Spaces:
Sleeping
Sleeping
Commit History
-
Fix G06 500 error: call env.reset() on session creation to init hidden state
echoboi +1: Claude Sonnet 4.6G06 (DirectionalFlow) requires orientation to be set before get_true_next() can be called. Session creation never called reset(), leaving _orientation=None and causing RuntimeError -> 500 on submit_rule. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>3dbfa3d
-
Add cell_errors, error_regions, common_error_patterns to submit_rule response
echoboi +1: Claude Sonnet 4.6Tracks worst-performing test state during evaluation and computes: - cell_errors: up to 10 wrong cells with 8-neighbor context - error_regions: quadrant error distribution + worst state accuracy - common_error_patterns: top-5 (predicted, expected, count) pairs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>5931b5f -
Fix submit_rule: random eval seed + held-out validation set
echoboi +1: Claude Sonnet 4.6- Use random.randint(1, 999_999) seed each submission to prevent overfitting to a fixed test set (was hardcoded seed=9999) - Add held-out val set (n=200, seed=42_000) for true accuracy estimate - Return val_accuracy, val_exact_accuracy, and eval_seed in response Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>86a79eb
-
.
echoboi591d55e -
.
echoboid2a501b -
.
echoboibfc20f2 -
.
echoboia2ab10e -
Remove Prime Intellect branding from dashboard
echoboi1dda0c3 -
Deploy Discovery Environment API + Dashboard with agent run logs
echoboid0e30e9
-
Strip docstrings from description length (parsimony scoring)
echoboi +1: Claude Sonnet 4.6strip_comments() now does a two-pass strip: 1. AST pass: removes module/class/function docstring nodes (Expr(Constant(str)) in first-statement position) 2. Tokenize pass: removes # inline comments Before: agent code with large docstrings was penalised vs clean code even when the actual algorithm was identical. After: stripped_code_length(code_with_docs) == stripped_code_length(code_clean) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>77b9e25 -
Make simulate endpoint robust to all state_json formats
echoboi +1: Claude Sonnet 4.6Models send state_json in many forms: - list/nested list (JSON array parsed by pydantic) - str containing JSON array "[[1,2],[3,4]]" - double-encoded str "\"[[1,2],[3,4]]\"" - str with Python repr prefix "array([[...]])" _parse_state() iteratively unwraps all of these. SimulateRequest.state_json changed from str to Any so pydantic doesn't reject array values before they reach the endpoint. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9e4cfa3 -
Fix scoring: use cell-level partial credit instead of binary exact-state match
echoboi +1: Claude Sonnet 4.6functional_accuracy now = mean(np.mean(predicted==expected)) across 500 test states so a rule that gets 90% of cells right scores 0.9 instead of 0.0. Old binary metric preserved as exact_accuracy for reference. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>bff2931
-
Add .gitignore, remove __pycache__
echoboi4ff2077 -
Deploy Discovery Environment API
echoboi8c464db