresh sessions

Each run below is a brand-new session, recorded for real in OMP. 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 practice folder. Read the prior frozen sources and outputs, then outcomes.txt and OUTCOME_SOURCES.csv. Do not change any input or any prior output, read outside this folder, access the network, install packages, or ask for a reply. Python 3 standard library only. Do not use em dashes or en dashes. Public copy calls the department Department of War (DoW), but quote historical source titles exactly as published. The earlier analysis remains a March 31, 2021 decision, not a prophecy.

Write only reveal.json, reveal.html and reveal-check.py. Compare two explicitly dated columns: what could have been known by March 31, 2021 and what later primary records establish. Extract the inspector general's audit date and stated concern, the 2022 operational demonstration and its soldier/comparative-performance and reliability findings, and the FY2023 procurement-request changes from the final appropriations table. Distinguish the maximum potential agreement value in later oversight from obligated spending, the original IVAS procurement request from the full Night Vision Devices account, the specific reduction from the remaining funding transferred to research, and the implications for IVAS-specific procurement from that request. The later test cannot retroactively prove a 2021 user-acceptance rate, and changes in budget are not themselves proof of a precise cause or cancellation of all future IVAS activity. Acknowledge what remained unknowable at the original decision.

reveal.html must be a beautiful self-contained offline two-column evidence room with a legible chronology, a source-linked inline SVG waterfall or horizontal bar visual for the comparable FY2023 funding components, ordinary semantic HTML holding all exact figures with scripts off, and a consequential native disclosure or touch control revealing cited test details. Opening it must change visible evidence. Every reader-facing monetary amount must use complete dollars with a dollar sign and comma separators, including table cells and labels, chart labels, SVG <title> and <desc>, and accessible text: $424,253,000, $334,229,000, and $400,024,000, not "424253 thousand dollars" or "$424,253 thousand". Original source units may appear only in a clearly labeled verbatim source quotation. Preserve the 2021 baseline call unchanged in its column. No external requests. At 390 pixels wide, the page fits the screen; no word or amount ever breaks mid-word; narrow tables become stacked cards. Give columns a readable stacked phone layout and a calm two-column desktop layout.

reveal-check.py independently parses both source sets, without importing or running audit.py or reveal generation scripts, recomputes decision-critical funding figures and original development arithmetic, verifies values and source URLs in reveal.json and tagged HTML, rejects 2022 test claims in the 2021 column, rejects claiming an award ceiling was paid, and rejects using the entire Night Vision Devices account as an IVAS-only appropriation. Its --self-test must exercise at least three distinct in-memory corruptions. Run python3 reveal-check.py and python3 reveal-check.py --self-test and report exit codes.

Do it for real

For a live equipment decision, obtain approved test plans, production-representative comparative trials, reliability and human-factors evidence, actual obligations, and soldiers' informed feedback. Protect operational details and personal health information before sharing a packet with any AI tool. This public historical exercise is not a substitute for acquisition, clinical or battlefield safety review.

Why an agent, not a chat window

The agent cross-checks a public test chronology against a budget transition and an award announcement, writes a source-derived evidence ledger, and independently checks its own arithmetic and citations. A fresh session reconciles later oversight, operational testing and enacted funding without rewriting the earlier call.