# Validation record — edition 0.2

5 October 2026. Scope: three small fictional cases matching the published WBS, launch-date and support Reels. No real customer outcome study has been conducted.

## What was executed

Three separate fresh-context Codex agent runs received the portable skill and designated raw sources only. They did not receive the expected answers or the parent audit. Each exported canonical records and a decision brief. The publisher then reviewed the claims against the supplied text and recalculated the reference figures. These are narrow forward execution tests, not blinded comparisons with naive/manual work.

| Run | Reviewed result | Final mechanical check |
|---|---|---|
| Harbor WBS | Approved discovery scope; draft additions excluded; identity reused; customer data/acceptance separate; owners and detailed tests labeled proposed; six supplied packages total 192h; roles 12/24/12/88/44/12; 2.4 capacity-weeks not a date; CRM unresolved | Passed; four disclosed unestimated-row warnings for customer/excluded work |
| Three launch dates | 2 Nov approved conditional target; 26 Oct unapproved draft; 9 Nov unsigned conditional proposal; gate omission and credentials open; sponsor decisions proposed; −7/+7/14 calendar-day differences | Passed; no warnings |
| Support review | 30 vs 60-day gap; hours/channel/acknowledgment/restoration/escalation uncommitted; listed base 2,400 vs stated 2,500, variance −100; optional 3×150=450; conditional combined 2,850; tax/currency/term unresolved | Passed; no warnings |

The downloadable examples preserve those outputs, with the final checker report rerun on the packaged sources. WBS references unlabeled calibration inputs through disclosed labeled copies. Sources and output hashes provide reproducibility, not proof of authentic approval.

## Friction found and corrected

The checker originally rejected blank economic inputs despite the instruction to preserve unknowns. It now permits blanks with warnings and prevents a known total from including an unknown amount. The record instructions now explicitly place proposed responsibility/acceptance tests in the obligation fallback field and scenario assumptions in the economic description. The checker now reports included role totals. These revisions were checked against the saved outputs and negative/unknown-value fixtures; the original forward tests preceded these small checker/reference changes.

Limitations that remain: unlabeled documents need a disclosed labeled extraction; a checker pass cannot certify semantic coverage, true source authority, complete dependency/calibration links, schedule realism or contractual interpretation. Intentional unestimated rows produce warnings. Actual approval documents may be absent even where a supplied file states approved. Questions and proposed owners still need a manager.

## Release checks

Skill frontmatter validated with skill-creator quick_validate.py. Packaged examples pass the final offline checker. Negative fixtures test incorrect arithmetic, a changed source hash, fabricated quotes/locators, missing ID links, approval without evidence and source-path escape; unknown-value fixtures preserve missing amounts and prevent misleading totals. PDF layout, site downloads, mobile/desktop controls and deployed asset contents are checked separately during release. Release evidence is retained locally; consult release-notes.md for the final scope.

## First ChatGPT UI walkthrough — 5 October 2026

The three core fictional cases were executed in separate ordinary chats in a signed-in ChatGPT account, using the displayed Chat mode and Medium setting. The exact model identifier was not exposed. Only raw designated sources and canonical schemas were supplied; completed examples and answer keys were withheld. The WBS and proposal used file uploads; dates used pasted source text after access errors. Sources pasted into both date conditions had identical XML boundaries and schemas. Project execution and the account's existing personalization were not separately controlled.

| Actual run | Result before external correction | Review intervention |
|---|---|---|
| WBS, published instructions | Final downloadable eight-file set passed the kit's independent checker: 21 exact excerpts, source hashes, linked requirements, six packages totaling 192h and role totals 12/24/12/88/44/12. Manual review retained all four journeys, exclusions, reuse, clean-data cap, proposed tests and unresolved CRM/acceptance authority. Five unestimated rows correctly produced warnings. | No external content correction; all three normal stages were sent. ChatGPT corrected its own dependency assumptions in the verification stage. |
| Dates, additional instruction headings | Source authority, all dates and conditions, sponsor ownership and calendar arithmetic were preserved. Independent checking failed two coverage rows referencing undeclared source ID ALL. | One coverage correction; the corrected set passed with three intentional missing-hash warnings for pasted sources. A separate text-export request was needed to inspect final records locally. |
| Dates, plain instruction paragraphs | Same source text, XML boundaries, schemas and task instructions. The final text export passed independently with three missing-hash warnings. Scope, authority and commitment blockers were retained. | No external content correction; the same text-export request was used for review. |
| Proposal, published instructions | Correct service gaps, base 2,400 vs stated 2,500, variance -100, optional 3x150=450 and conditional combined 2,850. The final independent checker rejected two invented header-locator labels. Manual review also found USD asserted where only the symbol $ was supplied. | One external correction for locators and currency. Corrected records passed independently: eight exact excerpts, source hashes and arithmetic, with three intentional incomplete-input warnings. Currency fields retain $, and the ISO code is an open supplier question. |

The original results were retained before feedback. The portable checker was not attached to ChatGPT; any model-generated checker was distinct from the kit's actual checker run by the reviewer. Early sign-in/session errors prevented generation until a normal page refresh; some code/file steps errored and were retried by ChatGPT. WBS and proposal ZIP downloads were retrieved; the structured date package required a text fallback. Browser automation failed to report download events even when native downloads existed. No task-time or savings comparison is inferred from these runs.

The heading comparison is one fixed-order pair on a small case, not a blinded or repeated experiment. It does not show that headings improve reliability: the plain condition passed before correction, while the extra-heading condition needed a schema correction. Both already had identical XML source delimiters, so this does not test XML against undelimited text.

Post-walkthrough changes add headings inside copyable instructions and explicitly require declared coverage IDs, exact printed header markers and unresolved ISO currency codes when only a symbol is supplied. These changes respond to observed failures. They have not received a new fresh complete UI walkthrough; a corrected existing conversation is not independent validation of revised prompts.

## What remains unverified

- Broader ChatGPT compatibility, Projects execution, Claude and Gemini full UI executions, and a Claude custom-skill import. Setup routes are based on current official documentation, not tested compatibility guarantees.
- Repeatability across cases, configurations and models; adversarial or large-document coverage.
- Five target managers completing unassisted and a blind comparison against naive prompting and manual work.
- Actual time savings, superior professional outcomes, repeat use or voluntary sharing.
- The full ten-issue Harbor reconciliation extension; its answer key is authored reference, not another execution result.

Focus next on actual service walkthroughs and independent manager usability, then compare outputs and full task time before scaling outcome claims.
