snf_experiments / method_v2 /evidence /code /suite /null_commitment.py
CharlesZhang-USTC's picture
Add Control v2 incremental experiment results (evidence, visuals, 42 fixed samples)
a64600b verified
Raw History Blame Contribute Delete
23.9 kB
"""Independent-token null for the interaction of control and commitment timing.
This is an analytical/synthetic experiment, not a generative-model evaluation.
Every unforced token independently matches its reference with probability b;
there is no context dependence and no learned model. If the reference is copied
at commitment, E[match_i] = b + (1-b) * g(alpha, p_commit_i). Without that copy,
showing the reference as context has no effect on this independent predictor.
Run: python -m suite.null_commitment --output artifacts/method_v2_local/null
"""
from __future__ import annotations
import argparse
import csv
import json
import math
import time
from datetime import datetime, timezone
from pathlib import Path
import numpy as np
ORDERS = ("target_first", "target_last", "random", "balanced")
SCHEDULES = ("duration", "amplitude")
def target_support(tokens: int, fraction: float = 0.25) -> np.ndarray:
if tokens < 8 or not 0 < fraction < 1:
raise ValueError("tokens >= 8 and 0 < target fraction < 1 are required")
count = max(1, min(tokens - 1, int(tokens * fraction)))
start = (tokens - count) // 2
mask = np.zeros(tokens, dtype=bool)
mask[start:start + count] = True
return mask
def fixed_order(mask: np.ndarray, name: str, order_seed: int = 5678) -> np.ndarray:
"""Fixed global priorities; balanced distributes target ranks uniformly."""
mask = np.asarray(mask, dtype=bool)
priority = np.random.default_rng(order_seed).random(len(mask))
ranked = np.argsort(priority, kind="stable")
target = ranked[mask[ranked]]
outside = ranked[~mask[ranked]]
if name == "target_first":
return np.concatenate((target, outside))
if name == "target_last":
return np.concatenate((outside, target))
if name == "random":
return ranked
if name != "balanced":
raise ValueError(f"Unknown order: {name}")
result = np.empty(len(mask), dtype=int)
ti = oi = 0
for rank in range(len(mask)):
# Exactly floor(k * target_fraction) target positions in the first k.
wanted = (rank + 1) * len(target) // len(mask)
if wanted > ti:
result[rank] = target[ti]
ti += 1
else:
result[rank] = outside[oi]
oi += 1
return result
def commitment_steps(permutation: np.ndarray, steps: int) -> np.ndarray:
"""Same cosine cumulative reveal and final-step completion as masked_decode."""
if steps < 2:
raise ValueError("steps >= 2 are required")
n = len(permutation)
if not np.array_equal(np.sort(permutation), np.arange(n)):
raise ValueError("permutation must contain each global position once")
committed = np.full(n, -1, dtype=int)
current = 0
for step in range(steps):
reveal = 1 - math.cos((step + 1) / steps * math.pi / 2)
wanted = n if step == steps - 1 else min(n, max(current, int(n * reveal)))
committed[permutation[current:wanted]] = step
current = wanted
return committed
def control_gate(alpha, progress, steps: int, schedule: str) -> np.ndarray:
a = np.asarray(alpha, dtype=float)
p = np.asarray(progress, dtype=float)
if np.any((a < 0) | (a > 1)):
raise ValueError("alpha must lie in [0, 1]")
if schedule == "duration":
return np.clip((a - p) * steps, 0, 1)
if schedule == "amplitude":
return a * np.clip((1 - p) * steps, 0, 1)
raise ValueError(f"Unknown schedule: {schedule}")
def exact_curve(alpha, commit: np.ndarray, mask: np.ndarray, steps: int,
baseline_match: float = 0.15, schedule: str = "duration",
reference_copy: bool = True) -> dict:
"""Integrates both independent token sampling and reference-copy uniforms."""
if not 0 < baseline_match < 1:
raise ValueError("baseline match probability must lie strictly in (0,1)")
a = np.atleast_1d(alpha).astype(float)
p = 1 - np.cos(np.asarray(commit) / steps * np.pi / 2)
gate = control_gate(a[:, None], p[None, :], steps, schedule)
gate[:, ~mask] = 0
if not reference_copy:
gate = np.zeros_like(gate)
match = baseline_match + (1 - baseline_match) * gate
return {
"target_match": match[:, mask].mean(axis=1),
"outside_match": match[:, ~mask].mean(axis=1),
"direct_copy_fraction_target": gate[:, mask].mean(axis=1),
"direct_copy_count_target": gate[:, mask].sum(axis=1),
"direct_copy_count_outside": gate[:, ~mask].sum(axis=1),
}
def response_statistics(alpha, response) -> dict:
a = np.asarray(alpha, dtype=float)
y = np.asarray(response, dtype=float)
gain = float(y[-1] - y[0])
result = {"endpoint_gain": gain, "alpha10": None, "alpha50": None,
"alpha90": None, "width80": None}
if gain <= 1e-12:
return result
z = np.maximum.accumulate((y - y[0]) / gain)
for q, name in ((0.1, "alpha10"), (0.5, "alpha50"), (0.9, "alpha90")):
result[name] = float(np.interp(q, z, a))
result["width80"] = result["alpha90"] - result["alpha10"]
return result
def oracle_remap(levels, commit: np.ndarray, mask: np.ndarray, steps: int,
baseline_match: float, schedule: str = "duration") -> dict:
"""Exact per-curve generalized inverse; illustration, NOT held-out calibration.
Continuous gates make every intermediate expectation reachable, even if the
original alpha response is narrow. Plateaus do not prevent this inversion.
This uses the same curve at evaluation time and provides no generalization
evidence, robustness guarantee, or statement about perceptual intermediate
states of an actual generator.
"""
q = np.asarray(levels, dtype=float)
low, high = np.zeros_like(q), np.ones_like(q)
for _ in range(60):
mid = (low + high) / 2
match = exact_curve(mid, commit, mask, steps, baseline_match, schedule)["target_match"]
normalized = (match - baseline_match) / (1 - baseline_match)
high = np.where(normalized >= q, mid, high)
low = np.where(normalized < q, mid, low)
alpha = (low + high) / 2
alpha[q == 0] = 0
achieved = exact_curve(alpha, commit, mask, steps, baseline_match, schedule)["target_match"]
achieved = (achieved - baseline_match) / (1 - baseline_match)
return {"requested_level": q, "remapped_alpha": alpha,
"achieved_level": achieved,
"max_absolute_error": float(np.max(np.abs(achieved - q))),
"caveat": "Oracle same-curve inversion only; not held-out calibration or learned-model evidence."}
def monte_carlo(alpha, commit: np.ndarray, mask: np.ndarray, steps: int,
baseline_match: float, schedule: str, seeds: int,
seed_start: int = 0, reference_copy: bool = True) -> dict:
"""Paired global-position noise across alpha/orders, independent across seeds."""
if seeds < 2:
raise ValueError("at least two independent Monte Carlo seeds are required")
a = np.asarray(alpha, dtype=float)
p = 1 - np.cos(commit / steps * np.pi / 2)
gates = control_gate(a[:, None], p[None, :], steps, schedule)
gates[:, ~mask] = 0
if not reference_copy:
gates = np.zeros_like(gates)
observations = np.empty((seeds, len(a)), dtype=float)
copy_observations = np.empty_like(observations)
outside_observations = np.empty(seeds, dtype=float)
outside_changed = 0
for si, seed in enumerate(range(seed_start, seed_start + seeds)):
# These streams are paired across configurations, not fitted to outcomes.
base_rng = np.random.default_rng(np.random.SeedSequence([seed, 0]))
ref_rng = np.random.default_rng(np.random.SeedSequence([seed, 1]))
global_uniforms = base_rng.random((steps, len(mask)))
baseline = global_uniforms[commit, np.arange(len(mask))] < baseline_match
copies = ref_rng.random(len(mask))[None, :] < gates
matches = baseline[None, :] | copies
observations[si] = matches[:, mask].mean(axis=1)
copy_observations[si] = copies[:, mask].mean(axis=1)
outside_observations[si] = baseline[~mask].mean()
outside_changed += int(np.count_nonzero(matches[:, ~mask] != baseline[None, ~mask]))
return {"target_match_mean": observations.mean(axis=0),
"target_match_seed_std": observations.std(axis=0, ddof=1),
"target_match_se": observations.std(axis=0, ddof=1) / np.sqrt(seeds),
"direct_copy_fraction_target_mean": copy_observations.mean(axis=0),
"outside_match_mean": float(outside_observations.mean()),
"outside_changed_token_observations": outside_changed}
def _jsonable(value):
if isinstance(value, dict):
return {k: _jsonable(v) for k, v in value.items()}
if isinstance(value, (list, tuple)):
return [_jsonable(v) for v in value]
if isinstance(value, np.ndarray):
return value.tolist()
if isinstance(value, np.generic):
return value.item()
return value
def _write_json(path: Path, data):
path.write_text(json.dumps(_jsonable(data), indent=2, ensure_ascii=False,
allow_nan=False), encoding="utf8")
def _write_csv(path: Path, rows: list[dict]):
with path.open("w", newline="", encoding="utf8") as handle:
writer = csv.DictWriter(handle, fieldnames=list(rows[0]))
writer.writeheader()
writer.writerows(rows)
def make_plots(output: Path, curves: list[dict], oracle: list[dict], config: dict) -> dict:
try:
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
except ImportError as exc:
return {"status": "unavailable", "reason": str(exc)}
colors = {"target_first": "#d95f02", "target_last": "#1b9e77",
"random": "#7570b3", "balanced": "#1f78b4"}
styles = {"target_first": "-", "target_last": "--", "random": "-.", "balanced": ":"}
fig, axes = plt.subplots(2, 2, figsize=(11.2, 7.5))
for c in curves:
if not c["reference_copy"]:
continue
panel = 0 if c["schedule"] == "duration" else 1
ax = axes.flat[panel]
a = np.asarray(c["alpha"])
y = np.asarray(c["exact"]["target_match"])
mc = np.asarray(c["monte_carlo"]["target_match_mean"])
name = c["order"]
ax.plot(a, y, color=colors[name], ls=styles[name], lw=2.2, label=name.replace("_", " "))
ax.scatter(a[::8], mc[::8], color=colors[name], s=13, marker="o", alpha=0.65)
if panel == 0:
p = np.array(c["commitment"]["target_progress"])
axes[1, 0].plot(a, [(p <= x).mean() for x in a], color=colors[name],
ls=styles[name], lw=2, label=name.replace("_", " "))
axes[1, 1].plot(a, mc - y, color=colors[name], ls=styles[name], lw=1.5)
b = config["baseline_match"]
for ax in axes.flat[:2]:
ax.axhline(b, color="#555555", lw=1, ls=":", label="no direct copy (exact)")
ax.set(xlabel="Control parameter alpha", ylabel="Reference-match probability",
xlim=(0, 1), ylim=(max(0, b - .05), 1.025))
ax.grid(alpha=.15)
axes[0, 0].set_title("A. Alpha changes exposure duration")
axes[0, 1].set_title("B. Alpha changes amplitude: all orders coincide")
axes[0, 0].legend(fontsize=8, loc="lower right")
axes[0, 1].legend(fontsize=8, loc="lower right")
axes[1, 0].set(title="C. Target commitment-time distribution", xlabel="Cosine progress",
ylabel="Fraction of target already committed", xlim=(0, 1), ylim=(0, 1.025))
axes[1, 1].axhline(0, color="#555555", lw=1)
axes[1, 1].set(title=f"D. Duration: MC minus exact ({config['seeds']} seeds)",
xlabel="Control parameter alpha", ylabel="Match-probability error", xlim=(0, 1))
for ax in axes.flat[2:]:
ax.grid(alpha=.15)
fig.suptitle(f"Independent-token analytical null: no model coupling ({config['steps']} steps)", fontsize=13)
fig.text(.5, .012, "Lines: exact expectation. Dots: Monte Carlo. Synthetic diagnostic; no generated media or model performance.",
ha="center", fontsize=9)
fig.tight_layout(rect=(0, .035, 1, .955))
files = []
for ext in ("png", "pdf"):
name = f"independent_token_null.{ext}"
fig.savefig(output / name, dpi=190, bbox_inches="tight")
files.append(name)
plt.close(fig)
fig, axes = plt.subplots(1, 2, figsize=(10.2, 3.8))
for rec in oracle:
name = rec["order"]
axes[0].plot(rec["requested_level"], rec["remapped_alpha"], color=colors[name],
ls=styles[name], lw=2, label=name.replace("_", " "))
axes[1].plot(rec["requested_level"], rec["achieved_level"], color=colors[name],
ls=styles[name], lw=2)
axes[0].set(xlabel="Requested normalized response", ylabel="Oracle remapped alpha", title="A. Exact inverse of each duration curve")
axes[1].set(xlabel="Requested normalized response", ylabel="Achieved normalized response", title="B. Every intermediate expectation remains reachable")
axes[0].legend(fontsize=8)
for ax in axes:
ax.set(xlim=(0, 1), ylim=(0, 1))
ax.grid(alpha=.15)
fig.suptitle("Oracle illustration only: same-curve access, no held-out calibration evidence", fontsize=11)
fig.tight_layout(rect=(0, 0, 1, .94))
for ext in ("png", "pdf"):
name = f"oracle_remapping_illustration.{ext}"
fig.savefig(output / name, dpi=190, bbox_inches="tight")
files.append(name)
plt.close(fig)
return {"status": "complete", "matplotlib_version": matplotlib.__version__, "files": files}
def run(output: Path, steps: int = 18, tokens: int = 256, seeds: int = 256,
baseline_match: float = .15, order_seed: int = 5678, seed_start: int = 0,
plots: bool = True) -> dict:
start = time.perf_counter()
output = Path(output)
output.mkdir(parents=True, exist_ok=True)
mask = target_support(tokens)
config = {"experiment": "independent_token_analytical_null", "synthetic": True,
"learned_model_used": False, "steps": steps, "tokens": tokens, "seeds": seeds,
"seed_start": seed_start, "baseline_match": baseline_match, "order_seed": order_seed,
"target_fraction": float(mask.mean()), "target_indices": np.flatnonzero(mask),
"alpha_points": 65, "gap": 1 / steps, "orders": ORDERS, "schedules": SCHEDULES,
"reference_copy_conditions": [True, False], "numpy_version": np.__version__,
"started_at_utc": datetime.now(timezone.utc).isoformat(),
"expectation_over": "independent categorical-match noise and reference-copy uniforms",
"formula": "E[match_i]=b+(1-b)*g(alpha,p_commit_i) with direct copy; b otherwise",
"scope": "Fixed commitment orders; does not simulate adaptive confidence ordering or real-model quality."}
_write_json(output / "config.json", config)
alpha = np.linspace(0, 1, 65)
curves, curve_rows, commitment_rows, oracle_records, oracle_rows = [], [], [], [], []
for order in ORDERS:
commit = commitment_steps(fixed_order(mask, order, order_seed), steps)
progress = 1 - np.cos(commit / steps * np.pi / 2)
target_p = progress[mask]
commitment = {"target_progress_mean": float(target_p.mean()),
"target_progress_std": float(target_p.std()),
"target_progress_min": float(target_p.min()),
"target_progress_max": float(target_p.max()),
"target_progress": target_p, "target_commit_steps": commit[mask],
"newly_committed_per_step": np.bincount(commit, minlength=steps),
"target_committed_per_step": np.bincount(commit[mask], minlength=steps)}
for i in range(tokens):
commitment_rows.append({"order": order, "position": i, "target": int(mask[i]),
"commit_step": int(commit[i]), "commit_progress": float(progress[i])})
for schedule in SCHEDULES:
for copy in (True, False):
exact = exact_curve(alpha, commit, mask, steps, baseline_match, schedule, copy)
mc = monte_carlo(alpha, commit, mask, steps, baseline_match, schedule, seeds, seed_start, copy)
error = mc["target_match_mean"] - exact["target_match"]
se = mc["target_match_se"]
stats = response_statistics(alpha, exact["target_match"])
stats.update({"mc_max_absolute_error": float(np.abs(error).max()),
"mc_rmse": float(np.sqrt(np.mean(error ** 2))),
"mc_max_absolute_zscore": float(np.max(np.abs(error[se > 1e-12] / se[se > 1e-12]))) if np.any(se > 1e-12) else 0.,
"outside_unchanged": mc["outside_changed_token_observations"] == 0})
record = {"order": order, "schedule": schedule, "reference_copy": copy,
"alpha": alpha, "exact": exact, "monte_carlo": mc,
"statistics": stats, "commitment": commitment}
curves.append(record)
for j, a in enumerate(alpha):
curve_rows.append({"order": order, "schedule": schedule, "reference_copy": copy,
"alpha": float(a), "target_match_exact": float(exact["target_match"][j]),
"target_match_mc": float(mc["target_match_mean"][j]),
"target_match_mc_se": float(se[j]),
"target_match_mc_seed_std": float(mc["target_match_seed_std"][j]),
"direct_copy_fraction_target_exact": float(exact["direct_copy_fraction_target"][j]),
"direct_copy_count_target_exact": float(exact["direct_copy_count_target"][j]),
"direct_copy_fraction_target_mc": float(mc["direct_copy_fraction_target_mean"][j]),
"outside_match_exact": float(exact["outside_match"][j]),
"direct_copy_count_outside_exact": float(exact["direct_copy_count_outside"][j])})
oracle = oracle_remap(alpha, commit, mask, steps, baseline_match)
oracle["order"] = order
oracle_records.append(oracle)
for q, a, achieved in zip(oracle["requested_level"], oracle["remapped_alpha"], oracle["achieved_level"]):
oracle_rows.append({"order": order, "requested_level": float(q), "remapped_alpha": float(a),
"achieved_level": float(achieved), "oracle_same_curve_only": True})
_write_json(output / "curves.json", curves)
_write_csv(output / "curves.csv", curve_rows)
_write_csv(output / "commitments.csv", commitment_rows)
_write_json(output / "oracle_remapping.json", oracle_records)
_write_csv(output / "oracle_remapping.csv", oracle_rows)
plot_result = make_plots(output, curves, oracle_records, config) if plots else {"status": "disabled"}
summary = {"synthetic": True, "learned_model_used": False,
"conditions": len(curves), "alpha_points_per_condition": len(alpha),
"elapsed_seconds": time.perf_counter() - start, "device": "CPU (NumPy)",
"model_evaluations": 0, "monte_carlo_seeds_per_condition": seeds,
"max_mc_absolute_error": max(c["statistics"]["mc_max_absolute_error"] for c in curves),
"max_mc_absolute_zscore": max(c["statistics"]["mc_max_absolute_zscore"] for c in curves),
"outside_unchanged_all_conditions": all(c["statistics"]["outside_unchanged"] for c in curves),
"oracle_max_error": max(x["max_absolute_error"] for x in oracle_records),
"plots": plot_result,
"statistics": [{"order": c["order"], "schedule": c["schedule"],
"reference_copy": c["reference_copy"], **c["statistics"]} for c in curves],
"interpretation": [
"Narrow raw-alpha response can arise from direct reference copying and fixed commitment timing with zero context coupling.",
"This null does not quantify how much of any learned-model effect it explains.",
"Amplitude control is linear in expectation for all orders; context-only control is dormant in an independent predictor.",
"Oracle inversion illustrates reachable intermediate expectations, not held-out calibration or perceptual states."]}
_write_json(output / "summary.json", summary)
lines = ["# Independent-token analytical null", "", "This is a synthetic CPU diagnostic. No learned model or generated media is used.", "",
"Every unforced token independently matches its reference with probability b. Context has no effect.",
"With direct reference copying at commitment, E[match_i] = b + (1-b) g(alpha, p_commit_i).", "",
"The two schedules use the same cosine reveal budget and fixed orders. Duration changes the release time; amplitude changes the copy probability over a fixed window.", "",
"`curves.csv/json` contain all exact and independent-seed Monte Carlo results, including no-copy controls and direct-copy doses.",
"`commitments.csv` records every position's commitment time. `config.json` records reproducibility settings; `summary.json` records statistics and CPU wall time.", "",
"`oracle_remapping.*` use exact access to the evaluation curve. They only illustrate that a continuous compressed curve can be reparameterized; they are not evidence of held-out calibration.", "",
"Width is measured in the original alpha coordinate and does not establish missing perceptual intermediate states.", "",
f"Steps: {steps}. Tokens: {tokens}. Target tokens: {int(mask.sum())}. Independent seeds: {seeds}. b: {baseline_match}.",
f"Maximum Monte Carlo absolute error: {summary['max_mc_absolute_error']:.6f}. Outside support unchanged: {summary['outside_unchanged_all_conditions']}.", ""]
(output / "README.md").write_text("\n".join(lines), encoding="utf8")
return summary
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--steps", type=int, default=18)
parser.add_argument("--tokens", type=int, default=256)
parser.add_argument("--seeds", type=int, default=256)
parser.add_argument("--baseline-match", type=float, default=.15)
parser.add_argument("--order-seed", type=int, default=5678)
parser.add_argument("--seed-start", type=int, default=0)
parser.add_argument("--no-plots", action="store_true")
args = parser.parse_args()
summary = run(args.output, args.steps, args.tokens, args.seeds, args.baseline_match,
args.order_seed, args.seed_start, not args.no_plots)
for record in summary["statistics"]:
if record["reference_copy"]:
print(f"{record['schedule']:9s} {record['order']:12s} width80={record['width80']:.6f} alpha50={record['alpha50']:.6f} gain={record['endpoint_gain']:.6f}")
print(f"MC maximum absolute error: {summary['max_mc_absolute_error']:.6f}; max |z|: {summary['max_mc_absolute_zscore']:.3f}")
print(f"Outside unchanged: {summary['outside_unchanged_all_conditions']}; oracle error: {summary['oracle_max_error']:.3g}")
print(f"Saved {summary['conditions']} conditions to {args.output}; CPU wall time {summary['elapsed_seconds']:.2f}s")
if __name__ == "__main__":
main()