# Validation and provenance

Research performed September 9, 2026. This page identifies what was actually checked and
where an implementation claim remains limited.

## Sources and collection

The source video is [xmGY276gEFY](https://www.youtube.com/watch?v=xmGY276gEFY), titled
*Claude Code's creator has some really good advice*, published July 21, 2026 by Theo. Its
duration is approximately 19 minutes 12 seconds. The complete captions were read; useful
ideas were extracted and reconciled against timestamps and visible frames. Segment
boundaries are approximate caption timestamps, not frame-accurate edit points.

The downloaded 1920×1080 video has SHA-256
`d81e3d3c29a6e02c9798e6fe37d362589e00d72ee4dcb95ab624f43e1a400944`. The video itself is
not redistributed in this repository.

For each frame, ffmpeg performed a timestamp seek and exported one JPEG:

```bash
ffmpeg -hide_banner -loglevel error -ss 383 \
  -i source/xmGY276gEFY.mp4 \
  -frames:v 1 -q:v 2 screenshots/06-23-upload-skill.jpg
```

(Paths above are relative; the extraction used a local copy of the video retained outside
this repository.) All 12 frames were visually inspected at full resolution. [The frame
manifest](evidence/frame-manifest.json) records each command, timestamp, output hash and
observation. The upload-skill screen at 06:23 is visible evidence of instructions, not
proof that an upload succeeded.

## Repository snapshots

| Source | Inspected revision | Scope |
|---|---|---|
| T3 Code | `6c583620ff7ad3235b135af7107c0543467eecfa` | Selected architecture, custom rules, launcher, tool adapters and PR evidence |
| T3 Code at end of publication day | `23c18fda7a969634a30888e36b2da45f6d66a83b` | Historical instruction, lint and test configuration comparison |
| Theo's Melee fork | `a276aeb70f9879204d891d967f1c9442523568e1` | Fork comparison, all 13 PR records, selected tools/native architecture and verifier execution |
| Original Melee verifier | `035d9711623a32fbe891cffdcf44a91a550c1947` | Exact linked commit and its nine verifier tests |
| Matt's Course Video Manager | `4c1f3f5d49417e54b185bfe737b1e1a56f29c7b8` | Latest 100 PR records, selected diffs and source; no application execution |
| Boris's json-schema-to-typescript | `5caacfc53671f9c891bb4e2a78bccc6190ed3ef4` | Compiler stages, validation, benchmarks, CI and selected PRs; no compiler test execution |

Moving PR/branch pages can change after this snapshot. Large PR file lists can be paginated
or truncated; the reports do not claim complete line-by-line review of every change. PR
performance figures remain author-reported measurements.

## Executed behavioral checks

**Lint demonstration.** Node 22.22.3, ESLint 10.10.0. `npm run demo` succeeded after
installing the lockfile dependencies. Without the restriction, the offending import passes
with exit 0. With it, ESLint exits 1 and reports exactly the forbidden UI-to-database
import; the approved API import produces no finding. The driver treats an execution/config
error as failure. [Red evidence](examples/recurring-rule/evidence/red.txt), [green
evidence](examples/recurring-rule/evidence/green.txt),
[reproduction instructions](examples/recurring-rule/README.md).

**Melee verifier.** `python3 -m unittest discover -s tools/tests -p test_verify.py -v`
passed at the original verifier snapshot (9 tests) and current fork (15 tests). A separate
fixture gave both verifiers matching synthetic bytes and successful mocked build/diff
commands, but incomplete source metrics. The initial version accepted it; the current
version rejected it. [Comparison result](evidence/melee-verifier-comparison.json).

The Melee build/diff subprocesses were mocked. No Nintendo assets were downloaded and no
game, native renderer or performance benchmark was executed. These results establish
verifier behavior only. The public applications, upload service and any private agent
systems were not executed.

## Static package checks

The final edition checks, run locally against this repository:

- All Markdown links and image targets inside the rendered pages resolve to files that
  exist; no link escapes the repository or, in `public/`, the generated root.
- Exactly 13 guide pages, 4 skill entrypoints and 12 full-size frames are present, with
  all screenshot SHA-256 hashes matching the extraction manifest.
- The renderer (`build/render.py`, Markdown pinned to 3.10.3 in
  `build/requirements.txt`) runs without network access and
  produces byte-identical `public/` output on a second run (checked by directory
  checksum before and after).
- The four `SKILL.md` files have valid Agent Skills frontmatter, and the
  `agents/openai.yaml` files parse with a display name, short description and a default
  prompt that references their own `$skill-name`.

These are repository and renderer checks, not tests of the inspected third-party
applications. See [CONTRIBUTING.md](CONTRIBUTING.md) to rerun them.

## What is not claimed

- No full transcript or copy of the video is redistributed, and no third-party repository
  is cloned into this repository.
- The handbook itself was deployed and checked as described below. The inspected third-party
  applications and upload service were not exercised.
- Sponsor claims and universal career/productivity claims remain unverified.
- Qualifications in the guides (for example, that a skill shown on screen does not prove a
  real upload succeeded) are carried through to the [detailed
  extraction](evidence/video-research.md) and the structured
  [ideas file](evidence/video-tips.json).

## Published handbook checks

The initial [public site](https://agent-engineering-handbook.vercel.app/) was deployed from this
repository on September 9, 2026. Vercel installed the pinned Markdown dependency,
regenerated the site and ran the package check. Anonymous HTTP requests returned 200 and
byte-identical content for the landing page, handbook, gallery, critical-journey guide,
feedback skill and all 12 screenshots. Desktop (1280×720) and mobile (390×844) inspection
found no page-width overflow; the live gallery had no failed loaded images and the guide
produced no console warnings or errors. Off-screen lazy images were verified separately
by their downloaded bytes.

The September 9 layout revision was inspected locally at desktop (1440px) and narrow
(390px and 320px) widths. The inspected pages had no horizontal overflow. The mobile
menu opened and closed with Escape, returning focus to its control; the guide contents
links reached their sections, and the desktop rail marked the selected section. The
guide produced no console warnings or errors. The gallery had no failed loaded images.
The integrated output matched the inspected build before this validation note was added.
The existing package check passed all 27 HTML pages. Design and interaction instructions
remain contributor source files and are excluded from the published output.

GitHub Actions is enabled on the repository, but the publishing token lacks its workflow
permission. An optional freshness workflow is retained under
`build/handbook-workflow.example.yml`; it is not an installed CI check. Vercel runs the
renderer and package check on deployment.