This repository was archived by the owner on Aug 3, 2026. It is now read-only.
Fix/validator scoring weights - #66
Merged
Merged
Conversation
- the rate-limit recovery merge re-added the previous king at king-level weight on top of the new king's, every epoch, until a set_weights finally landed and cleared pending_weights -> emission split across two kings (the "validator weights both kings after a change" bug). - _apply_pool_split already weights the CURRENT king authoritatively, so a recovered king-level weight (>= KING_POOL_FRACTION) for any non-current king is a stale reign -> drop it; keep sub-king meaningful_failure recoveries. - extract _merge_recovered_weights helper + regression tests. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- calibrate the H100 cost reference: matmul_ms_ref 5.0 (placeholder) -> 0.51 (measured on H100 PCIe, run_calibration). The old value inflated normalized H100-hours ~14x (a ~7 wall-clock-hour run read as 102 "H100h"). - net-score crown gate: a challenger that beats the king on quality/benchmark must ALSO be net-positive after compute is charged (score > 0) -> an equal- quality cheaper model wins, runaway compute is rejected, no hard hour cap. - tunable: RALPH_COMPUTE_COST_WEIGHT (default 0.002 = 500 H100h per bpb-point), RALPH_H100_MATMUL_MS_REF, RALPH_COMPUTE_CROWN_GATE escape hatch. - current king stays crowned (score +3.32); a 5x-compute clone is rejected. +6 gate tests; full suite 898 passed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
CI cloned recipe main, which tracks the latest king's auto-merged recipe; a U-Net skip-gate model there index-errors on the tests' small config, so main + every PR went red on 3 unrelated tests (multiseed + patched_eval). Pin to the canonical recipe the validator actually runs. Full suite: 905 passed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A challenger deferred by the king min-tenure guard sits in queue/pending and is re-scored EVERY epoch until tenure lapses (~300 blocks), re-running the ~90s GPU hidden-eval each time even though the bundle + eval shard are immutable — pure waste that also blocks the GPU from processing new submissions. Cache the op4 result as a per-bundle dotfile keyed on a fingerprint of the eval shard (auto- invalidates on eval rotation; op1 integrity is manifest-based so the file is ignored). +4 tests; full suite 909 passed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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.
1 participant
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.
No description provided.