mirror of
https://github.com/ruvnet/RuVector.git
synced 2026-07-10 01:38:44 +00:00
b2a32eae2f
* feat(sona): metaharness-Darwin evolves EWC++ config beyond hand-tuned SOTA examples/darwin_ewc: applies the Meta-Harness 'freeze the model, evolve the harness' pattern to SONA's continual-learning layer — frozen = the EWC++ algorithm (EwcPlusPlus), evolved = its EwcConfig genome (lambda schedule, Fisher decay, auto task-boundary threshold, learning rate). Benchmark: a single weight vector trained on a sequence of tasks (no replay, auto-detected boundaries) — the canonical plasticity-vs-forgetting frontier. Darwin (GA + coordinate-descent polish) evolves the genome on TRAIN task- sequences; results reported on HELD-OUT sequences (different seeds). Measured (deterministic), held-out: the evolved config beats EwcConfig::default() (the crate's hand-tuned 'OPTIMIZED' values) by 35% lower final loss and 98.6% less forgetting — a strict Pareto win (plasticity also improves), and it generalizes to unseen task sequences. clippy -D warnings clean, fmt clean. Co-Authored-By: claude-flow <ruv@ruv.net> * feat(sona): weightAdapter gene — Darwin selects/prunes a fine-tuned adapter Extends the metaharness-Darwin line: expose a fine-tuned adapter (e.g. a LoRA distilled from verified SWE-bench trajectories — the 'autonomous data engine') as a gene (which_adapter, alpha) so evolutionary selection decides whether/how much to apply it (w_eff = w_base + alpha·Δw) instead of assuming new weights are better. examples/darwin_weightadapter demonstrates it on two conflicting domains with a generalizing adapter and an overfit one. Key finding (sharpens the idea): 'selection prunes overfit adapters' holds ONLY under per-domain evaluation. Measured (held-out, in-dist-majority eval): overfit α=0.55 → ΔA +0.249 / ΔB -0.357 (regresses out-dist) AGGREGATE (volume-weighted) fitness → picks the overfit adapter (silent B regression) PER-DOMAIN (no-regression Pareto) → prunes it, keeps the generalizing adapter So: evolve the adapter as a gene, but score it per-repository. clippy/fmt clean. Co-Authored-By: claude-flow <ruv@ruv.net> * docs(adr): ADR-271 metaharness-Darwin for SONA self-improvement Documents the metaharness-Darwin-evolves-SONA architecture: EWC++ config evolution (PR #615), the weightAdapter gene (per-domain Pareto selection of fine-tuned adapters), the Autonomous Data Engine (execution-verified SWE-bench trajectories -> DPO pairs), and four Ornith-1.0 borrows (immutable-boundary + deterministic-monitor-with-exclude-from-advantage + frozen-LLM-judge-veto reward-hacking defense; per-task-category specialization; two-stage scaffold reward credit; staleness-weighted replay). Method-not-model: external evolutionary vs Ornith's in-weights RL. Co-Authored-By: claude-flow <ruv@ruv.net> * feat(sona): darwin-guard reward-hacking defense (Ornith-1.0 borrow, ADR-271) 3-layer defense for evolutionary config search: (1) immutable verifier boundary (screen is a pure fn of verifier output the candidate can't fabricate); (2) deterministic monitor — non-finite / out-of-bounds / degenerate candidates are EXCLUDED from selection (best_accepted), not zero-scored, so a hack can neither win nor bias the advantage; (3) IntentJudge trait = frozen-LLM veto-only layer. Wired into darwin_ewc: NaN/collapsed configs are excluded from the GA ranking (also fixes the partial_cmp().unwrap() NaN-panic). 4 unit tests; benchmark still reaches beyond-SOTA (35% lower loss, 98.6% less forgetting) unchanged. clippy -D warnings + fmt clean. Co-Authored-By: claude-flow <ruv@ruv.net> * feat(sona): per-task-category genome router beats single global config (ADR-271) Ornith-1.0 borrow #2 (per-category specialization): evolve a router task-class -> genome instead of one global EwcConfig. Two continual-learning workload classes with conflicting optima (STABLE wants high lambda / retain; VOLATILE wants low lambda / stay plastic). Guard-screened evolution. Measured (held-out, adequate per-class data): per-category router 0.1122 vs single best global genome 0.1144 -> router ~1.9% better on unseen sequences, because one config cannot serve conflicting workloads. Honest caveat (discovered + documented): the gain REVERSES when per-class data is scarce — a specialized config overfits while the pooled global generalizes. Per-category routing needs enough per-category samples (Ornith's regime). ADR-271 updated; clippy/fmt clean. Co-Authored-By: claude-flow <ruv@ruv.net> * feat(sona): online auto-tuner with staleness-weighted replay (ADR-271, Ornith borrow #4) auto_tuner module: StalenessSchedule (Ornith w(d_t): fresh<=k1, exp-decay, drop>k2) + StalenessWindow (staleness-weighted running estimate of recent config performance, evicts stale obs). 4 unit tests. examples/darwin_autotuner: a (1+1)-ES that adapts a DEPLOYED EwcConfig to a drifting workload stream (regime A -> B at the midpoint), scoring the incumbent on the staleness window and accepting a perturbation only when it beats the recent score. Measured: online tuner ~3% lower post-drift loss than the static deployment config (10 accepted re-tunes). Margin is modest on synthetic regimes; the durable win is the reusable staleness machinery + the online-adaptation principle (a fixed offline-tuned config goes stale under drift). Completes the four ADR-271 components. clippy --all-targets -D warnings + fmt clean; 102 sona tests pass. Co-Authored-By: claude-flow <ruv@ruv.net> * feat(sona): contamination/disjointness guard in darwin-guard (weight-eft/ADR-198 borrow) Adds the train/eval contamination guard — the gap @metaharness/weight-eft exposed in our reward-hacking-only guard. contamination()/assert_train_eval_disjoint() fail on any train∩eval instance-ID overlap (training/selecting on eval instances is fake lift); filter_holdout() partitions a set disjoint-by-construction and surfaces what was excluded. The SONA-side analog of weight-eft's assertTrainEvalDisjoint. 2 new tests (6 total in darwin_guard). ADR-271 updated: §3 Data Engine now cites @metaharness/weight-eft + adopts its RLHF-correct recipe (SFT distills ALL gold incl. off-policy frontier successes; DPO ON-POLICY cheap-vs-cheap only), and the darwin-guard borrow gains layer (iv) the contamination disjointness guard. clippy -D warnings + fmt clean. Co-Authored-By: claude-flow <ruv@ruv.net> * chore(release): ruvector-sona 0.2.1 — darwin_guard + auto_tuner modules Non-breaking minor feature release (new public modules darwin_guard, auto_tuner). Patch bump keeps the ^0.2 requirement of all in-workspace dependents (ruvllm, rvlite, mcp-brain, ...) satisfied. Co-Authored-By: claude-flow <ruv@ruv.net> --------- Co-authored-by: ruvnet <ruvnet@gmail.com> |
||
|---|---|---|
| .. | ||
| darwin_autotuner.rs | ||
| darwin_ewc.rs | ||
| darwin_router.rs | ||
| darwin_weightadapter.rs | ||