029-evaluation-repeatability
Evaluation repeatability, baselines, diffs and selection gates (amends 004 and 006)
An amendment of approved specs 004 and 006 (R-16), for #76 child E2. Evaluation runs each configuration N times from a declared repeat source (deterministic replay, captured historical run sets, or live runs once spec 026 is implemented), reports per metric and per slice the mean, standard deviation, minimum, maximum, a pooled estimate and a case-clustered percentile bootstrap interval computed with a named, versioned, seeded generator, and derives declared noise bands. Adds named baselines (`rustev.baseline/1`), a reproducibility check, diffs in JSON and Markdown, multi-candidate comparison, held-out accounting against a host-kept ledger file, and a declarative selection gate (`rustev.selection-gate/1`) with stated CI exit codes. Every existing document, gate and CLI output keeps its bytes. Answers spec 023's open question 3. Approved (A-25); approval authorizes no live run, provider call or spend, no quality claim, and no change to `rustev.eval-report/1` or any plan identity.
Read the full pinned specification ↗Relationships
depends on
depended on by
- 034-parameterized-definitions
- 039-model-backed-rank
- 040-decision-explanation
- 041-ranking-and-rubric-metrics
- 042-decision-service-and-remote-seams
- 043-live-re-execution-at-scale
- 061-host-scored-repeat-sets
- 063-node-binding-parity
- 064-selection-gate-case-regressions
- 065-selection-gate-run-integrity
- 067-node-live-repeat-runs
Claimed units
- extends
{ "nature": "amending", "spec": "004-evaluation-and-replay", "unit": { "kind": "directory", "path": "crates/rustev-eval/" } } - extends
{ "nature": "amending", "spec": "006-cli-surface", "unit": { "kind": "directory", "path": "crates/rustev-cli/" } } - extends
{ "nature": "additive", "spec": "001-boundaries-and-authority", "unit": { "kind": "file", "path": "Makefile" } } - extends
{ "nature": "additive", "spec": "001-boundaries-and-authority", "unit": { "kind": "directory", "path": "scripts/acceptance/" } } - extends
{ "nature": "additive", "spec": "027-node-and-wasm-bindings", "unit": { "kind": "directory", "path": "tools/rustev-conformance/" } } - extends
{ "nature": "additive", "spec": "031-public-api-and-release-family", "unit": { "kind": "directory", "path": "tools/rustev-release/" } }
Source headings
- 029: Evaluation repeatability, baselines, diffs and selection gates (amends 004 and 006)
- 1. Purpose
- 2. Territory
- 2.1 Binding principles of #76
- 3. Behavior
- 3.1 Repeat sources and runs
- 3.2 Per-case measures
- 3.3 Repeat configuration and statistics
- 3.4 Bootstrap intervals
- 3.5 Noise bands
- 3.6 Slices
- 3.7 Split roles and held-out discipline
- 3.8 Baselines
- 3.9 Reproducibility check
- 3.10 Diffs
- 3.11 Multi-candidate comparison
- 3.12 Selection gates
- 3.13 CLI surface (amends spec 006)
- 3.14 Exit codes
- 3.15 Compatibility
- 4. Out of scope
- 5. Observable negative cases
- Acceptance
- Verification
- Implementation record
- Decisions at approval
Frozen anchors
No frozen anchors declared.