gnomon-forecast 0.5.0__tar.gz

gnomon_forecast-0.5.0/.gitignore

@@ -0,0 +1,93 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.venv/
+dist/
+*.egg-info/
+gnomon-output/
+
+
+# secrets
+.env
+.env.*
+
+# Benchmark corpora are downloaded, not versioned. A `git add -A` swept
+# 582 MB of them into history once; GitHub rejects the 290 MB file among
+# them, and nothing in the codebase reads these paths.
+data/
+
+# Exception: the curated real-data corpora for DossierBench and
+# BreachBench ARE versioned. They are small (~220 KB total), the
+# benchmarks cannot run without them, and their sha256 is pinned into
+# every run's dataset identity — so the exact bytes must travel with the
+# code, not a download script that can drift or be unreachable. Both the
+# directory and its contents need re-including: git does not descend into
+# an excluded directory.
+!benchmarks/breachbench/data/
+!benchmarks/breachbench/data/**
+!benchmarks/dossierbench/data/
+!benchmarks/dossierbench/data/**
+
+# CiK's official result cache. Swept in by the same `git add -A`; the runs
+# that produced it used --no-cache because the cached path fails on
+# object-dtype columns.
+inference_cache/
+
+# Benchmark run outputs. Regenerable from benchmarks/configs/*.yaml, and
+# large enough (10.5 MB of JSON, PNG, and PDF) that carrying them makes
+# every clone pay for one machine's run.
+#
+# `results/*` rather than `results/`: git does not descend into an excluded
+# directory, so the negation below would never be reached under the shorter
+# form.
+results/*
+metric_scaling_cache/
+
+# Except the leakage-trap results — the evidence behind the McNemar figure
+# in docs/leakage-trap-results-2026-08.md. 62 KB of text, and a published
+# claim should not rest on output nobody can see.
+!results/leaktrap/
+
+# And except the future-context A/B pre-registration and its write-up: a
+# prediction registered before the run is only worth something if it is
+# versioned before the run.
+!results/future-context-ab/
+results/future-context-ab/cik-*
+
+# Except the proposer-trust warrant registration: the hypothesis is
+# versioned before the experiment exists, which is the point of it.
+!results/proposer-trust-warrant/
+
+# Same rule for the structural-effects registration.
+!results/structural-effects/
+!results/seasonal-regime-effects/
+
+# And except the news-regime exploration: pre-registration, summaries, and
+# the scripts that regenerate them are versioned; per-task raw outputs are
+# not (same rule as the benchmark corpora — regenerable, and bulky).
+!results/news-regime-explore/
+results/news-regime-explore/raw/
+
+# And the guardrail build's pre-registration and write-up, same rule.
+!results/short-history-guardrail/
+results/short-history-guardrail/raw/
+
+# Curated aggregate-only benchmark releases. Raw runs remain ignored and are
+# retained as CI artifacts; benchmarks.release validates size, provenance,
+# digests, and sensitive-field exclusions for these directories.
+!results/benchmark-releases/
+!results/benchmark-releases/**
+
+
+# Root-level run scratch: per-machine census JSONs, arm run directories,
+# their *-work task trees, and downloaded model checkpoints. The citable
+# versions of these live under results/<registration>/ and are
+# un-ignored above; these loose copies were drowning every git status.
+/census*.json
+/*-work/
+/results-*/
+/models/
+
+# Local content-addressed context receipts and numeric assessment cache.
+# These can contain operational source metadata and are never source files.
+/.gnomon/