Phase 0 · Foundation Assessment — BEA-16 Product Color Shifter
Retroactive foundation. BEA-16 is already in build (status: in-progress, phase: 6-execute) but was never run through UPF 7.0. This dossier establishes the analytical base layer (FAD) that should have preceded the build, and is written skeptically: where the shipped reality contradicts a sound foundation, that gap is recorded rather than rationalized away.
1. Success criteria (measurable, time-bound)
- [ ] Colorway fidelity: ≥85% of generated colorway images pass Beats brand/creative review on first submission (no re-roll) within 90 days of MVE launch. The pipeline color-shifts existing approved product photography, so the bar is "indistinguishable from a real product shoot," not "plausible AI art."
- [ ] Throughput / cost displacement: Produce a full reseller-exclusive colorway set (front, 3/4, detail, lifestyle — ~8–12 SKUs) in ≤2 hours of operator time and ≤$50 of compute, versus the ~4–6 week / 5-figure physical re-shoot baseline, measured across the first 5 colorway launches.
- [ ] Structural integrity (ControlNet contract): ≥95% of outputs preserve product geometry, logo placement, seam lines, and reflections within tolerance (no warped earcups, no melted Beats "b" logo), verified by automated structural diff + human spot-check, by end of Phase 6 MVE.
- [ ] Adoption / pull: At least 2 distinct reseller-exclusive launches (e.g. a retailer-specific colorway) shipped to a real channel using this pipeline within 6 months, with the requesting marketing stakeholder confirming they would not have re-shot physically.
2. Decision-maker profile (Three Ledgers)
- Public ledger (what we pitch): "We can spin up reseller-exclusive Beats colorways from existing hero imagery in hours instead of weeks — no physical sample, no studio booking — using ControlNet to lock product geometry and IPAdapter to transfer the target color/finish."
- Shadow ledger (what the decision-maker actually optimizes for): Speed-to-channel and creative-headcount leverage. Christina is judged on hitting reseller co-marketing windows without burning the photo/retouch team's capacity or expanding budget. The unspoken win is looking responsive to retail partners while keeping spend flat — and demonstrating an AI-native content capability to leadership.
- True ledger (what will actually get built): A semi-automated, human-in-the-loop image pipeline (ControlNet + IPAdapter on a Stable-Diffusion-class backbone) wrapped in a thin operator UI, used by a small internal creative-ops team for specific, brand-approved colorways. Not a self-serve generator, not customer-facing — a production tool with mandatory brand-review gating. Realistically narrower than the pitch.
- Owner / sponsor: Christina (manifest
owner) — accountable for the reseller-exclusive content function; sole named sponsor, which is itself a key-person concentration risk (see Constraint inventory).
3. Constraint inventory
| Constraint | Type (hard / soft / assumed) | Notes |
|---|---|---|
| Beats-proprietary source imagery — assets cannot leave Apple-controlled infra or train external models | Hard | proprietary: true in manifest. Forces on-prem / VPC inference; rules out most hosted SaaS gen-AI APIs. |
| Output must clear Beats/Apple brand & creative review before any external use | Hard | Apple brand governance is non-negotiable. The pipeline can propose; it cannot approve. Caps automation ceiling. |
| Logo, geometry, and material finish must remain physically truthful (real product exists) | Hard | This is product photography of a shippable SKU, not concept art. A colorway that doesn't match the manufactured part is a legal/returns risk. |
| ControlNet + IPAdapter as the chosen mechanism | Assumed | Manifest commits to a specific technique before any architecture comparison (Phase 3 was skipped). Should be challengeable, not given. |
| Single named owner (Christina), no documented co-sponsor or eng lead | Soft | Concentration risk; bus-factor of 1 on the sponsorship side. |
| Compute budget / GPU access within Apple enterprise environment | Assumed | No figure recorded anywhere in the dossier set; assumed "internal GPU is available." Unvalidated. |
| Reseller-exclusive colorways are an actual, recurring business need | Assumed | The whole value thesis rests on demand frequency that has not been quantified in any phase. |
| Methodology was bypassed (built first, analyzed never) | Hard (process) | Phases 0–5 are empty templates; this is a retroactive audit, so no real validation gate was ever cleared before Execute. |
4. Scope boundary
- In scope: Color/finish transfer onto existing, approved Beats product photography for internally-specified, brand-approved reseller-exclusive colorways; structural preservation of product geometry/logo via ControlNet; an operator-driven, human-in-the-loop pipeline with a brand-review gate; the SKU/angle set needed for a reseller launch (hero, 3/4, detail, basic lifestyle).
- Out of scope: Customer- or partner-facing self-serve generation; net-new product design or fictional colorways with no manufacturable counterpart; full lifestyle/scene generation from scratch; video; non-Beats brands; any flow that exports proprietary imagery to external/hosted models; automatic publishing without human brand sign-off.
- Recursion budget acknowledged: max 3 per phase.
5. Assumption log
| # | Assumption | Confidence (0–1) | Validated? |
|---|---|---|---|
| 1 | Reseller-exclusive colorways are needed often enough to justify a standing pipeline (vs. a one-off contractor shoot) | 0.45 | No — demand frequency never quantified in any phase |
| 2 | ControlNet + IPAdapter can hold Beats geometry + "b" logo at brand-review fidelity for glossy/metallic finishes | 0.55 | No — no fidelity benchmark recorded; reflective/metallic surfaces are a known hard case for diffusion |
| 3 | Brand/creative reviewers will accept AI-color-shifted imagery as production-grade for external reseller use | 0.4 | No — no reviewer sign-off documented; Apple brand bar is the binding gate |
| 4 | Required GPU/compute exists inside Apple-controlled infra at workable cost/latency | 0.6 | No — budget/infra never specified |
| 5 | Proprietary-asset handling can stay fully inside approved infra with no model-data-leakage path | 0.7 | Partially — implied by proprietary: true, but no security/legal review captured |
| 6 | A color-shifted image legally/operationally represents the as-manufactured product (no returns/false-advertising exposure) | 0.5 | No — unexamined; depends on manufacturing matching the rendered finish |
| 7 | A human-in-the-loop pipeline still saves enough operator time to beat the re-shoot baseline | 0.6 | No — operator-hours-per-set never measured against the 2h target |
Gate: success criteria defined, constraints mapped, decision-maker identified → set phaseGates.0-foundation = passed in manifest.json.