Reproduce paper results

SkillMonitoring & ops

Reproduce and compare a forecasting paper's reported experiments in ModernTSF by mapping its protocol to runnable configs and aligned metrics. Use for paper-result replication; not for implementing the model or designing an unrelated benchmark.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Reproduce paper results skill

What this skill tells your AI

The instructions your AI receives, as published by diaugeia/moderntsf in .agents/skills/reproduce-paper-results/SKILL.md and read by ahel’s review.

Require an identified paper and a runnable audited model. Verify the primary paper, supplement, authoritative source revision, and reported tables before spending compute. Route an existing implementation uncertainty to audit-model; when the model is absent, render the paper-to-model task instead of implementing it inside an experiment-reproduction run.

Map the protocol

Build a compact paper-to-repository map covering dataset name and version, split boundaries, scaling, feature mode, lookback and horizons, covariates, loss, optimizer and schedule, batch size, epochs and stopping, seeds, checkpoint selection, metric formula and aggregation, baselines, and hardware-sensitive settings. Label every field aligned, adapted, unknown, or blocked; do not silently fill paper gaps with repository defaults.

Encode aligned settings in dedicated run configs with inheritance. Keep faithful replication separate from controlled adaptations, and inspect the resolved matrix:

uv run tsf inspect --config <paper-run.toml>

Stop for a decision when missing data, licensing, ambiguous metrics, or infeasible compute would materially change the claim.

Execute and compare

Launch only when authorized, using run-experiment for concurrency and resource handling. Use the paper-reproduction Harness round for multi-run work and pass its id through tsf run --round. Preserve raw outputs, resolved configs, environment facts, seeds, and failed runs. Aggregate only compatible cells with analyze-results.

Compare each local value with the matching paper cell using the same metric direction and aggregation. Report paper value, local value, absolute and relative difference, run count, uncertainty when available, and all protocol deviations. Missing or failed cells remain visible.

Success means the experiment is rerunnable and the comparison is traceable; it does not require matching the paper. Describe the outcome as reproduced only when the recorded protocol and results support that claim, otherwise report a partial replication or a blocked attempt.

Signals

GitHub stars
65
Forks
8
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
reproduce-paper-results
Source
github.com/diaugeia/moderntsf