This repository was archived by the owner on Aug 3, 2026. It is now read-only.
op1: reject fabricated compute + config-vs-proof mismatch (compute-gaming) - #83
Merged
Merged
Conversation
…ming) - check_compute_plausibility: reject a bundle whose tokens_seen/wall_clock_s exceeds the declared GPU's bf16 MFU ceiling (>70%). wall_clock_s is miner-declared + not in bundle_hash, so it can be under-claimed to win the compute-weighted crown. implied MFU = 6*N*(tok/s) / gpu_bf16_peak. - check_recipe_config_matches_proof: reject when a submitted configs/*.json total_steps != the steps the proof actually ran (decoy recipe). - wired into op1_diff_and_integrity (reject before scoring). - catches the live king 5CtaoQf1: 818k tok/s on one H100 = 126% MFU (impossible); submitted muon/40000-step config but proof ran adamw/10600. - 8 tests incl. legit-30h + optimized-250k + unknown-GPU + incomplete guards. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
karpabot
force-pushed
the
op1-compute-config
branch
from
June 30, 2026 10:13
39c78ba to
acb23d8
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Two op1 gates against compute-gaming — the fraud the current live king
5CtaoQf1used.The king trained a real model but fabricated its compute to win the efficiency-weighted crown: it claimed 5.557B tokens in 6788s on a single H100 = 818k tok/s = 126% of H100 bf16 peak MFU — physically impossible (real sustained training MFU is 30–55%; the canonical recipe measures 51k tok/s).
wall_clock_sis miner-declared and not inbundle_hash, so it's free to forge. It also submitted amuon/40000-step config while the proof actually ranadamw/10600 steps.Checks (
validator/integrity.py)check_compute_plausibility(final_state, calibration)— implied MFU =6·N·(tokens/wall) / gpu_bf16_peak(gpu_name); reject if> 70%. Unknown GPU → assume the fastest known part, so it never false-rejects.check_recipe_config_matches_proof(patch, final_state)— a submittedconfigs/*.jsontotal_stepsmust equal the proof's recorded steps.Wired into
op1_diff_and_integrity(rejects before any compute is spent).Verified
5CtaoQf1→ rejected on both (126% MFU; 40000 ≠ 10600 steps).final_state→ skipped, not rejected. 15 tests pass.Caveat
The root issue is trusting a self-reported
wall_clock_s; the airtight fix is binding an attested duration (TDX trusted clock) into the compute cost. This gate makes the obvious fabrication non-viable.🤖 Generated with Claude Code