#!/usr/bin/env python3
"""Present retained protocol-check measurements as a supplement to the study.
Reads the existing pilot.html archive and writes supplement.html. Embedded case
data and explorer JavaScript are preserved exactly. No model or network calls.
"""
from __future__ import annotations
import argparse
import hashlib
import json
from pathlib import Path
import re
OUT = Path(__file__).resolve().parent.parent
def one(pattern, replacement, text):
result, count = re.subn(pattern, lambda _: replacement, text, count=1, flags=re.S)
if count != 1:
raise ValueError(f"Expected presentation element not found: {pattern}")
return result
def build(source=OUT / "pilot.html", destination=OUT / "supplement.html"):
source, destination = Path(source), Path(destination)
if source.resolve() == destination.resolve():
raise ValueError("The source archive must remain unchanged")
original = source.read_text()
evidence_match = re.search(r'', original, re.S)
if not evidence_match:
raise ValueError("The archived evidence payload is missing")
evidence = json.loads(evidence_match.group(1))
if len(evidence["cases"]) != 204 or len(evidence["records"]) != 441:
raise ValueError("Unexpected supplementary archive: expected 204 cases and 441 calls")
measured = sum(r["phase"] == "measured" for r in evidence["records"])
if measured != 432:
raise ValueError("Unexpected supplementary measured-call count")
# Only the presentation preceding the scripts is edited. The original JSON
# payload and all explorer/filter/pagination behavior remain byte-identical.
script_start = original.index(""
page = re.sub(r"]*>.*?", protect_pre, page, flags=re.S)
page = one(r".*?", "Supplementary measurements | Luna and Jev study", page)
page = one(r'',
'', page)
page = one(r".*?", """
Luna and Jev study · OpenRouterSupplementary measurements
Protocol checks and retained call-level evidence
Supplementary measurements
The study’s earlier protocol checks, with every input, reference target, response, timing and reported charge retained.
Measured: 2 October 2026 · 08:42–08:50 UTC / Transport: OpenRouter Configurations: GPT-6 Luna, reasoning disabled · Jev 1.13 Archive: 204 measured case IDs · 432 measured calls · 441 calls including setup, warmups and cache priming
""", page)
page = one(r'
.*?
\s*', """
Supplementary scope
These records document the study’s short-input and prefix-cache protocol checks. They are retained for transparency and inspection and are kept separate from the primary quality, timing and cache estimates in the study report.
The 180 short-input cases comprise 60 policy decisions, 60 BoolQ questions and 60 exact-probability problems, each sent to both models. A further 24 equipment-policy cases were sent to three cache-study arms. These account for 432 measured calls; four warmups, one cold cache prime and four setup probes bring the complete record to 441 calls.
Shared data and protocols. The 60 BoolQ rows also occur in the primary full split, some policy inputs overlap, and the synthetic task families and output contract are shared. Primary estimates use fresh recorded calls, but these datasets and designs are not independent. The supplementary observations are not pooled into the primary estimates.
""", page)
page = one(r'.*?', """
1Role of these measurements
The checks exercised the common probability-of-yes answer schema, the two OpenRouter request adapters, sequential timing and billing capture, and explicit prefix caching. The tables below describe their observed outcomes and preserve the original analysis choices.
The study report provides the primary estimates and overall interpretation. This supplement supplies the earlier call-level evidence and its limitations. Original run labels, case IDs, filenames and source hashes remain unchanged so records can be reconciled with their archived artifacts.
""", page)
discussion = re.search(r'.*?', page, re.S).group(0)
limits = re.search(r'.*?', discussion, re.S).group(0)
limits = limits.replace("per main task", "per short-input task").replace("this pilot", "these supplementary measurements")
limits = limits.replace("Primary results should be read as agreement with published annotations.",
"These BoolQ scores should be read as agreement with published annotations.")
page = one(r'.*?',
'
6Scope and interpretation limits
'
'
These small protocol checks support inspection of the recorded configurations. Their descriptive contrasts and intervals are supplementary evidence; the study’s primary estimates are reported separately. Shared source rows and task templates limit independence between the two sets of measurements.
'
+ limits + '', page)
replacements = {
"Experimental report 02 October 2026": "Study supplement 02 October 2026",
'