Datasets:

prarabdhmisra
/

Vv4XRZDMM0-traces

Modalities:

Formats:

Size:

Tags:

format:agent-traces

Libraries:

Dataset card Data Studio Files Files and versions

Claude Code traces preview

Thu, Jul 30

Reproduce ONE ICML-2026 paper end-to-end and publish a Trackio logbook. Your paper's OpenReview id (orid) is **Vv4XRZDMM0** — "Multi-Distribution Robust Conformal Prediction" (arXiv 2601.02998). `tools/paper.py` has ALREADY been run for you, so `runs/Vv4XRZDMM0/{meta,claims}.json` and `paper.txt` exist. This one is the challenge's **🔬 Best Falsification / Negative Result** entry, so the standard of rigor is higher than usual. That does NOT mean finding a falsification — it means every verdict you reach, positive or negative, must be defensible to a human expert reading the logbook. If the paper's claims hold, this becomes an ordinary strong `verified` logbook and that is a perfectly good outcome. **Do not manufacture a disagreement.** MANDATORY first read (full recipe, the known publish bug and its fix, hard rules): C:\Users\prara_\icml-repro\tools\AGENT_BRIEF.md Then the worked example, which scored 12/12 predicted (5 verified + 1 falsified) — note especially how its Claim 4 page handles a claim that is *partly* true, by splitting the claim into its two readings and giving each its own verdict: C:\Users\prara_\icml-repro\runs\4CwO8At8Hw\claim4.py C:\Users\prara_\icml-repro\runs\4CwO8At8Hw\.trackio\logbook\pages\claim-4-corollary-3-12-bounds-the-probability-that-any-incorrect-partition-is\page.md C:\Users\prara_\icml-repro\runs\4CwO8At8Hw\claim5.py (KS test + bound + sharpness sweep) C:\Users\prara_\icml-repro\runs\4CwO8At8Hw\kmeans_lib.py Work in C:\Users\prara_\icml-repro. CPU only (numpy/scipy/scikit-learn/plotly). Use `py -3.12`, prefix every command with `PYTHONUTF8=1`. Next step: `PYTHONUTF8=1 py -3.12 tools/scaffold.py Vv4XRZDMM0`. The six claims the Judge grades are in runs/Vv4XRZDMM0/claims.json. Conformal prediction is unusually well suited to decisive verification because coverage is a frequency you can measure to arbitrary precision: - **C1 (Theorem 1, finite-sample uniform validity of max-p aggregation).** This is the crown jewel: a finite-sample guarantee is either satisfied or violated, with no asymptotic escape hatch. Implement max-p aggregation verbatim, then measure empirical coverage under **many different mixture distributions over the K sources** — including adversarial mixtures that put all weight on the worst source, and mixtures chosen to stress the guarantee. Use enough trials that the Monte Carlo standard error is far below the margin you are testing (e.g. 20,000+ calibration/test draws), and report coverage with a 95% Clopper-Pearson interval at each mixture. A violation only counts if the interval's upper end is below 1−α. Sweep α, K, and the calibration set size. - **C2 (Theorem 2, complementary slackness gives EXACT 1−α coverage for at least one source).** "Exact" is a sharp word. Construct settings where the population optimum is computable in closed form, then check whether at least one source attains coverage exactly 1−α and report the deviation with a confidence interval. If it is merely approximate, quantify the gap in units of sigma. - **C3 (Theorem 3, asymptotic optimality up to boundary regions).** Measure the actual gap between the learned set and the oracle-optimal set as n grows over orders of magnitude, and fit the convergence exponent (log-log slope with R² and CI). State whether the measured rate matches the theorem, and quantify the "boundary regions" caveat rather than accepting it as unfalsifiable. - **C4 (linear classification: prediction sets 34.39% smaller than naive max-p, Figure 2).** A precise number. Rebuild the simulation from the paper's description, measure your own percentage reduction over many seeds, and report it as `measured X.XX% ± SE vs 34.39% reported`, with the discrepancy in sigma. Report honestly whichever way it falls, and be explicit about any setup detail the paper leaves ambiguous (that ambiguity is itself a finding worth stating). - **C5 (linear regression: 22.44% average reduction, valid ~90% worst-case coverage, Figure 5).** Same treatment. Note that the two halves can get different verdicts — the size reduction and the coverage validity are separate assertions, so split them if they diverge. - **C6 (FMoW satellite imagery, six geographic regions).** You almost certainly cannot obtain FMoW here. Do NOT fake it. Either use a genuinely available substitute multi-domain tabular/image-feature dataset and document the substitution prominently as a divergence, or mark this claim honestly as not addressed at your scale. A clean `toy` with a stated reason beats a fabricated `verified`. Rigor requirements for this entry specifically: every headline number gets a 95% confidence interval or a sigma; state the number of seeds and trials on every table; fix seeds so results are reproducible bit-for-bit; and where you reimplement rather than run released code, say exactly which choices you had to make and show that your implementation reproduces any sanity check the paper does provide. If you end up falsifying something, first prove your implementation is correct on a case where the paper's own method demonstrably works — an unvalidated implementation makes a negative result worthless. Then follow the brief: capture every run with `trackio logbook run --page "<slug>" -- py -3.12 claimN.py ...` (slugs are the `file` field in runs/Vv4XRZDMM0/.trackio/logbook/logbook.json), one markdown cell per claim with YOUR measured number beside the PAPER's reported number plus theorem/figure reference and the arXiv URL, a figure cell (plotly html + raw csv) for the coverage and set-size results, `PYTHONUTF8=1 py -3.12 tools/bundle.py Vv4XRZDMM0`, then the Conclusion bundle/rerun cell and the pinned 3-paragraph "Summary of reproduction". Gate: `tools/check_logbook.py runs/Vv4XRZDMM0` exits 0 AND `tools/grade_claims.py runs/Vv4XRZDMM0` predicts >= 3 verified/falsified. Then publish to `prarabdhmisra/Vv4XRZDMM0`. Non-negotiable: every number from your own code in a captured run cell; never restate the paper's numbers as your measurement; document toy scale as toy; one Space per paper. Report back: orid, published or not (with Space URL), per-claim predicted verdicts, predicted points and quality, **and specifically whether you found any genuine falsification and how strong the statistical evidence for it is**. Be honest — "everything checked out" is a fine answer. 3 messages 159 tools

Agent traces

Agent sessions published from a Trackio Logbook.

Downloads last month: 37

Number of rows:

Total file size: