Repository navigation
FAQ
Q: Why is accuracy different between two runs of the same prompt?
A: They likely ran on different surfaces — dataset, seed, or sample size.
Only same-surface comparisons (identical dataset_fingerprint + seed +
n_samples) are meaningful; the site's "Δ vs best" enforces this, and
bootstrap CIs tell you when a gap is just noise.
Q: Why is cost_total_usd $0.00 but the site shows an estimate?
A: OpenRouter usage payloads carry no cost field, so runs record $0.00 from
usage. The site computes deterministic estimates from token counts × verified
per-model prices (src/cost_models.py, a re-export shim for the
llm-dojo-scoring package's cost module), and shows real billed totals when
you rebuild with --openrouter-csv openrouter_activity.csv.
Q: Why does the markdown log look different after render_experiment_log.py?
A: The md is DERIVED whole from the JSONL — it always reflects every record,
in order. Never hand-edit it.
Q: Can I compare a 50-doc run to a 509-doc run? A: No. Deltas across different samples are meaningless (this is exactly how the v0.13.0 "regression" got misread). Compare only same-surface runs, and let the bootstrap delta CI decide significance.
Q: What is --handoff-scope ground_truth?
A: The chained error-propagation ablation: the specialist ALSO extracts the
same docs cued with the ground-truth subtype. scores.ablation splits sorter
routing loss from specialist error.
Q: What is the judge-calibration tracker?
A: With --judge, every ambiguous-band row's LLM verdict is compared to the
deterministic score (scores.judge_calibration): agree rate + lenient/strict
lean — before trusting the judge more broadly.
Q: How do I add a new prompt version?
A: Add the constant + register it in PROMPT_VERSIONS (src/prompts.py).
The version key IS the experiment identity. NEVER edit a prompt string after
it has run — a changed prompt needs a new version key.
Q: How does the wiki stay in sync?
A: docs/wiki/ is version-controlled here; ./docs/wiki/sync-wiki.sh pushes it to the
public GitHub wiki.
Q: Where does the GH Pages site come from?
A: docs/ on main — rebuilt with scripts/site/build_site.py, deployed on
every push, no Actions runners.
llm-entity-extraction · prompt experiment loop for llm-mailroom — repository · experiment-log site