This repository was archived by the owner on Aug 3, 2026. It is now read-only.
op1 anti-gaming hardening + pre-crown re-derivation (netuid 40) - #89
Merged
Merged
Conversation
- reject if declared wall_clock_s > canonical code age + slack - anchored to pinned RECIPE_DIR commit epoch (stable across validator deploys) - pairs with MFU gate: too-long trips this, too-short trips MFU - fail-open on missing wall_clock/epoch; RALPH_TIMING_GATE_OFF + RALPH_TIMING_SLACK_S knobs - catches the off-protocol crowns (6.95h run attested to 2.8h-old canonical code) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- add final_state to the required-artifact loop (reject if missing) - append final_state to bundle_hash UNCONDITIONALLY (match proof.runner) - closes one-file bypass: deleting final_state.json made the validator compute a self-consistent 4-component hash and SKIP every anti-gaming gate - verified live: no-final_state bundle -> "missing artifact final_state" Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
karpabot
changed the title
- reject any bundle whose (authenticated) checkpoint_sha256 is in chain/fraud_checkpoints.json - stops re-submission of an identical off-protocol checkpoint under a new bundle hash + adjusted metadata (the timing gate weakens as the canonical code ages, so a waited-out resubmit otherwise re-crowns) - fail-open on missing/malformed blocklist Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- compare_loss_trajectory (integrity.py): step-0 fingerprint + trajectory-band compare of declared vs re-derived training loss; the check off-protocol training cannot pass without actually running the canonical recipe - op_rederive_trajectory (validator.py): apply patch -> run real train.py on CANONICAL data with the miner config+seed -> watch first few log points -> compare. Sandboxed subprocess (reuses op4 pattern). Wired into judge_submission between op3 and op4. - RALPH_REDERIVE=1 to enable; fail-open (skip) when disabled or when canonical data_manifest.json is not materialized on the validator. UNVALIDATED until a validator has canonical data to run it against; calibrate the band before enabling. - tests: timing gate, fraud blocklist, trajectory compare Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
karpabot
changed the title
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.
op1/op-level anti-gaming hardening for the off-protocol-training class (netuid 40).
LIVE + deployed on the validator now:
Dormant / OFF by default (needs canonical data + calibration before enabling):
4. Pre-crown re-derivation (op_rederive_trajectory + compare_loss_trajectory) - re-run the first steps of the declared training on CANONICAL data and require the loss trajectory to reproduce. This is the only check off-protocol training cannot satisfy. Gated behind RALPH_REDERIVE=1; fail-open (skip) when disabled or when the validator has no materialized canonical data_manifest.json. Written but UNVALIDATED until a validator has canonical data to run it against.
Prereq for (4): materialize the canonical FineWeb data (data.prepare) on the validator, then calibrate the tolerance band on a known-honest bundle. Ideally the crown path invokes (4) only for a bundle that would beat the king (GPU-minutes).
All validator-only (no measured surface) -> no measurement cutover. Tests included for 1/3/4.