Contractor+ / Jobber source inventory — 2026-09-10

Acquisition complete: 5 of 5 original source items available locally. These are the full email attachment and all four linked Drive files. DEGRADED: 0-of-3 fresh video perception checks — no new private-pixel vision call or fresh listening pass in this lane. Metadata, archive membership, hashes, and client-supplied text were inspected; visual-quality judgments below are expressly attributed to that text.

Identity: a client-provided creative-production handoff comprising a reconstructed session history, the rejected Contractor+ Rands video, the real Rands testimonial, the Jobber reference screen recording, and a mixed archive of prior creative work. Medium/format: digital text, videos, and raster assets; no folded print artifact (fold count not applicable). Intent: inspect sources and diagnose/rebuild direction; not faithful duplication of competitor branding. Canvas dimensions come from fresh ffprobe/PNG headers below, not filenames.

Requested model/effort: GPT-6 Astra / ultra. Observed agent runtime effort: unavailable in this lane; configuration intent is not proof of runtime. No generation, external pixel vision, outbound email, source-code execution, credential access, or shared configuration edit was performed.

Provenance and local originals

Email: Justin Smith ceo@contractorplus.app to Jeremy, subject jobber style video ads, 2026-08-19; Gmail message/thread ID 1a01aefaf59a952a. Body is preserved separately in email-brief.txt at the task root. Session history was recovered by reading the authenticated Gmail message in raw format and decoding its MIME attachment locally; full raw email was not retained. This avoids the connector's approximately 30 KB extracted-text preview limit. The partial task-root history is superseded for reading by the complete original below, but retained as received.

Drive metadata was readable via the connected app; a streamed fetch returned a sediment:// reference. Local downloads used the supplied Drive file IDs through Google's normal publicly accessible download endpoint. The ZIP used Google's standard large-file confirmation form. No credentials were needed. Disk preflight showed 121 GB free.

All four downloaded Drive file sizes match source metadata. The Gmail attachment matches its declared 32,552 bytes. No original was silently truncated.

Fresh media metadata

The two ZIP videos are different assembled versions from the individually linked rejected render: each runs roughly half a second longer, and their audio sample rate differs. The ZIP versions were probed directly from archive bytes without creating large duplicate media files. No rotation side data appeared in the three original video probes. Audio streams exist, but codec/channel metadata alone cannot establish the presence of music, sound design, intelligibility, or quality.

Detailed receipts: source/media-summary.json, the three *.ffprobe.json files, and source/archive-media-summary.json.

Existing PromptLab source match

The downloaded Jobber MOV and existing [internal reference omitted] have different container hashes/sizes. Fresh ffmpeg stream-copy hashing established identical compressed video and audio payloads:

• Video SHA-256: 5be4ae0a8dd6b786b8a5097a4c8c7876a229cd111a7f1206aa0448b514817931.
• Audio SHA-256: 5d76beb68d16c418b9304b58a443031b265a4601bb6eabaf402fce2e582d74ad.

This establishes that the earlier PromptLab project's source media is the same reference content. It does not certify every prior analysis statement. Its old frame sidecars include incomplete OCR and interpretive descriptions; do not treat them as independently verified measurements. The tools-preflight lane was notified of the match and all three local source paths.

Archive: physically present versus promised

Outer ZIP: 79 entries, 227,642,902 uncompressed bytes. Excluding one directory and 39 macOS resource-fork sidecars, there are 39 real files: 31 PNG statics, two assembled MP4s, three Markdown documents, one JSON transcript, one text transcript, and one nested ZIP. No encrypted members or traversal/absolute paths were found.

Nested contractorplus-core-package.zip: 42 entries, 13,698,175 uncompressed bytes; 37 files plus five directories: ten Markdown documents (README plus nine strategy documents), two text files, one JSON transcript, six Python tools, and 18 PNGs. All 76 real outer-plus-inner members were read successfully with ZIP CRC checks and SHA-256 recorded in source/archive-hashes.json. Scripts were not executed.

Inventory receipts: source/zip-inventory.json, source/nested-core-inventory.json, and source/archive-hashes.json. Safe text copies are under source/archive-text/ (nested documents under core/export/). The original ZIP remains intact; bulky alternate renders and statics remain in it.

Present brand assets — do not request these again: 13 PNGs have been extracted under source/core-package/export/03_brand-assets/ for access. Their source identity is supplied by the client's package; fresh pixel classification is still open before creative use.

• Four primary-logo versions: 3085×599.
• Three wordmark versions: 1616×374.
• Three submark versions: 790×1075.
• Branded payment-reader plate: 1200×1068.
• App Store and Google Play badges: each 1080×322.

Nested research images are present in the original ZIP: rands_testimonial_shot_scan.png, jobber_ad_30_frames.png, jobber_end_card.png, F1_final_storyboard.png, and jobber_money_shot.png. They remain archive evidence, not fresh perception evidence.

Present planning/source text: nine strategy documents, including the full Jobber teardown, ten trade scripts, five real-customer film scripts, creative style system, and ad/copy libraries; supplied Jobber word timings and Scott's segment transcript; six Python production tools (layout_engine.py, animate.py, build_v3.py, export_ratios.py, fast_xfade.py, finish_ad.py).

Not physically included despite README/FILE_INDEX claims: the 04_video/ source hierarchy is absent from the nested core package. The promised 12 separate Veo plates, 18 beat outputs plus alternate beat, raw/final voiceover files, vo3_words.json, and animation-test renders are absent. The only video media in the outer ZIP are the two assembled F1 exports. The claimed full set of 107 statics is also absent; 31 are physically present, and contractorplus-static-ads.zip is not inside the supplied archive. Original brandbook.pdf is absent, although derived assets and palette/type notes are present. Real product screen recordings and customer-release documents are absent from these supplied items.

Evidence-backed content findings

1. The complete session history is explicitly a reconstruction, not a verbatim Claude transcript. It says no session log was available on disk. Its production and model claims remain author-reported historical claims, even where labeled measured or verified.
2. The rejected render matches the history's rough output spec (approximately 28.5 seconds, 1080×1920, 33 MB). Its physical delivery alone does not establish quality.
3. Client correction pattern: history records the wrong initial brand accent, stills-versus-motion misreading, lack of animated type, Scott's missing real footage, logo contrast failure, product placeholders, and dissatisfaction with the final sound/edit. These are historical source statements, not newly observed pixel/audio findings.
4. Scott's supplied transcript offers usable authentic source lines, subject to listening verification: 26.58–28.80 s, “it didn't do even a third of what Contractor Plus could do”; 73.74–82.24 s includes “saves us time, saves us money, and it helps us look good on paper.” Its spelling RANS reflects the supplied transcription; the source filename/package uses Rands. The supplied transcript describes over 30 years of combined experience (10.22–15.20 s), not nineteen years.
5. Stale source-script defect: archived contractor-plus-5-customer-films.md still writes “Scott's been in HVAC nineteen years,” although the later history says this was corrected. It also scripts the attributed complaint “NOBODY TOLD ME IT WAS LEAKING” and a five-star review “They sent me photos before I even called.” Those are not in Scott's supplied transcript. Do not reuse as authenticated customer quotes. Script's ✅ verified growth label is not a substantiation receipt; no account-data export is attached.
6. Reference timing is supplied, not freshly remeasured in this lane: archived Jobber transcript times the million-dollar wording at 6.36–6.80 s and the “one side” wording at 11.96–12.16 s, followed by the new-way narration at 13.76 s. It repeatedly transcribes the brand as Jabber; preserve original evidence while correcting branded text only in a separate reviewed working transcript.
7. Do not blindly carry historical model/pricing claims forward. The session's model credit prices, performance ratios, vendor comparisons, and offer amounts are August-session reports. The actual reference source is approximately 60 fps; the failed output is 24 fps, so the history's “24fps to match the reference” is not supported by the supplied source's encoding. Source recording frame rate also does not prove native underlying-ad cadence.
8. Claim and asset gaps remain distinct: source text flags financing/seat-price discrepancies, customer releases, and genuine product-screen footage. Brand logos and the real testimonial are available now; separate source plates/VO/editable timed elements are missing from the supplied package despite its inventory text.

Perception boundary and exact resume

Private pixels were not sent to any external service. The current image_context.py has both an optional external Astra support trial and a local Qwen route that sets think:false; that default is incompatible with this task's local-only direction and highest-supported-reasoning preference. No shared switch was edited, no lower-effort call was silently made, and no fresh visual acceptance is claimed. The documented ensure_context(..., vision_fn=...) injection seam suppresses external support; a compliant local thinking control still needs explicit verification before a new call.

Resume source checks locally: python3 -m json.tool project/source/media-summary.json. For creative review, use the tools-preflight lane's extraction output and arrange compliant local classification/listening before claiming visual or sound-quality validation. This inventory's acquisition and archive findings are complete; pixel/audio judgment remains visibly open.