Skip to content

scoring.md

Latest commit

 

History

History
80 lines (58 loc) · 2.87 KB

scoring.md

File metadata and controls

80 lines (58 loc) · 2.87 KB
Raw

Scoring and raw weights

Hypercluster uses a fixed four-factor product. Do not invent alternate product formulas.

Formula (per attempt)

composite = correctness × efficiency × fabric_gate × tee_bonus
Factor Type Definition
correctness gate {0,1} Golden microbench / output digest / schema validation. Zero zeroes all
efficiency continuous ≥ 0 Compute-normalized metric. Wall clock is a cap, not the primary score
fabric_gate gate {0,1} One when required fabric is present without forbidden fallback; else zero
tee_bonus multiplier ≥ 1 1.0 without verified TEE; higher only when proof verifies under policy

Integrity zero

Any integrity failure forces composite = 0 for that attempt, including attestation fail, image mutation, rank desync, inventory spoof, fabric lie, and GPU probe honesty fails (failed/error evidence under live-evidence policy, or claim-vs-measured mismatch vs last good GpuHostEvidence).

GPU probe does not add a fifth formula factor. Defaults leave the unprobed sim path green; operators may inject zeros with HYPER_SIM_GPU_PROBE_FAIL for CI drills. See GPU probe.

Aggregation → raw weights

  1. Select a scoring window (for example last N attempts via HYPER_SCORE_WINDOW_ATTEMPTS).
  2. Aggregate demand and/or supply composites per hotkey (default: sum of positive composites).
  3. Apply soft penalties (self-deal rate, spam fails) as policy.
  4. Emit dict[hotkey, float] with finite values ≥ 0.
  5. Prefer burn-safe empty epochs over inventing fake mass.
  6. Challenge never calls set_weights; Base master normalizes emission and maps hotkeys → UIDs.

Weight push

Background or CLI-triggered push builds a monochronic snapshot:

  • epoch, increasing revision
  • tz-aware computed_at / expires_at
  • payload_digest over canonical payload
  • POST raw weights to master (local mock supported for development)

Serve the same map via Base get_weights_fn / internal GET.

Suggested knobs

HYPER_TEE_BONUS_TDX=1.08
HYPER_TEE_BONUS_TDX_GPU=1.20
HYPER_WEIGHT_PUSH_INTERVAL_S=120
HYPER_SCORE_WINDOW_ATTEMPTS=50
HYPER_EFFICIENCY_FLOOR=0.0
HYPER_MASTER_BASE_URL=http://127.0.0.1:3201

HTTP routes

Method Path Purpose
GET /v1/leaderboard Aggregated composites
GET /v1/scores/{hotkey} History
GET /v1/weight-preview Pending/latest raw map (may be restricted in deploy)

CLI

hypercluster score show --hotkey HOTKEY
hypercluster score recompute [--epoch N]
hypercluster weights preview
hypercluster weights push --epoch E --revision R

Never print challenge tokens in CLI output.

Local validation

hypercluster sim run-scenario --name weights --url http://127.0.0.1:3200
# with mock master env if needed:
# HYPER_MASTER_BASE_URL=http://127.0.0.1:3201