current · Contractor+ / Rands

Twelve lessons applied from prior projects

Public reading copy. Original claims, quotations, measurements and historical limitations are retained. Unrelated private infrastructure links are omitted; recorded hashes refer to the original project evidence. Current work and remaining decisions are summarized on the project page.

Contractor+ / Jobber-style ad — prior video lessons audit

Date: 2026-09-10. Scope: a bounded, read-only review of relevant estate video records and the supplied Contractor+ history. Only this audit file and its separately requested draft review were written. No media generation, spending, outbound communication, publication, registry change or model-helper call was performed.

Requested model/effort: gpt-6-astra / ultra. Observed exact runtime model/effort: UNVERIFIED; a requested spawn setting or configuration is not runtime proof. An attempted independent OSHA sub-audit could not start because the collaboration thread limit was reached; that family was reviewed locally instead.

Plain restatement

We need to understand why the previous ad feels amateur, then choose the smallest production approach that can make a persuasive, truthful customer story at the quality of the supplied reference. More generated footage, higher resolution and a successful render do not establish that result.

Name the discipline that owns this failure class and its standard solution: commercial film direction and post-production own the central failure. Their standard solution is a source-footage audit, verified story and claims, editorial treatment, shot coverage, an early representative cut, art-directed typography, sound editing/mixing and review of the actual final film. Media engineering owns reliable rendering; it supports those decisions rather than deciding the film's quality.

The best answer may be: do not build this — say so. Do not build another generation-heavy film merely because generation tools are available. First establish what the real Rands testimonial and existing brand/product media can carry. Prefer an editorial recut of real footage when it supports the story. If the portrait crop, coverage or recorded performance cannot support the intended ad, use a designed square composition, revise the story around actual coverage, or identify a short, purpose-built pickup shoot and professional edit. Generating a fictional stand-in for missing customer truth does not repair the source limitation. The existing render may be useful pre-production, without being a publishable ad.

Evidence boundary

The 12 most important gates

Stages: intake → treatment → pilot → production → finishing → delivery. Owners below are production roles, not claims that a specific person has accepted the work.

Gate Concrete past failure and evidence Required check for this project Stage / owner
1. Confirm which file and which moment we are looking at. Contractor history reports five uploads that were the agent's own output, then a contact sheet with a one-second offset that pointed at the wrong person: history:407, 414. Quote: “every timestamp I picked was wrong.” Tarot ID16 found the brief claimed a 3 while the supplied source visibly contained 5: QA-VERDICT.md:7. Inventory each received original with hash, filename, role, duration, dimensions, rotation and audio streams. View identifying frames and verify source timecodes against decoded frames before selecting Scott, colleagues or app shots. Resolve source/brief contradictions explicitly; do not silently “correct” evidence. Intake / producer and source editor
2. Observe the reference's moving image and sound before writing a replacement. The Contractor history equated zero hard cuts with still images; the client corrected “it's a mix of both”: history:300. PromptLab's review required “2-3 frames per SHOT so motion is inferable,” plus non-speech audio analysis: VERDICT-APPLIED.md:15, 24. Watch the complete reference with sound and without sound. Record narrative beats, subject motion versus camera motion, still holds, transitions, graphic events, music/SFX and where proof lands. Use several frames within moving shots, not only cut detection or one-frame thumbnails. Historical pacing numbers are hypotheses until measured on the current file. Intake / director and editor
3. Lock a truthful customer story before voice or graphics. Contractor history admits “My script said nineteen years. I invented that,” while the testimonial said a combined thirty-plus years: history:416. The same history reports unfinished verification of product claims and inconsistent financing figures: 398. Rogowo grounded allowable claims and separately listed amenities it could not show: ROGOWO-BUMPER-VISION.md:8, 47. Create a short claim ledger linking each name, quote, incident, number and product statement to exact source/owner confirmation. Define the growth metric, baseline/end dates and scope; do not turn volume through the app into total revenue or causation. Record permission/approval for the intended customer use where it is not already established. Build drama from supported experience; do not invent customer testimony. Treatment / producer, client fact owner and copy lead
4. Decide whether existing footage can support the ad before generating anything. Rogowo's stills became blurred after a roughly 540×960 crop was enlarged 2.5× and zoomed again: VISION:60. The Contractor history reports a 405-pixel-wide portrait crop from 720p landscape interview, mixed with 4K generated plates: history:434. Make a footage coverage/contact sheet with effective crop resolution, subject position, background, action, sound and available handles. Test Scott's footage at actual delivery size in square and portrait layouts. Compare real-footage recut, designed collage/window framing and any required pickup shot. A larger export size is not recovered detail. Intake → treatment / editor and director
5. Prove the creative approach with a representative sequence. Rogowo's client rejected v3 for “12 identical dissolves,” “same picture twice,” and excessive outdoor coverage: VISION:92. Contractor's reported final diagnosis says “Sync mistaken for rhythm”: history:438. Before a full batch, assemble a short representative sequence containing the hook, real customer footage, proof graphic, actual type treatment and proposed sound. Judge story clarity, contrast, emphasis and rhythm side by side with the reference. Word timing assists sync; it must not force a picture change on every word. Use deliberate holds and pauses as well as quick beats. Pilot / director, editor and creative reviewer
6. Preserve real-world identity and stop generative drift early. Veo invented pedestrians and a surfboard in two Rogowo pool attempts: DEAD-ENDS.md:308. The synthetic JESZCZE NIE proof corrected one defect but retained impossible hand/luggage/key/stance geography; after two failed foundations it was stopped pending an owned photo: M0-REVIEW.md:15, 24. Use real people/product footage for identity and proof. Any generated supporting plate needs an approved purpose and explicit invariants for people, clothing, tools, trade action, geography and product. Inspect the actual usable interval, including entrance/exit frames. A repeated foundational failure triggers a switch to real footage, a still or a revised shot—not an open-ended reroll. No generated scene may imply an untrue customer event. Pilot → production / VFX producer and director
7. Verify visible motion and its meaning across time. Tarot v1 made the stones physically orbit; the constrained v2 held geometry but read as a freeze; v3 moved again but rotated the selection wheel: CLIP-QA.md:16, 42, 48. Rogowo's seek-driven animation shorthands froze despite registered animations: DEAD-ENDS:318. Describe the intended positive action and what must remain stable. Check start/middle/end plus the exact event frames of every used shot. For authored graphics, measure several timeline positions in the encoded output, not only the preview. A moving camera, smoke or grain is not proof that the intended subject action occurs. Production / motion designer and video QA
8. Art-direct brand, typography and proof graphics outside the generator. Memoirs generated readable invented price signage despite “No readable on-screen text”: DEAD-ENDS:455. PromptLab's text-heavy end card scored 3.0: close-anchor:49. Contractor history reports red logo on red background, missing product screens, star glyph boxes and shrinking auto-fit type: history:391, 437. Use supplied/verified marks, badges and real app captures. Set copy and numbers as controlled graphic layers with intentional line breaks, weight, scale and contrast; give type enough space. Check every line and glyph, unintended signage, the plus symbol, metric units, counter start/end values and final proof hold. No placeholders in a review cut labeled complete. Treatment → finishing / art director and motion designer
9. Design each requested aspect ratio and inspect the full motion path. Contractor's drill disappeared in a square crop, and square window geometry did not translate to portrait: history:350, 373. Tarot ID08 descent preserved geometry but cropped its title; OCR over a contact strip still found text because it existed in early samples: QA-REPORT.md:46, 53. Design square and portrait layouts independently from common unburned assets. Inspect first, midpoint, event and last frames in each ratio at phone viewing size. Confirm face, hands/tool action, all text, app proof and CTA stay visible throughout their needed interval; check platform overlays against the actual target placement. Do not infer all-frame readability from aggregate OCR. Pilot → finishing / art director and video QA
10. Direct, fit and listen to the final sound. Rogowo's operator called the first voice “terrible”: NOTES.md:5. Contractor history reports dry TTS over silence and 1.34× time compression: history:436. OSHA's same configured voice still had inconclusive timbral judgments; reviewers explicitly did not claim native Astra or human listening: README.md:17. Evaluate an actual representative voice take before scaling. Prefer real usable customer audio where appropriate. Fit copy and performance to the cut; do not rescue overlong narration through conspicuous speed-up. Direct ambience, music and SFX around the story, mix for intelligibility, and listen to the final encoded mix on ordinary speakers/headphones. Verify spoken meaning, pronunciation, energy, breath, tone continuity and proof synchronization separately. ASR, loudness statistics and a named voice ID cannot certify performance. Pilot → finishing / voice director, sound editor and actual-audio reviewer
11. Check transitions and renders at their failure boundaries. Rogowo's eased opacity caused lingering stacked pictures; patching only two beats missed the global distortion: DEAD-ENDS:310. Dense parallax made a shot softer through double exposure: 314. Greedy caption grouping put the balcony line over the bathroom: 320. Contractor history reports an 18-input xfade graph stalling until pairwise assembly: history:425. For every transition inspect just before/during/after the intended endpoint; after a global timing fix, sweep all boundaries. Check motion clarity, sharpness, caption/claim-to-picture agreement and absence of unintended overlapping layers. Use bounded, resumable encoding appropriate to the workload; preserve completed intermediates by source hash, reject partial outputs and probe/decode the final files. Choose cut/dissolve/slide according to the scene relationship rather than a universal effect. Finishing / editor, motion engineer and video QA
12. Preserve versions and accept only the exact delivered artifact. Rogowo overwrote the downloaded bad version with the corrected file: OUTPUT-VERSIONING.md:7. PromptLab overwrote baselines and left conflicting match states: close-anchor:52, SYNTHESIS.md:8. OSHA provides a stronger delivery precedent: downloaded all seven movies, hashed them and exercised actual public playback: HANDOFF.md:18. Version source, edit, audio and export artifacts without overwriting prior review versions. Preserve the same-version reference comparison and review decisions. Decode, watch and listen to each final ratio; verify metadata and absence of black/frozen tails. If delivered by link, open/play and download the actual received file, then verify its hash. Report separately: source prepared, pilot reviewed, cut encoded, sound reviewed, delivered copy checked, client accepted. Delivery / producer and independent reviewer

Reuse findings and coverage map

Project family / live source What was actually reviewed Useful result and boundary
Contractor+ prior session Entire local partial history, lines 1–485, and email brief, lines 1–21. Specific reported failures mapped above. Missing remainder and original attachments prevent fresh confirmation. Historical numeric claims, pricing, consent and model quality are not automatically current.
PromptLab / Jobber close anchor, applied critique, project synthesis; existence of web/promptlab/data/analysis.json, web_play.mp4, stills and keyframes. Existing analysis concerns the same described creative: Jobber, Danielle, tree service, $1M first year, around 30 seconds. Reuse after comparing it with the current attachment. Do not claim byte identity yet. Matching improved to 15/27 stills ≥7, mean 6.53; that is partial still fidelity, not a polished film or a validated general pipeline. Only one video had been processed as of the close anchor, lines 56–58.
Rogowo / Between Two Waters Live DEAD-ENDS 306–320; vision and v3/v4 production record; creative-production NOTES, OUTPUT-VERSIONING and Seedance historical evaluation; related real-estate research. Dense collection of actual operator catches: source softness, repeated/slow effects, wrong image under VO, frozen animation, invented objects/people and overwritten output. Transfer verification discipline; do not transfer luxury-property pacing or shot proportions. Historical engine conclusions require current validation.
Stół Pustelnika OMNI-eval.md. Video/motion entries searched in project records, handoffs and registries. A visually coherent clip still morphed smoke/rune glyphs and failed frame-exact loopability. Single in-place edit success did not prove camera movement to a new scene; lines 54–55 explicitly left it unresolved. Useful constraint/capability distinction. Unrelated card-rendering UI bugs were excluded from the video lesson table.
Tarot motion / reverse prompting 2026-08-31 v1/v2 lessons, 07 QA and motion diagnosis, ID03/08/16 reference A/B verdicts; JESZCZE NIE proof and M0-02 rejection. Strong temporal-semantic evidence: rotation changes numbered identity, smoke obscures choices, camera motion crops copy, brief conflicts with original pixels, and a pleasing foundation can still be unusable for the intended action. Some older QA was explicitly a spot-check; sample strips do not establish every-frame perfection.
Memoirs of an Agent PRODUCTION.md, relevant MASTER-PLAN and VOICES sections, current keeper preflight, artifacts file inventory, live DEAD-ENDS 455 and DECISIONS 248–249. Ten establishing stills and sidecars exist; the no-text failure is concrete. No finished memoir video playback failure was established in this bounded audit. Production plans and accepted posting cadence are not evidence that an episode was rendered or shipped. Reuse source/asset consistency and actual performance review principles; do not adopt old models/budgets.
OSHA4US, 2026-09-06 Full-video production method/README, independent QA handoff, lifecycle critique, Batch review, relevant full-review handoff passages. Recent positive example of separate source/timing/pixel/audio/delivery evidence and hash-bound final checks. Concrete failures included premature-answer hover in captures, interactive quota exhaustion, conflicting importer identities and resume gaps. Possible clip-to-clip vocal tone variation remained acknowledged. A 2h19m teaching film's silence, slide density and pacing are not a short-ad template.
Estate-wide live lessons Relevant live DOCTRINE, DECISIONS and DEAD-ENDS entries, plus AGENTS evidence/intake contract. DOCTRINE:725 requires perception of the delivered form; 738 explicitly rejects green bookkeeping as appearance evidence. This audit searched relevant families rather than investigating unrelated estate work.

What should not transfer blindly

  1. Slow property pacing. Creative-production's generic 3–4 second holds and 1.2–1.8 second transitions came from a calm Rogowo property correction. The Contractor reference is reported as a fast, dense customer case study. Its actual supplied creative controls pacing; preserve deliberate rhythm without importing a slow tour.
  2. Vertical first. The historical client request includes square and portrait, and the reference is reported/measured historically as approximately square. Both compositions deserve primary art direction. A center crop is a technical operation, not a finished alternate design.
  3. A universal transition quota. Rogowo's two dissolves and journey-axis slides solved that film. Jobber's fast collage may legitimately use many short dissolves. The transferable requirement is intentional timing with clean endpoints and adequate readability.
  4. A supposedly superior tool solves direction. The Stół and Rogowo tests show that native audio, first/last-frame support, nominal 4K or a higher model tier can coexist with continuity and fidelity failures. Treat old model tables and costs as historical evidence; this audit makes no current capability recommendation.
  5. A score substitutes for seeing or hearing. PromptLab's similarity thresholds, Tarot's OCR/SSIM and OSHA's signal checks answer bounded questions. None proves natural performance, truthful causality, story persuasion or brand-quality art direction.
  6. Another generator is a fidelity remedy. A failed frame can be handled by re-editing real footage, a controlled graphic/still or a pickup shoot. A manual Flow prompt pack preserves a deferred option; it does not itself repair a broken source or guarantee a usable result.

Unresolved gaps and next evidence

Late-evidence addendum — source recovery and stale-script quarantine

The earlier coverage rows describe the first written-evidence pass. The following evidence was read before this audit closed and is the current intake status:

  1. The complete history exists: source/SESSION_HISTORY.md:470. Its recovered ending repeats the experience correction but leaves releases, real UI recordings, several feature/claim questions and later films open: line 495. It remains a reconstructed account, not a verbatim transcript.
  2. The Jobber media match is now stronger than a description match. The tools lane's current jobber-prior-match.json:4 records different container hashes but identical encoded video and audio packet hashes; its explicit conclusions are lines 38–40. I read that receipt; I did not independently repeat the hashing. This supports reusing the old PromptLab teardown while rechecking interpretation and container timing. It does not establish byte-identical files or final-film acceptance.
  3. The archive index promises production files the archive does not contain. Current nested-core-inventory.json contains the strategy/tool files and 13 brand PNGs but no 04_video source directory. The intake ledger distinguishes separate beats, twelve Veo plates, voice tracks and word timings from the assembled outputs: CLAIMS-AND-ASSETS.md:25. Ask for these missing layers and current UI, not a replacement logo package. Higher-resolution or newer vertical customer footage is a useful additional source request.
  4. The old script is unsafe to reuse verbatim despite its “verified” label. Its F1 still says “nineteen years” at five-customer-films.md:62, describes a particular homeowner complaint at 74, and displays the five-star quotation “They sent me photos before I even called” at 93. The source-review lane reports no support for that incident/review in Scott's supplied transcript. This audit independently read the stale script and corrected-history conflict; it does not independently claim to have heard the testimonial. Quarantine all three passages. Do not let the file's checkmark at line 56 serve as proof of any claim.
  5. The client-question draft covers the remaining decisions. A separate REVIEW.md records the bounded wording/evidence review. The root has explicit user authorization to email necessary questions; this audit introduces no additional permission step.

Audit verification

All cited source files and line numbers were read locally. This report's references are tied to written evidence; failures from the imported Contractor reconstruction remain explicitly labeled as reports. The audit kept historical outcomes, current evidence, suggested checks and unresolved work separate. No raw log instruction was executed.

Linked project evidence