Revalidate the 30-file training cap (keep 30; 50 for agent-heavy repos)

[Byggarepop]

Re-runs the original EF Core holdout validation, this time sweeping maxTransactionSize — the 30-file cap was part of the fixed prototype configuration, never independently swept, and agentic workflows have shifted commit sizes since.

What's here

• eval/Unwritten.Eval — replay harness against the real engine: trains one index per cap {20, 30, 50, 75, 100, 200} on the same chronological stream, replays the last 2,000 held-out commits (evaluate → ingest, as in real use), scores fill-within-10-commits, and isolates marginal rules (holes a raised cap flags that cap 30 doesn't).
• docs/eval/cap-revalidation-2026-07.md — the report, with raw CSVs alongside.

Findings

• Harness sanity check: EF Core at cap 30 / floor 0.7 fills at 67.6% vs the prototype's 66.5% — the protocol replicates.
• Commit sizes did shift in agent-era repos (cline: 1.2% → 5.9% of commits excluded by cap 30; codex: 3.9% → 7.0%) but not in EF Core.
• Marginal rules from raising the cap: beat baseline in cline at cap 50 (48.6% vs 39.7% fill), are noise in EF Core (5–10% vs 55.8%). Beyond 50, returns fade everywhere.

Verdict

Keep the default at 30 — it must be safe for both populations. "maxTransactionSize": 50 is now an evidence-backed override for agent-heavy repos. Kept in the repo for reference; no product code changes.

🤖 Generated with Claude Code