601 of 6,146 captured versions have reviewed comparisons at this snapshot. Additional execution checks are below; the complete queue continues across every domain.
SEOTwo inactive package copies
Supplemental search-data helpers
73 / 73 repaired-helper cases
The original PAA parser and supplemental-file validator pass 27 of 73 cases. Their inactive repairs pass all 73 cases and 341 checks, including quoted multiline CSV, Unicode, missing fields, strict section markers and size limits. Both complete package copies retain the earlier sitemap repair.
Limits and evidence
Local synthetic inputs only. These results do not establish current search facts, live crawling, WordPress upload or full plugin operation. Fifteen original companion files remain unchanged. Earlier failures and repair attempts are retained.
Result file SHA-256: 932a7c03e537912f4ad485cc5f6dd925a381e355e67e1cfa1cf893379f3cb487
DESIGNAll 18 fresh requests completed
Diagram imports and paired outputs
7 / 12 generated diagrams pass
Three inactive skill versions received fresh same-provider baselines, guided outputs and independent reviews. All 12 unchanged diagrams were imported and exported in Excalidraw 0.18.1, then visually inspected. Seven pass every local check; five retain connector or layout defects. The suite passes 146 of 156 checks.
Limits and evidence
One refund-flowchart fixture only. Requested Astra outputs pass with and without the skill, so these runs do not establish added skill value or general model superiority. Hosted imports, VS Code extensions and other diagram formats remain untested. Earlier root-authored repairs are separate.
Result file SHA-256: b8482e0ec0cbbd2059d37afb9ae1a14800e4f00e6cb5ec90fef87ec38d5083fe
DOCUMENTSThree inactive package copies
Sheets command construction
4 / 4 repaired drafts pass local checks
All three source versions and four retained Claude drafts were checked. One original draft passes every local payload check; four separate root-authored copies pass all 51. The isolated gws 0.22.5 CLI accepts 96 of 96 repaired example/draft dry runs. Helper controls expose comma splitting and input-mode defaults.
Limits and evidence
No live Sheet or real credentials were used. curl and gcloud were inert capture functions. The four original model drafts cover two versions; the third source was checked but its model comparison remains separate. Repairs do not replace model grades, and successful request previews do not prove live integration.
Result file SHA-256: b1b6dce9a94173d6877e54643b5d9b9923b1f8d7b95601bb97077e05ff791fd3
MARKETINGTwo inactive package copies
GHL message-listener repair
32 / 32 repaired browser scenarios
Both source versions and four retained Claude answers were checked. Inactive package and output repairs pass 136 listener cases plus two diagnostic cases, and 32 real browser scenarios. Tests cover duplicate IDs, exact message origins, absent/replaced iframes and preserving the data layer. Originals pass 60 of 138 local cases and 8 of 32 browser scenarios.
Limits and evidence
Synthetic forms and messages only: the real GHL payload contract, deployed GTM tags and GA4 collection/key events remain unverified. Browser pages were supplied locally; no external service was loaded. Corrected reviews retract unfair variable and length objections but still fail all four original answers. Root repairs do not inherit model grades.
Result file SHA-256: 520b9992df6103aca82421c272fb1ed71d3f257ba49c4866db510560891ac4dc
SEOTwo inactive package copies
Sitemap URL discovery repair
48 / 48 repaired-script cases
The original helper shared by two versions passes 20 of 48 cases. Its inactive repair passes all 48 cases and 383 checks, including 27 real loopback HTTP requests. Checks cover XML namespaces, escaped URLs, metadata exclusion, gzip, recursive indexes and output handling.
Limits and evidence
Synthetic files and a local test server only. The first repair still crashed on corrupt compressed data; that failure and the final correction are retained. Page facts, HTTPS/TLS, other supplemental-file scripts and WordPress upload remain unverified. Twenty-one companion files stay unchanged.
Result file SHA-256: 7903a361537a3ffedde37c68b764a2677794d9f53c993f20f63d9a1039db73e7
SEOSeven inactive package copies
Image metadata repair
45 image cases + 31 follow-up checks
All seven archived EXIF examples fail Python parsing. After restoring indentation, real writes reproduce wrong GPS hemispheres, omitted zero coordinates and joined keywords. The final helper passes 45 JPEG, PNG and WebP cases plus replacement, repeat-write and error checks.
Limits and evidence
The 45 cases include 27 real writes and 18 rejected invalid requests. Follow-up checks include 12 real update writes and three injected failure contracts. The full renaming CLI, image analysis, profile storage and MCP integration remain missing or untested. Originals stay unchanged.
Result file SHA-256: f123c7108f6e95ee67d57f4a85a664aa6d304c0c1081ed5fbb053e8985e9daf2
SEOSix inactive package copies
HTML audit parser repair
114 / 114 repaired-script checks
Searching all 6,209 ledger rows found two distinct parser files across six source versions. Original scripts pass 76 of 114 checks; repaired copies pass all 114. Checks cover URL resolution, readable text, hidden content, JSON-LD syntax and malformed links.
Limits and evidence
These are real local CLI checks on synthetic HTML. JavaScript and CSS rendering, live crawling and search eligibility are untested. Seven other seo-audit versions contain no parser. The six repaired packages preserve 77 unchanged companion files.
Result file SHA-256: 3b5eb75390ae2492b9e4c805329c62cf2a4ecd701085a9f4670377f392f06772
VIDEOBoth original examples pass
React animation execution
160 / 160 browser checks
Both unchanged Astra hero components ran in real browsers at phone and desktop sizes, in development Strict Mode and production. Eight runs verify animation, no replay, listener removal and cleanup. Four screenshots were inspected.
Limits and evidence
Only the supplied hero components ran. Other animation modes, missing implementations, reduced-motion behavior and a complete accessibility audit remain unverified. An initial checker syntax error and 16 incorrect expectations are retained; source components were unchanged.
Result file SHA-256: 6e7a2f5401d423dd4c2fe98e78b768beda2a5d220f4255206a381977c9a09d42
TESTSFresh model pairs reviewed
Nine corrected task contracts
18 answers + 152 bounded checks
An audit scanned 831 planned source versions and examined all 84 flagged contracts. Nine confirmed defects received new tasks with complete inputs. Astra generated 18 fresh guided and baseline answers; Claude reviewed every pair. All nine pairs passed their corrected offline cases.
Limits and evidence
The 152 checks cover fields, isolated hook inputs and command text. Live media, provider calls, publishing and other named integrations remain untested for all nine versions. Eight extractor controls are separate. Original sources, tasks and failures remain preserved.
Result file SHA-256: d297fcc2fdc99efc43608e9a1c688c666cfed50685d45cf51dffdc15e8f680a3
TESTSFive integration flags corrected
Offline pass scope corrected
21 prior pass labels screened
The audit examined all 21 pass labels at its snapshot. Five versions now explicitly retain missing references or untested operations, including the partial animation example. Original task text, model answers and model verdicts were unchanged.
Limits and evidence
This corrects coverage labels, not model grades or implementations. A successful bounded task still does not establish that missing SOPs, external frameworks or live workflows work. No skill was activated.
Result file SHA-256: 6fc9e91f81de67c2ece229e4e8251490f96abba9832eea898306571dc6cfea46
TESTSAll four answers still fail
Video-ad and campaign regrades
Two fresh independent reviews
Two hidden rubric requirements were removed before Astra regraded four unchanged Claude answers. The local campaign answers contain 409 and 382 words against a fewer-than-350 limit. Both video storyboards add B-roll beyond the requested two scenes.
Limits and evidence
These are offline answer comparisons. Unsupported claims and omitted qualifications still fail. Independent word measurements are retained separately from reviewer judgments. No media, account creation or publication was tested; all four unrelated negative requests passed.
Result file SHA-256: 494486b43f05e8f1858a86e2f1a624e9fd397cf2143a3f3af8cd18a0a520d8f2
VIDEOOriginals fail; repairs pass
Video concatenation
72 / 72 repair checks
Four original executions failed. A requested nine-second sequence became a 614.546-second container in the completed guided run. Four selected repaired executions preserve all 270 frames and pass timing, audio and clip-order checks.
Limits and evidence
Repairs change copies of generated scripts. The original skill remains failed and inactive. Two audio scenarios and this FFmpeg version were checked; arbitrary formats and perceptual audio quality were not.
Result file SHA-256: eecc73257cb23f3de4c9b6700a662eee94f3f3a6f45a051cf4faec46691d80d4
Repaired guided example, mixed-audio fixture. Nine seconds, 270 frames, with synthetic test tones. The original output failed.
VIDEOOriginal examples pass
ASCII animation
32 / 32 artifact checks
Both original Astra scripts rendered 15-second, 1920 by 1080 GIFs. FFmpeg and Pillow each decoded all 300 frames of both files. Claude passed both unchanged answers after correcting the rubric.
Limits and evidence
This checks the generative GIF example. Eight referenced files are missing; other modes and full package behavior remain unverified.
Result file SHA-256: d8a7b3fe4be9728156c82ee82c4c22dac08f6a1fe14d2b5a666f3dcd7fc244cd
VIDEOInactive example repair passes
Readable video on phones
115 / 115 browser checks
The guided example repair keeps required copy at least 12 pixels across five tested widths, from 320 to 1280 pixels. Each run observes two full 20-second loops and the next opening.
Limits and evidence
Playback used a test-owned host simulation. The baseline mobile issue, actual recording hook and exported video remain unverified. The original output is preserved.
Result file SHA-256: dd9b9962cb5e286de337805dd15874b5d39365f24b63930c8ff1e4ff401d17ed
SEOInactive example repair passes
Homepage keyboard access
38 focus instances checked
Both originals passed 92 homepage checks each. Extra keyboard checks found invisible or obscured focus in the guided output. A three-rule CSS repair passes the same 92 checks and all 38 focus instances.
Limits and evidence
One fictional homepage across four widths was checked. Live SEO performance and the complete workflow are unverified. Original content, links and markup are preserved.
Result file SHA-256: ad2d0559d1c8208b35641bbb99778e2a0d8517d8db2ac439e11fd7277afca21f
SEOSeven unknown-type rows remain
Structured data audit
1,071 saved answers scanned
The snapshot contains 121 answers with JSON-LD markers: 114 rows use recognized vocabulary; seven retain the unknown ElectricalContractor type. Six selected answers pass 76 task-fact checks.
Limits and evidence
This is a pinned-vocabulary and task-fact audit, including restricted interpretation of JSX data. Live URLs, complete JSON-LD semantics and search eligibility were not tested. Repeated rows are not additional skills.
Result file SHA-256: 32605d696ec9a50201f8fa86ed2dda2ef80e1702bc5acca93e347142be764199
SEOOriginal examples pass
Campaign tracking links
32 / 32 URL checks
Both retained Astra answers preserve the booking path, service selection and availability fragment. Four external placements use distinct tracking; the internal homepage link stays untagged.
Limits and evidence
These are parser checks of two answers. No URL was visited and no analytics or attribution was verified.
Selected additional execution and mechanical checks from the full collection. Counts measure checks or answers, never additional skill comparisons. Original failures are retained. No package is activated or approved for its full workflow. Evidence published from checkpoint 2026-09-12T10:43:42.721675+00:00.
Coverage by domain
6,146 captured source versions are in the test queue, plus 63 alias and missing-source records. 600 comparisons have valid, settled pass/fail reviews at this snapshot. Valid comparisons include failures; they do not mean the skill passed.
See every domain
Current source-version coverage
Domain
Captured versions
Comparisons reviewed
Valid comparisons
Reviews unfinished
AI workflows
812
23
23
789
Business & sales
292
22
22
270
Coding
1,286
35
35
1,251
Content & marketing
615
65
65
550
Design & media
669
132
132
537
Documents & data
163
15
15
148
Infrastructure
859
42
41
817
Needs sorting
33
5
5
28
Research & productivity
261
23
23
238
SEO & local search
1,156
239
239
917
Reviewed totals can include invalid fixtures or unresolved judgments. Those remain outside the valid-comparison count and retain their original evidence.
Every version is accounted for
6,209 version/reference records remain in the ledger. 1,892 CLI requests are recorded at this snapshot; copied earlier receipts do not count as new requests.
Queued and generated outputs are not passes. Invalid fixtures, skill failures, and missing integrations remain distinct. A successful text fixture does not prove a live account action or media workflow. No test result automatically activates a skill.
This page is a published snapshot. The laptop progress page shows newer activity between publications.
Same-model baseline comparisons
Every active version test is scheduled with and without its skill on the same model and task, followed by a blind review from the other model. Astra and Claude use signed-in CLI subscriptions through ORCA and Herdr. No metered LLM APIs are used.
Earlier guided answers and reviews remain preserved. Exact existing answers are reused where the task and batch context match; changed case groups receive fresh answers. Up to three independent cases share a batch. These results measure the supplied fixture, not general model superiority or real-world outcomes.
Bundled tests from 7 source versions ran on this laptop. The results include documented Windows test-driver adapters and pinned upstream dependencies.
69 / 73test files passing
636 / 641reported checks passing
5checks unresolved
0checks skipped
Original helpers and earlier failures are preserved. The reported count includes checks reached by these runs; a file that fails during setup may contain further checks. This does not verify live media providers or the complete video workflow.
Unresolved checks in 4 files
media-use/audio/scripts/lib/audio-meta.test.mjs: 2 unresolved. File-symlink fixture unavailable in this Windows environment.
media-use/scripts/lib/heygen-video-provider.test.mjs: 1 unresolved. File-symlink fixture unavailable in this Windows environment.
media-use/scripts/lib/lut-preset-provider.test.mjs: 1 unresolved. File-symlink fixture unavailable in this Windows environment.
media-use/scripts/lib/probe.test.mjs: 1 unresolved. Windows cannot create this quoted filename.
Runtime checkpoint: 2026-09-12T00:50:06.174835+00:00. No skills were activated by these tests.
Structure checked across the collection
6,146packages checked
1,704distinct scripts parsed
546packages with issues
526instruction repair candidates
Prepared repairs include 502 scalar-format fixes, 21 additions of missing metadata, and 3 explicit metadata conflict resolutions. Separately, JSX filename and reference patches cover 2 package versions. Instruction bodies and originals are preserved; metadata choices are recorded for review.
These checks do not establish runtime behavior, live integrations, or approval to activate a skill. 18 structural records still need resolution; the linked audit records their reasons.
Relative resource links were inspected across 6,146 of 6,146 captured packages. 231 packages have links needing resolution or version review; this alone does not prove a capability failure.
22 complete review copies restore 313 captured resources while preserving original entrypoints. 0 other candidates still have conflicting resource versions. These copies are not activated.
6 invalid fixtures received replacements with complete inputs and consistent requirements. Five incorrect word-limit findings were withdrawn after measuring visible Markdown text. Auditing all 25 completed Python outputs at that checkpoint also corrected 7 false syntax failures caused by explanatory prose or multiple code blocks.
56 recorded requests through ORCA/Herdr using Astra and Claude subscriptions, including baselines and retained failures. Seven current packages have text checks; Remotion has a verified local render. Gemini remains signed out. Small fixtures do not prove general model superiority or SEO results.
What changed after testing
SEO: the repaired versions preserve source metrics and distinguish missing evidence from real defects. Both models passed the selected skill checks.
Scripts: Astra passed both durations with the current version. Claude still strengthened a supplied claim, so Astra is the recommended route for this skill.
Scoring: equivalent notes arrays and hours objects were initially rejected by the grader. Corrected scores retain their original receipts; format errors and factual errors remain separate.
Many baselines also passed. These are small constraint checks, not a general model ranking. Earlier failed versions remain in the table and archive.
A real laptop render
720×1280 · 30 fps · 150 frames · five seconds · H.264 · silent. Opening, middle, and ending frames were inspected. This is a local render check, separate from the paired text tests.
Completed laptop pilot runs
Model requested
Case
Mode
Result
Seconds
gpt-6-astra
code-zero
baseline
Pass
8.32
gpt-6-astra
code-zero
skill
Pass
7.3
gpt-6-astra
code-negative
skill
Pass
6.62
gpt-6-astra
code-negative
baseline
Pass
8.48
sonnet
seo-links
baseline
exact_orphans
10.93
sonnet
seo-links
skill
Pass
17.48
sonnet
video-script
skill
Pass
30.02
sonnet
video-script
baseline
Pass
14.26
gpt-6-astra
seo-links
baseline
Pass
11.23
gpt-6-astra
seo-links
skill
Pass
13.82
gpt-6-astra
video-script
skill
spoken_word_budget
18.8
gpt-6-astra
video-script
baseline
Pass
18.5
sonnet
links-repaired
baseline
Pass
17.07
sonnet
links-repaired
skill
Pass
20.1
sonnet
script-repaired
skill
Pass
22.01
sonnet
script-repaired
baseline
Pass
16.57
sonnet
keywords-export
baseline
Pass
15.08
sonnet
keywords-export
skill
Pass
15.9
sonnet
youtube-timestamps
skill
Pass
8.7
sonnet
youtube-timestamps
baseline
Pass
10.83
gpt-6-astra
links-repaired
baseline
Pass
14.06
gpt-6-astra
links-repaired
skill
Pass
25.61
gpt-6-astra
script-repaired
skill
Pass
20.96
gpt-6-astra
script-repaired
baseline
Pass
19.34
gpt-6-astra
keywords-export
baseline
Pass
12.41
gpt-6-astra
keywords-export
skill
Pass
12.97
gpt-6-astra
youtube-timestamps
skill
Pass
13.44
gpt-6-astra
youtube-timestamps
baseline
Pass
16.17
sonnet
links-holdout
baseline
Pass
19.84
sonnet
links-holdout
skill
Pass
21.55
sonnet
script-holdout
skill
no_unsupplied_urls, no_claim_expansion
33.58
sonnet
script-holdout
baseline
valid_json
16.1
sonnet
seo-evidence
baseline
Pass
15.32
sonnet
seo-evidence
skill
Pass
9.5
sonnet
gbp-facts
skill
Pass
6.63
sonnet
gbp-facts
baseline
profile_shape
15.56
gpt-6-astra
links-holdout
baseline
Pass
12.33
gpt-6-astra
links-holdout
skill
Pass
14.99
gpt-6-astra
script-holdout
skill
Pass
28.22
gpt-6-astra
script-holdout
baseline
Pass
24.09
gpt-6-astra
seo-evidence
baseline
Pass
11.82
gpt-6-astra
seo-evidence
skill
Pass
14.3
gpt-6-astra
schema-facts
skill
Pass
13.95
gpt-6-astra
schema-facts
baseline
Pass
14.81
sonnet
schema-facts
baseline
valid_json
24.03
sonnet
schema-facts
skill
Pass
10.44
gpt-6-astra
gbp-facts
baseline
Pass
8.97
gpt-6-astra
gbp-facts
skill
Pass
11.65
sonnet
script-repaired
baseline
no_unsupplied_urls
15.75
sonnet
script-repaired
skill
Pass
14.84
sonnet
script-holdout
skill
no_claim_expansion
29.16
sonnet
script-holdout
baseline
no_claim_expansion, fictional_label
17.19
gpt-6-astra
script-repaired
baseline
Pass
18.15
gpt-6-astra
script-repaired
skill
Pass
21.67
gpt-6-astra
script-holdout
skill
Pass
27.84
gpt-6-astra
script-holdout
baseline
Pass
23.93
Text checks cover declared constraints, not audience performance or live SEO outcomes. Astra did not report a resolved model ID; Claude usage reported Sonnet 4.6 and auxiliary Haiku 4.5. Eight exact package versions received a separate review for the stated use and models. Gemini is not tested.