Experiment Logs

This directory holds dated records of experiments, evaluations, and design decisions for the Oczy project.

Each log should be timestamped and include:

• hypothesis or goal
• method
• results
• conclusion / next steps

Curated campaign logs

When a batch of experiments is executed together as a remote campaign (multiple jobs across kaggle/colab providers from a shared source commit), a single curated campaign log consolidates the evidence. The filename format is YYYY-MM-DD_campaign_<short-commit>.md.

A curated campaign log must include:

• Goal — what the campaign set out to adjudicate
• Immutable source commits and providers — commit hashes, provider(s), CPU/GPU contract
• Lawful scope — statement that no metrics, thresholds, baselines, or manifests were modified
• Concurrency — batch sequence and execution model (kernels, notebooks, cross-batch concurrency)
• Per-run results — per-experiment outcome table with primary metrics, seeds, and provider
• Seed distributions — cross-seed variance data for multi-seed experiments (e.g. Exp06, R18)
• Nulls and refutations — preserved as prominently as positives; metricless NULL distinguished from scientific NULL
• Infrastructure blockers — experiments that did not run, with reason (not a scientific verdict)
• Non-runnable inventory — which catalogued experiments did not produce results
• Artifact provenance paths — report paths relative to the campaign working directory
• Infrastructure fixes — what broke and was fixed between retry batches
• Next steps — concrete follow-up actions

Do not copy transient raw logs (/tmp/ execution summaries, *.json batch files) into the curated log. Reference them by relative path. The curated log is the durable record; the raw summaries are ephemeral.

The ledger (LEDGER.md) classifies each campaign log as a single VALID row pointing to the curated file, replacing per-experiment row references.