was mache ich hier?: & { Set-StrictMode -Version Latest ...

Prompt

was mache ich hier?: & { Set-StrictMode -Version Latest $ErrorActionPreference = 'Stop' $root = 'C:\Users\user\downloads\rlcard-master\rlcard-master' $py = 'C:\Users\user\AppData\Local\Python\pythoncore-3.14-64\python.exe' $path = Join-Path $root ` 'general_research_factorized_recurrent_semantics_v0_1\d6b_task_fixed_arm_diagnostic_run_v0_1.py' $outDir = Join-Path $root ` 'experiments\reasoning\semantic-mechanism-diagnosis-v0_1\stage-d6b-task-fixed-arm-diagnostic-v0_1' $stdout = Join-Path $outDir 'RUN_STDOUT.txt' $stderr = Join-Path $outDir 'RUN_STDERR.txt' $text = @' def sign_counts(values): values = tuple(float(v) for v in values) return { "negative": sum(v < 0.0 for v in values), "zero": sum(v == 0.0 for v in values), "positive": sum(v > 0.0 for v in values), } def summarize_delta(values): values = tuple(float(v) for v in values) require( bool(values), "cannot summarize empty values", ) require( all(math.isfinite(v) for v in values), "non-finite summary value", ) return { "mean": sum(values) / len(values), "median": statistics.median(values), "negative_zero_positive_counts": sign_counts(values), } def selected_arm_counts(rows): counts = Counter( row["task_conditioned_arm"] for row in rows ) return { arm: counts.get(arm, 0) for arm in core.ARMS } def aggregate_rows(rows): require( bool(rows), "cannot aggregate empty rows", ) delta = [ row["task_conditioned_minus_global"] for row in rows ] task_excess = [ row["task_conditioned_excess_over_oracle"] for row in rows ] global_excess = [ row["global_excess_over_oracle"] for row in rows ] task_oracle_matches = sum( row["task_conditioned_arm"] == row["oracle_arm"] for row in rows ) global_oracle_matches = sum( row["global_fixed_arm"] == row["oracle_arm"] for row in rows ) mean_task_excess = ( sum(task_excess) / len(task_excess) ) mean_global_excess = ( sum(global_excess) / len(global_excess) ) return { "cell_count": len(rows), "selected_arm_counts": selected_arm_counts(rows), "mean_task_conditioned_minus_global": sum(delta) / len(delta), "median_task_conditioned_minus_global": statistics.median(delta), "negative_zero_positive_counts": sign_counts(delta), "task_conditioned_oracle_match_count": task_oracle_matches, "global_oracle_match_count": global_oracle_matches, "mean_task_conditioned_excess_over_oracle": mean_task_excess, "mean_global_excess_over_oracle": mean_global_excess, } def build_result(rows): require( len(rows) == 32, "expected exactly 32 diagnostic rows", ) overall = aggregate_rows(rows) global_gap = float( overall[ "mean_global_excess_over_oracle" ] ) task_gap = float( overall[ "mean_task_conditioned_excess_over_oracle" ] ) if global_gap > 0.0: recovery_fraction = ( global_gap - task_gap ) / global_gap elif global_gap == 0.0: recovery_fraction = ( 0.0 if task_gap == 0.0 else None ) else: recovery_fraction = None per_task = {} for task in d6b.TASKS: subset = [ row for row in rows if row["task"] == task ] require( len(subset) == 8, f"task {task}: expected 8 cells", ) per_task[str(task)] = ( aggregate_rows(subset) ) return { "schema_version": 1, "record_type": "d6b_task_fixed_arm_diagnostic_result_v0_1", "status": "posthoc_development_diagnostic_only", "preregistration_freeze_sha256": EXPECTED_PREREG_FREEZE_SHA256, "source_failure_audit_sha256": EXPECTED_AUDIT_RESULT_SHA256, "method": "task-conditioned leave-one-seed-out fixed-arm selection", "overall": { "mean_task_conditioned_minus_global": overall[ "mean_task_conditioned_minus_global" ], "median_task_conditioned_minus_global": overall[ "median_task_conditioned_minus_global" ], "negative_zero_positive_counts": overall[ "negative_zero_positive_counts" ], "task_conditioned_oracle_match_count": overall[ "task_conditioned_oracle_match_count" ], "global_oracle_match_count": overall[ "global_oracle_match_count" ], "mean_task_conditioned_excess_over_oracle": overall[ "mean_task_conditioned_excess_over_oracle" ], "mean_global_excess_over_oracle": overall[ "mean_global_excess_over_oracle" ], "oracle_gap_recovery_fraction": recovery_fraction, }, "per_task": per_task, "per_cell": rows, "interpretation_contract": { "confirmatory": False, "can_rescue_d6b": False, "causal_claim_permitted": False, "future_mechanism_requires_fresh_seeds": True, }, "new_seeds_used": False, "new_rl_outcomes_generated": False, "features_used": False, "model_trained": False, "hyperparameter_tuning": False, "threshold_search": False, } def main(): rows = build_loso_cells() result = build_result(rows) digest = write_once( RESULT, result, ) print( "D6B_TASK_FIXED_RESULT_SHA256", digest, ) overall = result["overall"] print( "D6B_TASK_FIXED_MEAN_MINUS_GLOBAL", format( overall[ "mean_task_conditioned_minus_global" ], ".17g", ), ) print( "D6B_TASK_FIXED_MEDIAN_MINUS_GLOBAL", format( overall[ "median_task_conditioned_minus_global" ], ".17g", ), ) counts = overall[ "negative_zero_positive_counts" ] print( "D6B_TASK_FIXED_NEG_ZERO_POS", counts["negative"], counts["zero"], counts["positive"], ) print( "D6B_TASK_FIXED_ORACLE_MATCHES", overall[ "task_conditioned_oracle_match_count" ], "/ 32", ) print( "D6B_GLOBAL_FIXED_ORACLE_MATCHES", overall[ "global_oracle_match_count" ], "/ 32", ) print( "D6B_TASK_FIXED_EXCESS_ORACLE_MEAN", format( overall[ "mean_task_conditioned_excess_over_oracle" ], ".17g", ), ) print( "D6B_GLOBAL_FIXED_EXCESS_ORACLE_MEAN", format( overall[ "mean_global_excess_over_oracle" ], ".17g", ), ) print( "D6B_TASK_FIXED_ORACLE_GAP_RECOVERY_FRACTION", ( format( overall[ "oracle_gap_recovery_fraction" ], ".17g", ) if overall[ "oracle_gap_recovery_fraction" ] is not None else "UNDEFINED" ), ) for task in d6b.TASKS: value = result[ "per_task" ][str(task)] counts = value[ "negative_zero_positive_counts" ] print( "D6B_TASK_FIXED_TASK", task, "ARMS", json.dumps( value[ "selected_arm_counts" ], sort_keys=True, ), "MEAN_MINUS_GLOBAL", format( value[ "mean_task_conditioned_minus_global" ], ".17g", ), "NEG_ZERO_POS", counts["negative"], counts["zero"], counts["positive"], "TASK_ORACLE", value[ "task_conditioned_oracle_match_count" ], "GLOBAL_ORACLE", value[ "global_oracle_match_count" ], "TASK_EXCESS", format( value[ "mean_task_conditioned_excess_over_oracle" ], ".17g", ), "GLOBAL_EXCESS", format( value[ "mean_global_excess_over_oracle" ], ".17g", ), ) print( "D6B_TASK_FIXED_CONFIRMATORY = False" ) print( "D6B_TASK_FIXED_CAN_RESCUE_D6B = False" ) print( "D6B_TASK_FIXED_NEW_SEEDS_USED = False" ) print( "D6B_TASK_FIXED_NEW_RL_OUTCOMES_GENERATED = False" ) if __name__ == "__main__": main() '@ [System.IO.File]::AppendAllText( $path, $text, [System.Text.UTF8Encoding]::new($false) ) $env:PYTHONPATH = $root try { & $py -B -m py_compile $path if ($LASTEXITCODE -ne 0) { throw 'Task-fixed diagnostic runner py_compile failed.' } Write-Host 'D6B_TASK_FIXED_RUNNER_PY_COMPILE_OK' $result = Join-Path $outDir 'ANALYSIS_RESULT.json' if (Test-Path -LiteralPath $result) { throw "Unexpected existing result: $result" } Remove-Item -LiteralPath $stdout -ErrorAction SilentlyContinue Remove-Item -LiteralPath $stderr -ErrorAction SilentlyContinue $process = Start-Process ` -FilePath $py ` -ArgumentList @( '-u', '-B', '-m', 'general_research_factorized_recurrent_semantics_v0_1.d6b_task_fixed_arm_diagnostic_run_v0_1' ) ` -WorkingDirectory $root ` -RedirectStandardOutput $stdout ` -RedirectStandardError $stderr ` -Wait ` -NoNewWindow ` -PassThru Write-Host '' Write-Host '================ TASK-FIXED STDOUT ================' Get-Content -LiteralPath $stdout if ( (Test-Path -LiteralPath $stderr) -and ((Get-Item -LiteralPath $stderr).Length -gt 0) ) { Write-Host '' Write-Host '================ TASK-FIXED STDERR ================' Get-Content -LiteralPath $stderr } if ($process.ExitCode -ne 0) { throw "Task-fixed diagnostic failed with exit code $($process.ExitCode)" } Write-Host '' Write-Host 'D6B_TASK_FIXED_PROCESS_EXIT_CODE 0' } finally { Remove-Item Env:PYTHONPATH -ErrorAction SilentlyContinue } } Poste danach den Output ab D6B_TASK_FIXED_RESULT_SHA256. Dann können wir direkt beurteilen, ob Task-Konditionierung einen relevanten Anteil des Oracle-Gaps zurückgewinnt oder ob die entscheidende Variation tatsächlich innerhalb der Tasks zwischen Seeds liegt. D6B\_TASK\_FIXED\_RUNNER\_PY\_COMPILE\_OK \================ TASK-FIXED STDOUT ================ D6B\_TASK\_FIXED\_BOUND\_LOSO\_CELLS 32 D6B\_TASK\_FIXED\_RESULT\_SHA256 54b3afb3c671f85fd43151e133c400846e884e00a19ec48d7a2b5d6a4d4bc858 D6B\_TASK\_FIXED\_MEAN\_MINUS\_GLOBAL 0.017648228410462895 D6B\_TASK\_FIXED\_MEDIAN\_MINUS\_GLOBAL 0 D6B\_TASK\_FIXED\_NEG\_ZERO\_POS 7 16 9 D6B\_TASK\_FIXED\_ORACLE\_MATCHES 10 / 32 D6B\_GLOBAL\_FIXED\_ORACLE\_MATCHES 12 / 32 D6B\_TASK\_FIXED\_EXCESS\_ORACLE\_MEAN 0.063214806432874443 D6B\_GLOBAL\_FIXED\_EXCESS\_ORACLE\_MEAN 0.045566578022411555 D6B\_TASK\_FIXED\_ORACLE\_GAP\_RECOVERY\_FRACTION -0.38730642449785779 D6B\_TASK\_FIXED\_TASK 3 ARMS {"ALL\_LEGAL\_RIDGE": 0, "SEMANTIC\_COVERAGE": 0, "TOP2\_RIDGE": 0, "UNIFORM": 8} MEAN\_MINUS\_GLOBAL -0.029804041635493092 NEG\_ZERO\_POS 5 0 3 TASK\_ORACLE 3 GLOBAL\_ORACLE 2 TASK\_EXCESS 0.027684003399192511 GLOBAL\_EXCESS 0.057488045034685603 D6B\_TASK\_FIXED\_TASK 14 ARMS {"ALL\_LEGAL\_RIDGE": 6, "SEMANTIC\_COVERAGE": 1, "TOP2\_RIDGE": 0, "UNIFORM": 1} MEAN\_MINUS\_GLOBAL 0.072744274312089685 NEG\_ZERO\_POS 0 6 2 TASK\_ORACLE 0 GLOBAL\_ORACLE 1 TASK\_EXCESS 0.14417579887857201 GLOBAL\_EXCESS 0.071431524566482357 D6B\_TASK\_FIXED\_TASK 15 ARMS {"ALL\_LEGAL\_RIDGE": 8, "SEMANTIC\_COVERAGE": 0, "TOP2\_RIDGE": 0, "UNIFORM": 0} MEAN\_MINUS\_GLOBAL 0 NEG\_ZERO\_POS 0 8 0 TASK\_ORACLE 6 GLOBAL\_ORACLE 6 TASK\_EXCESS 0.0063604070735948648 GLOBAL\_EXCESS 0.0063604070735948648 D6B\_TASK\_FIXED\_TASK 16 ARMS {"ALL\_LEGAL\_RIDGE": 2, "SEMANTIC\_COVERAGE": 0, "TOP2\_RIDGE": 0, "UNIFORM": 6} MEAN\_MINUS\_GLOBAL 0.027652680965255 NEG\_ZERO\_POS 2 2 4 TASK\_ORACLE 1 GLOBAL\_ORACLE 3 TASK\_EXCESS 0.074639016380138395 GLOBAL\_EXCESS 0.046986335414883396 D6B\_TASK\_FIXED\_CONFIRMATORY = False D6B\_TASK\_FIXED\_CAN\_RESCUE\_D6B = False D6B\_TASK\_FIXED\_NEW\_SEEDS\_USED = False D6B\_TASK\_FIXED\_NEW\_RL\_OUTCOMES\_GENERATED = False D6B\_TASK\_FIXED\_PROCESS\_EXIT\_CODE 0 PS C:\Users\user\downloads\rlcard-master\rlcard-master> Weniger anzeigen