diff options
Diffstat (limited to 'experiments/analyze_kp_dynamic_projection_full.py')
| -rwxr-xr-x | experiments/analyze_kp_dynamic_projection_full.py | 223 |
1 files changed, 223 insertions, 0 deletions
diff --git a/experiments/analyze_kp_dynamic_projection_full.py b/experiments/analyze_kp_dynamic_projection_full.py new file mode 100755 index 0000000..90b70df --- /dev/null +++ b/experiments/analyze_kp_dynamic_projection_full.py @@ -0,0 +1,223 @@ +#!/usr/bin/env python3 +"""Audit the conditionally frozen D3 full dynamic-projection endpoint.""" +import argparse +import json +import math +import os + + +SPLIT_HASH = "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b" +KP_FULL_ACCURACY = 0.9126 + + +def numeric_leaves(value): + if isinstance(value, bool) or value is None: + return + if isinstance(value, (int, float)): + yield float(value) + elif isinstance(value, dict): + for child in value.values(): + yield from numeric_leaves(child) + elif isinstance(value, (list, tuple)): + for child in value: + yield from numeric_leaves(child) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--input", default="results/kp_dynamic_projection_full/dynamic.json") + parser.add_argument( + "--d2_gate", default="results/kp_dynamic_projection_short_gate.json") + parser.add_argument( + "--bp_selection", default="results/oral_a_bp_selection.json") + parser.add_argument( + "--out", default="results/kp_dynamic_projection_full_gate.json") + args = parser.parse_args() + with open(args.d2_gate) as handle: + d2 = json.load(handle) + if (d2.get("protocol") != "kp_dynamic_neutral_projection_short_v1" + or d2.get("status") != "passed" + or d2.get("full_validation_opened") is not True): + raise ValueError("D3 requires the audited D2 pass") + with open(args.bp_selection) as handle: + bp_selection = json.load(handle) + if bp_selection.get("status") != "passed_primary": + raise ValueError("D3 requires the frozen full BP reference") + with open(bp_selection["selected"]["path"]) as handle: + bp = json.load(handle) + bp_accuracy = float(bp["final"]["accuracy"]) + bp_macs = int(bp["work"]["total_macs_estimate"]) + if bp_accuracy != 0.9162: + raise ValueError("D3 BP accuracy reference drift") + + with open(args.input) as handle: + record = json.load(handle) + expected = { + "mode": "kp_traffic", "traffic_rule": "innovation", + "predictor_mode": "closed_form", "neutral_projection": 1, + "depth": 20, "width": 16, "seed": 0, "loader_seed": 0, + "batch_size": 128, "epochs": 200, "train_limit": 0, + "val_examples": 5000, "split_seed": 2027, + "eval_split": "validation", "eval_every": 0, + "augment_train": 1, "lr": 0.1, "output_lr": 0.1, + "lr_schedule": "step", "lr_milestones": "100,150", + "lr_gamma": 0.1, "warmup_epochs": 0, "momentum": 0.9, + "weight_decay": 1e-4, "normalization": "batchnorm", + "a_scale": 1.0, "traffic_seed": 4000, "traffic_ratio": 4.0, + "traffic_calibration_examples": 64, "learn_P": 1, + "eta_P": 0.1, "predictor_warmup_steps": 1, + "predictor_every": 0, "alignment_probe": 32, + } + for key, value in expected.items(): + if record["args"].get(key) != value: + raise ValueError(f"D3 {key} drift") + if record["provenance"]["git_tracked_dirty"]: + raise ValueError("tracked-dirty D3 record") + if record["split"]["validation_index_sha256"] != SPLIT_HASH: + raise ValueError("D3 split drift") + evaluation = record["evaluation_protocol"] + if (evaluation["validation_evaluations"] != 1 + or evaluation["test_evaluations"] != 0 + or evaluation["test_used_for_selection"] is not False): + raise ValueError("D3 evaluation boundary drift") + if record.get("calibration_metric_space") != ( + "reciprocal_local_activity_products_with_mixed_apical_traffic"): + raise ValueError("D3 metric-space drift") + warmup = record.get("predictor_warmup", {}) + if (warmup.get("mode") != "closed_form" + or warmup.get("steps") != 1 + or warmup.get("examples") != 64 + or warmup.get("instruction_present") is not False + or warmup.get("task_loader_state_restored") is not True + or warmup.get("reuses_traffic_calibration_forward") is not True): + raise ValueError("D3 neutral slow-fit invariant failed") + epochs = record["epochs"] + if len(epochs) != 200 or any(row["epoch"] != index + 1 + for index, row in enumerate(epochs)): + raise ValueError("D3 epoch trajectory is incomplete") + projection = [row.get("neutral_projection") for row in epochs] + mixed = [row.get("mixed_apical") for row in epochs] + tracking = [row.get("feedback_tracking") for row in epochs] + if any(value is None for value in projection + mixed + tracking): + raise ValueError("D3 audited trajectory is incomplete") + + final = record["final"] + diagnostics = record["diagnostics"] + accuracy = float(final["accuracy"]) + early = float(diagnostics["early_third_mean"]) + final_feedback = float(diagnostics["mean_feedback_forward_cosine"]) + late_feedback = sum(float(value["mean_feedback_forward_cosine"]) + for value in tracking[150:]) / 50 + maximum_signal_ratio_error = max(abs( + float(values["teaching_rms"]) + / max(float(values["instruction_rms"]), 1e-30) - 1.0) + for values in mixed) + maximum_post_ratio = max(float(value[ + "maximum_post_projection_traffic_rms_ratio"]) + for value in projection) + maximum_post_slope = max(float(value[ + "maximum_absolute_post_projection_soma_slope"]) + for value in projection) + maximum_pre_ratio = max(float(value[ + "maximum_pre_projection_traffic_rms_ratio"]) + for value in projection) + instruction_observations = sum(int(value["instruction_observations"]) + for value in projection) + initial_ratio_error = max(abs(float(value) - 4.0) for value in + record["traffic_calibration"][ + "realized_traffic_instruction_rms_ratio"]) + work = record["work"] + counters = record["counters"] + mac_ratio = float(work["total_macs_estimate"]) / bp_macs + all_finite = bool(final["finite"]) and all( + math.isfinite(value) for value in numeric_leaves({ + "final": final, "epochs": epochs, "diagnostics": diagnostics, + "warmup": warmup, "traffic": record["traffic_calibration"], + "work": work, + })) + + checks = { + "record_trajectory_and_diagnostics_finite": all_finite, + "accuracy_at_least_0p89": accuracy >= 0.89, + "within_1p5_points_of_bp": accuracy >= bp_accuracy - 0.015, + "within_1p5_points_of_clean_kp": ( + accuracy >= KP_FULL_ACCURACY - 0.015), + "early_alignment_at_least_0p90": early >= 0.90, + "final_feedback_cosine_at_least_0p98": final_feedback >= 0.98, + "epoch151_to200_feedback_cosine_at_least_0p97": late_feedback >= 0.97, + "used_instruction_rms_ratio_within_1e_minus_4": ( + maximum_signal_ratio_error <= 1e-4), + "post_projection_traffic_ratio_at_most_1e_minus_5": ( + maximum_post_ratio <= 1e-5), + "post_projection_soma_slope_at_most_1e_minus_5": ( + maximum_post_slope <= 1e-5), + "zero_instruction_observations_in_fast_fit": ( + instruction_observations == 0), + "one_frozen_64_example_slow_fit": ( + counters["predictor_update_examples"] == 64 + and counters["predictor_warmup_examples"] == 0 + and record["args"]["predictor_every"] == 0), + "projection_observes_each_ordinary_example": ( + counters["neutral_projection_examples"] + == counters["ordinary_examples"] == 9_000_000), + "initial_traffic_ratio_error_at_most_1e_minus_5": ( + initial_ratio_error <= 1e-5), + "zero_task_loss_queries": work["logical_batch_loss_queries"] == 0, + "macs_at_most_1p34x_bp": mac_ratio <= 1.34, + "elementwise_and_neutral_cost_reported": ( + work["elementwise_operations_estimate"] > 0 + and work["neutral_projection_observations"] == 9_000_000), + "peak_allocated_memory_at_most_2p5_gib": ( + record["hardware"]["peak_memory_allocated_bytes"] + <= int(2.5 * 1024 ** 3)), + "one_validation_and_zero_test_evaluations": ( + evaluation["validation_evaluations"] == 1 + and evaluation["test_evaluations"] == 0), + } + passed = all(checks.values()) + output = { + "protocol": "kp_dynamic_neutral_projection_full_v1", + "status": "passed" if passed else "failed", + "checks": checks, + "metrics": { + "accuracy": accuracy, + "loss": float(final["loss"]), + "bp_accuracy": bp_accuracy, + "clean_kp_accuracy": KP_FULL_ACCURACY, + "early_third_alignment": early, + "final_feedback_forward_cosine": final_feedback, + "epoch151_to200_feedback_forward_cosine": late_feedback, + "maximum_used_instruction_rms_ratio_error": ( + maximum_signal_ratio_error), + "maximum_pre_projection_traffic_rms_ratio": maximum_pre_ratio, + "maximum_post_projection_traffic_rms_ratio": maximum_post_ratio, + "maximum_post_projection_soma_slope": maximum_post_slope, + "total_macs": int(work["total_macs_estimate"]), + "bp_total_macs": bp_macs, + "mac_ratio_to_bp": mac_ratio, + "elementwise_operations_estimate": int( + work["elementwise_operations_estimate"]), + "peak_memory_allocated_bytes": int(record["hardware"][ + "peak_memory_allocated_bytes"]), + "wall_s": float(record["timing"]["total_timed_wall_s"]), + "source_commit": record["provenance"]["git_commit"], + }, + "independent_confirmation_opened": passed, + "confirmation_test_seeds_touched": False, + "review_score_before": 5, + "review_score_after": 6 if passed else 5, + "score_change_rule": ( + "a fully passed frozen near-BP standard-ResNet innovation endpoint " + "resolves the primary accept objection; confirmation remains open"), + } + os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True) + with open(args.out, "w") as handle: + json.dump(output, handle, indent=2, sort_keys=True) + handle.write("\n") + print(json.dumps(output, indent=2)) + + +if __name__ == "__main__": + main() + |