| |
| """Render the rubric library + experiment metrics as one self-contained page. |
| |
| python3 make_browser.py --out browser.html |
| |
| Everything inlined -- no server, no fetches: open the file. Regenerate after |
| any pipeline run; the page states which corpus and truth source produced each |
| number so a stale copy cannot masquerade as current. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import html |
| import json |
| import re |
| import time |
| from pathlib import Path |
|
|
| HERE = Path(__file__).resolve().parent |
|
|
| CSS = """ |
| body{font-family:-apple-system,Segoe UI,Helvetica,Arial,sans-serif;margin:0;background:#0f1115;color:#e6e6e6} |
| header{padding:24px 32px;background:#161a22;border-bottom:1px solid #2a2f3a} |
| h1{margin:0 0 6px;font-size:22px} .sub{color:#9aa4b2;font-size:13px} |
| main{max-width:1100px;margin:0 auto;padding:24px 32px} |
| h2{font-size:17px;border-bottom:1px solid #2a2f3a;padding-bottom:6px;margin-top:36px} |
| table{border-collapse:collapse;width:100%;font-size:13px;margin:12px 0} |
| th,td{border:1px solid #2a2f3a;padding:6px 10px;text-align:left} |
| th{background:#161a22} .good{color:#7ee08a} .bad{color:#ff7d7d} .warn{color:#ffd479} |
| .note{color:#9aa4b2;font-size:12px;margin:4px 0 16px} |
| details{border:1px solid #2a2f3a;border-radius:6px;margin:8px 0;background:#141822} |
| summary{cursor:pointer;padding:10px 14px;font-size:14px} |
| summary .cls{display:inline-block;min-width:170px;color:#8ab4ff} |
| summary .prov{float:right;color:#9aa4b2;font-size:12px} |
| .body{padding:0 18px 14px;font-size:13px;line-height:1.5} |
| .body pre{background:#0b0d12;border:1px solid #232838;border-radius:4px;padding:10px;overflow-x:auto;font-size:12px} |
| .body code{color:#ffd479} |
| .filter{margin:12px 0} |
| .filter button{background:#161a22;color:#e6e6e6;border:1px solid #2a2f3a;border-radius:4px;padding:4px 10px;margin-right:6px;cursor:pointer;font-size:12px} |
| .filter button.on{background:#2a3550;border-color:#8ab4ff} |
| """ |
|
|
| JS = """ |
| function filt(cls,btn){ |
| document.querySelectorAll('.filter button').forEach(b=>b.classList.remove('on')); |
| btn.classList.add('on'); |
| document.querySelectorAll('details.rubric').forEach(d=>{ |
| d.style.display=(cls==='all'||d.dataset.cls===cls)?'':'none';}); |
| } |
| """ |
|
|
|
|
| def md_to_html(text: str) -> str: |
| """Just enough markdown for these rubric files.""" |
| out, in_code = [], False |
| for line in text.splitlines(): |
| if line.strip().startswith("```"): |
| out.append("</pre>" if in_code else "<pre>") |
| in_code = not in_code |
| continue |
| if in_code: |
| out.append(html.escape(line)) |
| continue |
| esc = html.escape(line) |
| esc = re.sub(r"\*\*(.+?)\*\*", r"<b>\1</b>", esc) |
| esc = re.sub(r"`([^`]+)`", r"<code>\1</code>", esc) |
| if esc.startswith("### "): |
| esc = f"<h3>{esc[4:]}</h3>" |
| elif esc.strip().startswith("- "): |
| esc = f"<div style='margin-left:1em'>• {esc.strip()[2:]}</div>" |
| elif re.match(r"^\s+\d+\.", esc): |
| esc = f"<div style='margin-left:2em'>{esc.strip()}</div>" |
| out.append(esc) |
| return "\n".join(out) |
|
|
|
|
| def load_json(path: Path): |
| try: |
| return json.loads(path.read_text()) |
| except (OSError, json.JSONDecodeError): |
| return None |
|
|
|
|
| def fmt(v, pct=False): |
| if v is None: |
| return "—" |
| return f"{v:.1%}" if pct else f"{v:.3f}" |
|
|
|
|
| def metrics_section() -> str: |
| pr = load_json(HERE / "data" / "pr_results.json") |
| assist = load_json(HERE / "data" / "assist_results.json") |
| cf = load_json(HERE / "data" / "counterfactual.json") |
| truth_rep = load_json(HERE / "data" / "truth_report.json") |
| parts = ["<h2>Experiment metrics</h2>"] |
|
|
| if pr: |
| parts.append(f"<div class=note>truth source: {html.escape(pr.get('truth_source','?'))} " |
| f"· {pr.get('decisions','?')} blind decisions across " |
| f"{len(pr.get('eval_competitions',[]))} held-out competitions</div>") |
| rows = ["<table><tr><th>family</th><th>precision</th><th>95% CI</th>" |
| "<th>recall</th><th>95% CI</th><th>tp/fp/fn</th></tr>"] |
| for fam in ("conceptual", "control_generic", "control_shuffled", |
| "control_conceptual"): |
| m = (pr.get("by_family") or {}).get(fam) |
| if not m: |
| continue |
| cls = "good" if fam == "conceptual" else "warn" |
| rows.append( |
| f"<tr><td class={cls}>{fam}</td>" |
| f"<td>{fmt(m.get('precision'))}</td><td>{m.get('precision_ci95') or '—'}</td>" |
| f"<td>{fmt(m.get('recall'))}</td><td>{m.get('recall_ci95') or '—'}</td>" |
| f"<td>{m.get('tp')}/{m.get('fp')}/{m.get('fn')}</td></tr>") |
| rows.append("</table>") |
| parts.append("".join(rows)) |
| parts.append("<div class=note>Controls are the floor: vague criticism and " |
| "domain-swapped rubrics score zero precision; whatever the real " |
| "rubrics score above that is earned signal.</div>") |
| if truth_rep: |
| parts.append(f"<div class=note>ground-truth annotator agreement: " |
| f"{fmt(truth_rep.get('annotator_agreement_per_group'), pct=True)} per group of " |
| f"{truth_rep.get('annotation_batch','?')} (reference protocol reports 97%)</div>") |
|
|
| if assist: |
| p = assist.get("paired") or {} |
| parts.append("<h2>Utility: does giving the rubrics to a coding agent help?</h2>") |
| parts.append( |
| f"<table><tr><th>paired states</th><th>guided better</th><th>revise better</th>" |
| f"<th>ties</th><th>mean advantage</th><th>median advantage</th></tr>" |
| f"<tr><td>{p.get('paired_states','—')}</td>" |
| f"<td class=good>{p.get('guided_better','—')}</td>" |
| f"<td class=bad>{p.get('revise_better','—')}</td>" |
| f"<td>{p.get('ties','—')}</td>" |
| f"<td>{fmt(p.get('mean_paired_advantage'), pct=True)}</td>" |
| f"<td>{fmt(p.get('median_paired_advantage'), pct=True)}</td></tr></table>") |
| parts.append("<div class=note>guided = revision with retrieved rubrics; revise = " |
| "same revision without them (the control). The rubrics' contribution " |
| "is the difference, paired on the same starting program.</div>") |
|
|
| if cf: |
| parts.append("<h2>Causal: does removing the flagged pattern improve the score?</h2>") |
| parts.append( |
| f"<table><tr><th>fires measured</th><th>improved</th><th>causal precision</th>" |
| f"<th>mean improvement</th><th>outcomes</th></tr>" |
| f"<tr><td>{cf.get('measured')}</td><td>{cf.get('helped')}</td>" |
| f"<td>{fmt(cf.get('causal_precision'))}</td>" |
| f"<td>{cf.get('mean_improvement')}</td>" |
| f"<td>{html.escape(json.dumps(cf.get('outcomes', {})))}</td></tr></table>") |
| parts.append("<div class=note>Patch out exactly what the rubric flagged, re-run, " |
| "re-grade against the official answers. Failed edits are excluded " |
| "from the denominator; they are facts about the editor.</div>") |
| return "\n".join(parts) |
|
|
|
|
| def rubric_section() -> str: |
| files = sorted((HERE / "out" / "conceptual").glob("*.md")) |
| classes = sorted({f.name.split("__")[0].upper() for f in files}) |
| buttons = ["<div class=filter><button class=on onclick=\"filt('all',this)\">all " |
| f"({len(files)})</button>"] |
| for cls in classes: |
| n = sum(1 for f in files if f.name.split("__")[0].upper() == cls) |
| buttons.append(f"<button onclick=\"filt('{cls}',this)\">{cls} ({n})</button>") |
| buttons.append("</div>") |
|
|
| cards = [] |
| lib = {} |
| for row in (load_json(HERE / "data" / "conceptual_applicable.jsonl") or []): |
| pass |
| try: |
| for line in (HERE / "data" / "conceptual_applicable.jsonl").read_text().splitlines(): |
| if line.strip(): |
| r = json.loads(line) |
| lib[r["id"]] = r |
| except OSError: |
| pass |
|
|
| for f in files: |
| cls = f.name.split("__")[0].upper() |
| text = f.read_text() |
| title = re.search(r"### P1 — (.+)", text) |
| cand = lib.get(f.stem.rsplit("__", 1)[-1], {}) |
| prov = cand.get("provenance", "") |
| eff = cand.get("measured_effect") |
| prov_txt = prov + (f" · effect {eff:+.4f}" if eff is not None else "") |
| cards.append( |
| f"<details class=rubric data-cls='{cls}'>" |
| f"<summary><span class=cls>{cls} · {html.escape(f.name.split('__')[1])}</span> " |
| f"{html.escape(title.group(1) if title else f.name)}" |
| f"<span class=prov>{prov_txt}</span></summary>" |
| f"<div class=body>{md_to_html(text)}</div></details>") |
| return ("<h2>Rubric library (conceptual / performance errors)</h2>" |
| + "".join(buttons) + "\n" + "\n".join(cards)) |
|
|
|
|
| def main() -> int: |
| ap = argparse.ArgumentParser(description=__doc__) |
| ap.add_argument("--out", default=str(HERE / "browser.html")) |
| args = ap.parse_args() |
|
|
| n = len(list((HERE / "out" / "conceptual").glob("*.md"))) |
| page = f"""<!doctype html><html><head><meta charset=utf-8> |
| <title>rubric_gen — conceptual rubric library</title><style>{CSS}</style> |
| <script>{JS}</script></head><body> |
| <header><h1>Conceptual rubric library & experiment metrics</h1> |
| <div class=sub>{n} rubrics · split rubric-gen-full75-v1 (15 train / 60 eval) |
| · generated {time.strftime('%F %T UTC', time.gmtime())} · regenerate with |
| <code>python3 make_browser.py</code></div></header> |
| <main>{metrics_section()}{rubric_section()}</main></body></html>""" |
| Path(args.out).write_text(page) |
| print(f"{n} rubrics + metrics -> {args.out}") |
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|
