Five kinds of evidence
| Kind | What is here | What it cannot show |
|---|---|---|
| Made-up practice | 20 cases, 66 claim findings, and four accurate drafts. | Real AI error rates, real employer errors, or actual approval to publish. |
| Examples from real sources | Three dated examples, 18 claims, and four original source pages. | That jobs are open now, company claims are independently proven, or the tool works well in all uses. |
| Software tests | Checks of the review rules, files, saving, and browser controls. | Whether a person judged a real claim correctly. |
| Agent checks | Separate checks of content, sources, NIST links, and files. | A human study or proof of human approval. |
| Human study | Not done. No participants were recruited or observed. | No human success rate, feedback, or time savings are claimed. |
Do not combine these into one accuracy score. The answer key is a teaching guide. It is not test data for an automatic error detector.
What happens to your data
Choose how to keep each review. Save on this device keeps a copy in this browser. The page says saved only after the save finishes. Open saved reviews lists the copies in this browser. When you choose one, device saving resumes for it.
Private session keeps new work in this tab and stops new device saves. It does not remove a copy already saved. Removing a saved copy is a separate choice. The toolkit keeps no browser copy until you choose Save on this device or open a saved review.
Nothing you type in the review is uploaded. The toolkit has no account, AI service, analytics, or remote review database. It does not put your review text in a web address or cookie. Device copies use this browser’s own storage. Normal page and download requests still reach the host, which may keep access logs such as your IP address and the file requested. Opening a source sends a request to that source. Your browser, extensions, and device are outside the toolkit’s control.
Keep a separate backup file too. Where your browser supports it, Link a backup file lets you choose a file to update while this tab is open. Otherwise choose Download backup. Backup-file status and device-save status are separate. Check that you have the file and keep it somewhere you can find again.
Clearing this browser’s site data can erase device copies. A separate file remains if you keep it. Private-session work can be lost when the tab closes. No tool can recover work after every copy is erased. Use only the information you need, especially on a shared device.
The worksheet shows entered text as text. It makes links only from safe http or https addresses with no hidden sign-in details. You can save unfinished addresses and dates, but they cannot pass an evidence check. A review file may be up to 8 MiB (8,388,608 bytes). The tool checks its structure and linked records. If a file fails to open, the review already in the tab stays in place.
Review files and workbooks can be edited. Their history and file hashes help you inspect them. They do not prove who acted, when they acted, or that no one changed the record. NIST action MS-2.8-003 is therefore marked Not covered.
Sources: SRC06: MS-2.8-003; content provenance.
What a review record keeps
A review has its own ID and a current version ID. It keeps the original and final draft, use, date or period, audience, and reviewer. It also records whether the reviewer checked the full claim list. Sources, claims, issues, decisions, and saved versions have IDs that link them together.
A web source needs a safe address. A document, interview, email, or other non-web source needs a useful record reference, owner, and place to find the evidence. Keep the relevant evidence text. Do not invent a website. A person still needs to check the source.
A Material claim could change a reader’s choice. Unsure is treated the same way. Each such claim that stays in the draft needs evidence, all three source judgments, a reason, a result, and a check of the exact final words. A removed claim keeps its removal check. Do not invent a claim just because a source leaves out a fact such as pay.
The field dictionary explains each saved field. The JSON schema gives the file structure. The import code also checks links between records and whether past decisions fit their saved versions.
Which version does a decision cover?
Ready means the required checks are done and someone recorded that decision. Approved needs a separate decision by a named human. These are ArizonaAI workflow rules, not NIST rules for certification. The name and computer clock are not independently verified.
Each decision keeps a full copy of the inputs it used. Later edits change the current version, so an old approval cannot approve the new draft. Changing the full draft clears its completed claim-list check. Changing a source clears its check and linked claim checks. Changing the use, audience, or date context requires new timeframe judgments.
The tab keeps your current edits, including an unfinished decision reason when you switch steps. It is not a record of every keystroke. Each decided version stays in the history. A backup also includes the current version. Keep a separate backup before a major edit if you need that undecided version too.
You can discard an unused source or claim card. Removing an item you used needs a reason and keeps earlier versions. If a current claim still needs a source, unlink or replace it with care first.
Version 1.0.0 review files still open. Their original decisions and saved versions stay intact. The repaired rules check current readiness again. An old approval that fails those rules is shown as history and does not approve the current draft.
The workbook needs manual version records. Before a decision, save a dated full copy and fill in Revisions. After an edit, clear the old approval, check the work again, and keep the new copy. Its formulas only provide small counts. They do not enforce the web worksheet’s approval or history rules.
Choose the right backup or download
- JSON backup: the full review, including unfinished fields, an unfinished decision reason, all decisions, and saved versions. You can download one before checks are done. Open the file in the worksheet to keep working.
- Audit CSV: rows for reviews, sources, claims, issues, decisions, and versions. The payload_json column keeps the full data for each row. Use it to inspect the record. Keep the JSON backup too; the worksheet does not import CSV.
- Review summary: the final draft, checks, open issues, and current decision. This is useful for discussing publication. Keep the full backup for its history.
- Full audit print: the current record, missing checks, decisions, and earlier saved versions. A large review can make many pages.
- Workbook: manual forms, examples, answers, and references. It cannot be opened as a JSON review.
CSV adds an apostrophe before text that a spreadsheet might read as a formula. The JSON keeps the original text. Check a downloaded file before opening it in another app.
The 8 MiB limit includes file formatting and all saved versions. If an edit would exceed it, the tool refuses that edit and keeps the last accepted record. Back up the record and start a separate follow-up review that names it. Do not delete decision history to make space.
Review format and toolkit version: 1.1.0. Event times include a timezone offset. The page shows Arizona time. A date with no time is just a calendar date. Unknown source dates stay unknown.
Practice dates and addresses are made up. Recording a check in practice is an exercise. It does not mean you opened a real website.
What has been checked
Version 1.1.0 has passed the recorded repair and saving checks. The records below name the versions and files tested. The owner approved this release on October 8, 2026. The CC BY 4.0 content and MIT code licenses remain active. Nothing has been sent to NIST.
The dated test records below keep their original wording and counts. They apply only to the versions they name.
| Area | Result | Evidence |
|---|---|---|
| Version 1.1.0 saving, recovery and clearer instructions | Recorded checks passed; publication approved October 8, 2026 | The website build passed 1,970 code tests, including 39 file-saving tests. A real backup file kept later edits and decision history after browser data was cleared and the browser was closed and opened again. Earlier checks passed 24 review cases, 25 recovery cases and two window-focus cases. Separate checks covered the new guide and workbook. Four final file requests confirmed the tested preview and its updated downloads. These are checks of the recorded versions, not a study of how people use the tool. Earlier failed attempts and their limits remain in the full test record. |
| Version 1.1.0 audit repairs A01 through A11 | Local repair checks passed; deployment proof recorded separately | 86 engine tests passed. The full local browser suite passed 24 tests; final interaction polish passed 5 targeted tests. Coverage includes blocked empty Minor approvals, four original 1.0.0 recovery files with unchanged history, partial and near-capacity save/reopen, focused typing and pending explanations, safe discard, all nine branded pages, first-field visibility, guided errors, separate completed examples and practice, compact summaries, final-text copy, short names and non-web evidence. Axe 4.10.3 reported no violations across nine pages and three worksheet stages in Chromium 155. These scripted checks and independent code review do not constitute a human study or complete accessibility certification. |
| Retained 1.0.0 baseline: Decision engine and import/export rules | Passed final local run; exact receipt retained | 50 engine tests passed, 0 failed, on Node v22.22.2. Cases cover blank defaults, material Holds, rechecks, edited-input invalidation, snapshots, relationships, imports, Unicode, every ASCII control prefix before a CSV formula, and Arizona dates. Targeted ESLint reported 0 errors and 0 warnings in all three implementation scripts. Commands, exits and tested hashes are retained. |
| Retained 1.0.0 baseline: Browser worksheet workflow | Passed final local run; exact receipt retained | 10 browser tests passed in Chromium 155.0.8059.39 on Node v22.22.2. Real controls exercised decisions, JSON/CSV/print, retained history after edits, rejected imports, local network/storage behavior, static reading without JavaScript and targeted access checks. Direct-file load, edit and JSON export also passed in this browser. |
| Retained 1.0.0 baseline: Scoped accessibility and reflow | Targeted checks and enumerated needs-review follow-up completed | Axe 4.10.3 found 0 violations in 9 static pages and 3 workflow stages. The toolbar group correction removed the ARIA review item. All 14 remaining contrast occurrences received computed-style, five-point unobscured-hit and scoped visual review: text areas measured 16.22:1; decorative arrows 7.73:1. Keyboard/error focus, 1440/390/320 px widths and enlarged CSS text passed. Complete WCAG conformance, native zoom and formal screen-reader coverage were not established. |
| Retained 1.0.0 baseline: Local input handling and exports | Passed within the tested local scenario | Entering review text caused no requests after initial static loads, no browser-store entries or cookies, and no text in the URL. JSON roundtrip retained Unicode and snapshots; CSV formula-prefix protection and failed-import preservation were tested. Deployed host integrations and user-saved copies remain outside that observation. |
| Retained 1.0.0 baseline: Static artifacts, archive assembly and local links | Static checks passed; archive assembly validated on the recorded snapshot | 2,242 static assertions passed and 323 local links were checked, including the reader ZIP and manifest targets. Separate archive assembly validation checked ZIP CRC, unique safe paths, exact inner and public manifest inventories, and every recorded byte size and SHA-256 on a complete toolkit snapshot. Owner-only packaging preserved the public files and their modification times. The accompanying ArizonaAI_Release_Verification_1.0.0.json identifies the exact final owner distribution. |
| Retained 1.0.0 baseline: Selected NIST source and artifact mapping | Scoped agent review completed | All 18 identifiers/locators and finished mapping records were reviewed, with 47 primary/supporting artifact targets checked. Scope: 14 practical connections to selected aspects, 2 contextual connections and 2 Not covered. This is not external certification, NIST staff review, comprehensive coverage or compliance. |
| Retained 1.0.0 baseline: Dated observed illustrations | Scoped agent review completed | Three observed examples contain 18 reviewed claims tied to four primary-source pages. All 18 internal claim/source links and four external source links were checked. Their dates, attribution and limits remain part of the evidence. |
| Retained 1.0.0 baseline: Fictional practice corpus and answer keys | Authored and checked within the documented scope | 20 original packets, 66 claim findings and 4 valid controls are present. The recorded corpus review and D03 follow-up are separate from a real-world accuracy benchmark. Corpus text and answers are preserved in the workbook and static exports. |
| Retained 1.0.0 baseline: Separate agent J01 walkthrough | Completed retest; two earlier observations fixed | Eight real-UI steps passed in the final source retest, including the named five-button file-action group, original Hold, correction, rechecks, inventory review, explicit fictional Ready and JSON export. The original Hold snapshot remained intact; no approval decision or human participant result was created. Changed toolbar views were inspected, and the unchanged claim views retained their prior review by exact image identity. |
| Retained 1.0.0 baseline: Guide, checklist and mapping PDFs | Document checks and all-page visual review completed | The 38-page guide, 2-page checklist and 10-page mapping were rendered and reviewed on all 50 pages. All 48 bookmark destinations resolved; 446 expected text segments were present. No open visual defect is recorded. PDF/UA and formal screen-reader reading order are not certified. |
| Retained 1.0.0 baseline: Workbook content, formulas and native print | 579 saved-file assertions passed; native print checked | The 12-sheet workbook preserves 20 cases, 66 findings and 18 mappings. All 160 recorded working-input locations are blank; its two formulas are limited diagnostics. All sheets were rendered and 17 selected views inspected or retained by byte identity. LibreOffice conversion produced 154 Letter portrait pages, with no blank/sparse pages or off-page text; all 2,238 long text cells were present. |
| Retained 1.0.0 baseline: Actual human usability and effectiveness evaluation | Pending; no participant results claimed | No human participants were recruited or observed. Agent findings, scripted workflows, layout heights and software timings are not human completion, accuracy, time-saving or effectiveness results. |
| Retained 1.0.0 baseline: Production publication, rights and NIST submission | Version 1.0.0 published; licenses active; NIST correspondence unsent | The owner authorized and published version 1.0.0 with active CC BY 4.0 original-content and MIT original-code licenses. The publication receipt and PR385 record that original deployment. The owner separately approved the 1.1.0 repair release on October 8, 2026. No NIST message has been sent. |
Each test result applies only to the version and files it names. Device saving starts only after you choose it. Private session stops new saves but does not remove an older copy. Browser site-data clearing can erase device copies. Keep a separate backup file. No tool can recover work after every copy is gone.
The full review file is limited to 8 MiB, including history and formatting. The tool refuses an edit that would make it too large. Download a backup, then start another review if needed. Name the earlier review in your notes. Original 1.0.0 files are still supported under the stated file rules.
The sharing tags, fonts, and layouts have limited software checks. No social-platform card preview, publisher study, full screen-reader study, or all-browser certification is claimed.
The 1.0.0 test records and old PDF and workbook files are kept as history. They do not test the current application or new downloads. Scripted and agent checks are not a human study.
Software tests, document checks, agent checks, and made-up exercises are different kinds of evidence. None proves real-world model accuracy, automatic error detection, or measured human results.
Archive and file-hash checks apply only to the exact files recorded. The original final owner package is identified by ArizonaAI_Release_Verification_1.0.0.json. A hash shows which bytes were checked. It does not prove truth or approval.
The accessibility goal is WCAG 2.2 AA. Automated passes and reviewed warnings cover only the stated checks. A 320 px screen width is not a native browser zoom test. Larger CSS text is also a separate check.
The workbook needs manual dated copies and new checks after edits. Review files, names, and times can be edited. They do not prove identity or protect history from tampering.
The original workbook print check used the recorded LibreOffice build. It did not test use in Microsoft Excel, a physical printer, full PDF accessibility, or actual human use.
Production access, site analytics or ads, license terms, and NIST submission are separate matters. No acceptance, endorsement, compliance, public listing, or reply date is promised.
Our accessibility target is WCAG 2.2 AA. Reports state which automated, keyboard, text-size, and layout checks were done. They do not certify full compliance. Read the complete HTML guide if the PDF is hard to use. We do not claim PDF/UA certification.
Tests of this worksheet do not test the instructions for a generative AI system. That task, MS-2.8-004, is Not covered. These checks also do not prove AI reliability, less bias, or better business results.
Reuse, rights, and upkeep
You can use this toolkit for free with no account. Cenk Brown, publisher ArizonaAI.com, licenses the original guide and templates under CC BY 4.0 and the original code under MIT. The owner approved these licenses on October 7, 2026. They remain active. They do not cover third-party sources, trademarks, fonts, code dependencies, your review records, or the whole ArizonaAI website.
Read the content license and credit rules and the full MIT terms. Keep the notices when you reuse files. The source list records past checks. It does not show NIST review. Sending anything to NIST needs separate owner approval. NIST decides what to list and when.
Name an owner for upkeep. Check official guidance changes and fix broken source links. Record changes to answers or decision rules. Choose a review schedule that fits your risks. The toolkit does not run scheduled checks for you.