Scoring
Impact = how much it moves the core loop (cataloging speed, review trust, conversion, defensibility). Effort = build cost to a usable v1. Feasibility = what it depends on.
| # | Idea | Impact | Effort | Feasibility | Verdict |
|---|---|---|---|---|---|
| 1 | Voice cataloging | High | Med | model · local free | Do first — best pain-per-dollar. |
| 2 | Cross-run dedupe | High | Med | model · cheap | Do first — unlocks scale + recovery. |
| 3 | PII / privacy scrubber | Med-High | Med-High | model (vision) | Do, fast-follow — differentiator. |
| 4 | Confidence-tiered promotion | High | Low | deterministic | Sleeper pick — ship before models. |
| 5 | Provenance timeline | Med | Low | deterministic | Quick win; trust + debugging. |
| 6 | Throughput / SLA dashboard | Med (agency) | Low-Med | deterministic | Cheap; pair with agency push. |
| 7 | Comparable-evidence cards | High | Med | surface free deep metered | High-trust; phased. |
| 8 | Outcome feedback loop | High (compounding) | Med-High | deterministic · needs data | Defer until pilot data. |
| 9 | Auction-pipeline + channel export adapters | Highest value | High | file earlier API needs Auth | Phase-2 — reseller/auction payoff. Spec. |
| 10 | Photo decomposition / video-sweep | Med-High | High | model (segmentation) | Defer — accuracy risk. |
| 11 | Active-learning review | Med | High | model + data | Defer — needs edit corpus. |
| 12 | Authenticity detector | Med | High | model | Defer — false-positive risk. |
Reading the 2×2
- High impact / low effort / deterministic — do now, no model needed: confidence-tiered promotion (4), provenance timeline (5), throughput dashboard (6). Ship keyless, stay green in CI, and together materially improve throughput and trust before spending on inference. #4 is the most underrated — pure policy logic over the existing promote path that multiplies reviewer throughput.
- High impact / medium effort / free-capable model — the headline bets: voice cataloging (1), cross-run dedupe (2). Both have a genuine free local path (Whisper, perceptual hash), so they fit the cost-controlled rule and only escalate to paid inference when quality demands. This is where the "kills the pain" story lives.
- Highest value but mostly Auth-gated — stage it: auction-pipeline + channel export adapters (9). The literal "catalog once, list everywhere" promise and the reseller/auction pitch. The file exports (LiveAuctioneers / Proxibid / Auction Flex spreadsheet + lot-numbered image bundle) can ship without third-party Auth; the API + sale-results pull-back need Auth + connectors. See the auction-outbound integration spec.
- Defer on feasibility risk, not value: decomposition/video-sweep (10), active-learning (11), authenticity (12). All need models and either accuracy guarantees or a training corpus not yet available, with no free fallback.
Recommended sequence
- Deterministic trio — confidence-tiered promotion → provenance timeline → throughput dashboard. Fast, keyless, de-risks the queue.
- Voice cataloging + cross-run dedupe — the two free-capable model wins that carry the marketing story.
- PII scrubber — fast-follow trust differentiator.
- Comparable-evidence cards, then outcome feedback loop once sold-price data accumulates.
- Auction-pipeline + channel export adapters — file exports first (no Auth), then the API + sale-results pull-back when Auth + connectors are live.
Trap to avoid: the flashiest ideas (video-sweep, decomposition, authenticity) are the worst first bets — model-dependent, no free fallback, quality-sensitive. The deterministic policy work is unglamorous but it's where the cheap throughput gains actually are.