Phase 0 · Foundation Assessment — BEA-15 2D / Motion Assets to Spec
The permanent base layer (FAD) for all downstream work. Written before analysis; reconciled against Phases 1–5 so it remains the stable foundation those phases build on. One-liner under assessment: "One approved creative → every platform spec (social, Amazon, print, motion) with product-aware cropping." Owner: Sam. Category: Production Automation. Build type: backend-automation.
1. Success criteria (measurable, time-bound)
The central finding carried through every later phase is that this project's value lives almost entirely in one empirical question — can Beats' product-framing judgment be reduced to a brand-ratified, auditable preservation envelope? — and that the render/conform pipeline is largely commodity. Foundation success is therefore defined around retiring that question cheaply before any build, plus the operational floors that prevent "innovation theater."
- [ ] Pre-build validation gate cleared (binding, before any pipeline build). Within 6 weeks of project start, replay ≥150 historical Beats approved/rejected framing decisions through a candidate geometric preservation scorer and measure agreement; gate passes only at ≥80% agreement AND a brand owner countersigns a ratified threshold document. Below 80% or no ratification → project is reclassified commodity-buy and rescoped/killed, not built.
- [ ] CV-recall benchmark passed on a real Beats catalog (parallel pre-build gate). Within the same 6 weeks, off-the-shelf detection (MediaPipe / YOLO / Apple Vision) achieves ≥85% recall on a labeled set of ≥200 real Beats SKU/logo assets (including glossy-black-on-dark and lifestyle-occlusion cases). Below bar → fall back to fast manual-ROI draw, do not claim "confirm in seconds."
- [ ] Production-usage floor met post-launch (anti-theater kill switch). By 3 months after first production release, ≥30 approved masters/week flow through the system end-to-end measured from pipeline logs, not demos, and ≥60% of crop-required jobs (not just scale/pad jobs) are auto-cleared or resolved in-system. Below either floor → declare theater and sunset.
- [ ] Brand-incident-free auto-clear in production. Over the first 90 days of production auto-clear, zero brand-flagged "product cropped / off-brand" incidents on auto-cleared (non-human-reviewed) outputs; first such incident triggers immediate threshold review.
- [ ] Producer time genuinely removed, not relocated. By month 3, median producer time per approved master to obtain all platform variants drops from the ~25–40 min/variant × 6–10 variants baseline to ≤30% of that baseline, measured by role-segmented time tracking — the benefit must land on the same role (producers) that bears the new ROI-confirm cost.
2. Decision-maker profile (Three Ledgers)
- Public ledger (what we pitch): "One approved creative becomes every platform deliverable automatically — social, Amazon, print, motion — and it is product-aware, so the headphones and logo never get cropped wrong. We save producer hours and kill the Amazon/social rejection-retry loop." This is the visceral, fundable pitch that Finance (Ben Coleman archetype) and producers (Sofia Reyes archetype) can both restate in one sentence.
- Shadow ledger (what the decision-maker actually optimizes for): the sponsor (Sam) is optimizing for a defensible, demonstrable internal-tooling win that survives the funding review and does not become an embarrassment — i.e., a tool that visibly removes producer drudgery and that the brand team will stand behind. Underneath: avoiding the two failure shapes the org has seen before — (a) a tool that demos well and is quietly unused (Réti's "tool graveyard"), and (b) an automated output that ships an off-brand crop and becomes someone's named incident (Whitfield's accountability fear). The real optimization is risk-adjusted credibility, not raw automation coverage.
- True ledger (what will actually get built): realistically, Architecture B's tractable core on top of Architecture A's commodity substrate — a spec-registry + conformance + ingest-normalization batch layer (mostly bought/assembled, off-the-shelf FFmpeg/libvips/Sharp + a conformance gate), with a CV-ROI + geometric preservation envelope + brand-owned auto-clear threshold as the only genuinely Beats-proprietary layer. Motion will likely ship in a later increment than static (it is the weak flank everywhere). Architecture C (slots + constraint solver + edge render) will not get built as a funded program — at most its slot/intent-contract kernel survives as a research direction. What ships is gated behind the pre-build validation probe; if that probe fails, the "true ledger" collapses to "buy Auto Reframe + a folder watcher + a normalization step."
- Owner / sponsor: Sam (manifest owner). Accountable for the funding case and for sequencing the pre-build validation as a gate, not a milestone. Critical dependency: a named brand-governance owner (Whitfield archetype) who must agree to ratify and personally stand behind the auto-clear threshold — without that role committed, the proprietary value is un-bankable.
3. Constraint inventory
| Constraint | Type (hard / soft / assumed) | Notes |
|---|---|---|
| Brand-acceptability of a crop is contextual and may be irreducibly aesthetic | hard | The dominant systemic risk across all phases. If "on-brand" cannot be captured as measurable geometry, the entire proprietary value evaporates; this is what the ≥80% historical-replay gate exists to test. |
| Brand owner must personally ratify and stand behind any auto-clear threshold (accountability transfer) | hard | Whitfield's escalation of B#13: even an accurate number is worthless if no one will sign it, because the one bad auto-clear becomes their incident. Governance artifact, not an engineering setting. |
| Approved finals do NOT arrive in one tidy location/format (PSD, layered TIFF, AE projects, ad-hoc "final_FINAL_v3") | hard | Confirmed by Nair (Tier 1). Any watch-folder ingest needs a normalization step or it fails on day one; this is a real cost line, not an afterthought. |
| Apple-internal security review of pre-release product imagery on any edge/CDN render path | hard | C#31. Pre-release Beats imagery on external edge infrastructure is a security/governance minefield outside the project's control — a hard block on the zero-render edge premise. |
| Creative designers will not re-author masters as typed slots (entrenched Photoshop/After Effects throughput) | hard (org-cultural) | Nair declared this flatly and unprompted; it is C's primary kill candidate (C#22). Treated as immovable: do not build anything depending on a creative-workflow tool change. |
| Spec registry must be machine-readable data with a named human owner and review cadence | soft | Okafor + Nair both volunteered that an unowned registry is stale within two quarters and "conformant" becomes a lie. Maintainable only with a staffed operating model. |
| CV detector recall on adversarial Beats SKUs (glossy black on dark, embossed logos, occlusion) | assumed (must be benchmarked) | Vance: off-the-shelf recall will be materially worse than demo numbers on exactly the hero product. Treated as an assumption to validate pre-build, not a given. |
| Motion conformance (R128 loudness, color-gamut, codec-profile across ProRes/H.264/AE exports) | assumed (high QC cost) | Okafor: ~40% of broadcast QC engineering went here; false-quarantine on motion (A#6) is real. Budget explicitly or scope motion to a later increment. |
| Fully-loaded internal eng cost ≈ $220K/eng-year; internal-tool NPV hurdle 12% | soft (synthetic) | Phase 4 synthetic anchor, wide error bars — no real Beats financials were available. Used for comparability, not budgeting. |
| Producers (who bear the new ROI-confirm step) are not the role that reaps the brand/legal benefit | soft (incentive) | Cross-cutting adoption risk. The new cost must be near-zero-effort (one-click ROI accept) and the time-saved must be felt by producers, or adoption stalls regardless of code quality. |
4. Scope boundary
- In scope: Backend automation that takes one approved master 2D/motion creative and produces conformant platform deliverables (social, Amazon, print, motion) via master-plus-delta rendering; a machine-readable spec registry consulted at job-time; a conformance gate with structured per-clause error reports; ingest normalization for the real mix of approved-final formats; a CV-ROI + geometric product-preservation envelope with a brand-owned auto-clear threshold and a human-review queue for boundary cases; and — as the gating first work item — the historical-replay validation + brand-ratification + CV-recall pre-build probe. Static 2D is the primary launch surface; motion is in scope but explicitly sequenced behind static.
- Out of scope: Slot-based master authoring and any tool change forcing creatives off Photoshop/After Effects (C's premise — declared dead); a bespoke constraint-solver / "executable intent contract" formal layer (research direction only); zero-render edge/CDN delivery of pre-release imagery (security-blocked, and infeasible for motion); a self-discovering/auto-updating spec registry (registry is human-owned with a cadence); building the commodity render/conform engine from scratch where an off-the-shelf component (FFmpeg/libvips/Sharp/MediaConvert-class) suffices; the speculative "specs that negotiate back" idea.
- Recursion budget acknowledged: max 3 per phase. (Phase 5 explicitly recommended no phase recursion — the validators confirmed Phases 3–4.5 rather than invalidating them — so recursion budget is reserved for genuinely new disconfirming evidence, e.g., the pre-build probe failing.)
5. Assumption log
| # | Assumption | Confidence (0–1) | Validated? |
|---|---|---|---|
| 1 | Beats product-framing judgment is reducible to a geometric preservation envelope that agrees with historical human approve/reject decisions at ≥80%. | 0.45 | No — this is the binding pre-build gate; untested. Whitfield bets it comes back <80%. |
| 2 | A named brand-governance owner will ratify and personally stand behind a documented auto-clear threshold (accountability transfer, not just accuracy). | 0.40 | No — Whitfield's challenge says she would not sign without eyes on every output; must be secured as a governance artifact before build. |
| 3 | The crop-required share of real jobs is the majority (55–75%), so a tool that only auto-clears scale/pad cases is theater. | 0.70 | Partially — Phase 4 modeling assumption (Triangular 45/65/80%); consistent across phases but not measured on real Beats volume. |
| 4 | Off-the-shelf CV detectors reach ≥85% recall on real Beats SKUs/logos so "confirm ROI in seconds" holds. | 0.50 | No — Vance flags adversarial conditions; benchmark is a parallel pre-build gate. |
| 5 | The render/conform pipeline is commodity (master-plus-delta is solved at planetary scale via IMF), so the build effort and risk are not there. | 0.90 | Yes (qualitatively) — independently confirmed by Okafor and Tan; the proprietary surface is only the brand layer. |
| 6 | Approved finals require an ingest-normalization step (mixed PSD/TIFF/AE/ad-hoc names) rather than landing in one clean watch-folder. | 0.80 | Yes (qualitatively) — Nair confirmed from the creative-ops chair; treated as a real cost line. |
| 7 | The spec registry will go stale and "conformant" will become false unless a named owner with a review cadence is staffed. | 0.85 | Partially — pattern-confirmed by Okafor + Nair from prior shops; mitigated by making registry ownership a launch condition. |
| 8 | Motion conformance (loudness/color/codec across master formats) is materially harder than static and risks false-quarantine. | 0.75 | Partially — Okafor quantifies it (~40% of broadcast QC effort); not yet measured on a Beats motion corpus. |
| 9 | The new ROI-confirm step lands on the most deadline-pressured role (producers) and will be skipped or done carelessly unless near-zero-effort. | 0.70 | No — adoption hypothesis; mitigated by one-click accept design and role-segmented usage instrumentation. |
| 10 | Synthetic economics (≈$220K/eng-yr, throughput ~40–120 masters/mo, B-NPV slightly positive) are order-of-magnitude correct for go/no-go comparison. | 0.35 | No — explicitly AI-estimated, no Beats financial/telemetry data; usable only for relative ranking, not budgeting. |
Gate: success criteria defined, constraints mapped, decision-maker identified → set phaseGates.0-foundation = passed in manifest.json.