feat: resolution tier rewards

[dylanuys]

Problem

Veo C2PA manifests don't expose the model variant, so every Veo submission (lite, fast, full) was priced at the cheapest variant's rate ($0.03/s). Miners had no incentive to run anything above veo-lite, despite a 20x cost spread across the family. Separately, resolution wasn't priced at all — for any model, minimum-resolution output earned the same reward as top-tier output.

Approach

Resolution and audio presence are covered by the C2PA hard binding we already require (assertion.bmffHash.match), so they are cryptographically attested signals: a miner can't upscale, re-encode, or mux in audio without invalidating the manifest. We use them to price rewards by delivered resolution tier.

Video — tiers 480p / 720p / 1080p (challenge mix 20/40/40, CHALLENGE_TIER_WEIGHTS). 4K is deliberately not a tier for now: it is never requested, and 4K output classifies and prices as 1080p. The Veo 4K price entries remain in MODEL_TIER_PRICES (the live SKU merge re-adds them anyway) but are unreachable until 4K rejoins the tier list.

Image — tiers 1K / 2K / 4K by minimum pixel dimension (nominal 1024/2048/4096, challenge mix 40/40/20). Image C2PA manifests rarely expose a usable model variant and image APIs price per image (typically tiered by output size), so images are priced purely by delivered tier with relative prices 1x/2x/4x → after the sqrt taper, multipliers of 1x/1.41x/2x.

Shared mechanics for both modalities:

• Challenges request a tier via generation parameters["resolution"]; rewards price at min(observed, requested):
  • Overshooting a request never pays more, so ignoring requests is never profitable.
  • A miner whose model can't reach the requested tier delivers a lower one and gets that tier's price — graceful degradation, no failure state, no capability negotiation.
• Veo (C2PA-blind, model_name=None) is priced as the fast variant at its (tier, audio) combination. Flooring at lite was maximally arbitrage-safe but credited every real fast/full generation at lite rates, capping the family at 1.63x; the fast floor bounds a lite miner's overcredit at ~2x on price while Google-certified content earns a fair mid-family rate (720p 1.63-1.83x, 1080p 1.83-2.0x). Veo can't produce sub-720p, so a 480p-tier challenge prices at Seedance fast's 480p rate ($0.07/s -> 1.53x) — the bottom tier pays the same regardless of family. Tier/audio price tables are parsed from OpenRouter pricing_skus key suffixes at module load (hardcoded fallbacks included), so capability and price drift self-correct.
• Named models (Seedance exposes model_name via C2PA) use explicit per-tier price tables (NAMED_MODEL_TIER_PRICES) from provider quotes, since OpenRouter publishes no Seedance pricing SKUs: seedance-2-0 $0.50/$2.00/$2.40 per 5s at 480p/720p/1080p, fast $0.35/$0.76 at 480p/720p. The spread is not pixel-proportional, so tables beat a shared scale. Resulting multipliers: fast 1.53x/2.25x; full 1.83x/3.65x/4.0x. A tier absent from a model's table pays its top priced tier (no extrapolated premium); named models without a table fall back to pixel-ratio scaling of their flat 720p reference, floored at baseline.

Changes

• gas/evaluation/resolution_tiers.py (new): per-modality tier definitions, orientation-agnostic min-dimension classification, weighted sampling, effective_tier.
• gas/evaluation/generative_challenge_manager.py: samples and sends the requested tier for both modalities, records it on the challenge outcome.
• gas/evaluation/rewards.py: MODEL_TIER_PRICES (Veo floor pricing), IMAGE_TIER_PRICES, per-generation multipliers for both modalities; falls back to model-name pricing for stats that predate tiered challenges.
• gas/cache/: schema migration adding requested_resolution to generator_challenge_outcomes and has_audio to media; miner video writes detect the audio track (ffprobe); reward stats join media onto outcomes and emit structured image_generations / video_generations lists.

No miner changes required — the reference miner already forwards parameters["resolution"] and snaps to the nearest supported value. Requested vs. observed tier is stored per outcome, so compliance is queryable before deciding whether penalties are ever warranted.

Testing

• Full test suite passes (37 tests).
• End-to-end check: challenge outcome → media join → stats → tier-priced reward multiplier for both modalities, including overshoot capping, undershoot degradation, unknown-resolution baseline, 4K-video-classifies-as-1080p, and image tier classification (portrait/landscape, sub-1K floor).

Deploy notes

• Schema migration auto-applies on validator start.
• Worth confirming with one live generation per variant that OpenRouter's un-suffixed duration_seconds SKU is 1080p output (price tables self-correct via the live fetch; that naming assumption doesn't).
• IMAGE_TIER_PRICES ratios (1/2/4) and the challenge tier mixes are the tuning knobs; both are deliberately conservative first guesses.