[P1] Add skill evolution orchestration over rlm_evolution + eval evidence #446

New issue

Closed

opened 2026-09-10 21:18:19 +00:00 by nsaspy · 1 comment

nsaspy commented

2026-09-10 21:18:19 +00:00

Owner

Parent: #167
Related: #142, PR #148, #168, #169, #170

Goal

Build the orchestration layer that turns the merged pure rlm_evolution kernel into a bounded skill experiment loop without contaminating that kernel with providers, scheduling, authority, or product code.

Pipeline

trusted skill genotype/schema
-> candidate variants
-> mutate/crossover through rlm_evolution
-> bounded candidate materialization
-> selection + outcome eval suites
-> vector fitness
-> Pareto selection
-> lifecycle promotion proposal

Requirements

keep rlm_evolution pure closed-data machinery;
define a trusted closed genotype/schema for skill experiments;
materialization may vary allowed skill metadata/instruction fields only within the declared schema;
optional model-assisted text proposal must run through normal provider/budget/trace semantics and remain candidate data;
evaluator execution reuses existing async/Future/cancellation/runtime contracts;
fitness is multi-objective and can include correctness, trigger quality, tokens, cost, latency, regressions and robustness;
lineage/evidence from mutation/crossover is retained;
immutable constraints and authority/capability ceilings cannot be mutated away;
promotion is proposed to #170 lifecycle policy, never self-applied by candidate/model output.

Acceptance

deterministic end-to-end experiment starts from at least two skill candidates;
mutation and/or crossover uses the canonical rlm_evolution API;
candidates are evaluated through #168/#169-compatible records;
vector fitness is computed without collapsing all objectives into one magic scalar;
Pareto/non-dominated selection is reproducible;
lineage and evaluation evidence survive the full cycle;
winner can be proposed for promotion or all candidates can be rejected;
cancellation/budget exhaustion produce structured incomplete experiments, not false winners;
no generated arbitrary Prolog is executed.

Non-goals

no model-weight/parameter-space ES;
no DeepSeek/Cordis/AgentProlog product schema in core;
no second scheduler/verifier/effect system.

Reconcile #142 research and merged PR #148 against current main before implementation. Do as much coherent work as possible per cycle.

Parent: #167 Related: #142, PR #148, #168, #169, #170 ## Goal Build the orchestration layer that turns the merged pure `rlm_evolution` kernel into a bounded skill experiment loop without contaminating that kernel with providers, scheduling, authority, or product code. ## Pipeline ```text trusted skill genotype/schema -> candidate variants -> mutate/crossover through rlm_evolution -> bounded candidate materialization -> selection + outcome eval suites -> vector fitness -> Pareto selection -> lifecycle promotion proposal ``` ## Requirements - keep `rlm_evolution` pure closed-data machinery; - define a trusted closed genotype/schema for skill experiments; - materialization may vary allowed skill metadata/instruction fields only within the declared schema; - optional model-assisted text proposal must run through normal provider/budget/trace semantics and remain candidate data; - evaluator execution reuses existing async/Future/cancellation/runtime contracts; - fitness is multi-objective and can include correctness, trigger quality, tokens, cost, latency, regressions and robustness; - lineage/evidence from mutation/crossover is retained; - immutable constraints and authority/capability ceilings cannot be mutated away; - promotion is proposed to #170 lifecycle policy, never self-applied by candidate/model output. ## Acceptance - [ ] deterministic end-to-end experiment starts from at least two skill candidates; - [ ] mutation and/or crossover uses the canonical `rlm_evolution` API; - [ ] candidates are evaluated through #168/#169-compatible records; - [ ] vector fitness is computed without collapsing all objectives into one magic scalar; - [ ] Pareto/non-dominated selection is reproducible; - [ ] lineage and evaluation evidence survive the full cycle; - [ ] winner can be proposed for promotion or all candidates can be rejected; - [ ] cancellation/budget exhaustion produce structured incomplete experiments, not false winners; - [ ] no generated arbitrary Prolog is executed. ## Non-goals - no model-weight/parameter-space ES; - no DeepSeek/Cordis/AgentProlog product schema in core; - no second scheduler/verifier/effect system. Reconcile #142 research and merged PR #148 against current `main` before implementation. Do as much coherent work as possible per cycle.

~~nsaspy referenced this issue 2026-09-10 21:20:47 +00:00~~

[P1] Add skill evolution orchestration over rlm_evolution + eval evidence #171

nsaspy commented

2026-09-10 21:21:32 +00:00

Author

Owner

Duplicate of #171 (pre-existing Forgejo mirror). Closing this accidental duplicate created by today's open-state sync; #171 stays canonical on Forgejo.

nsaspy closed this issue

2026-09-10 21:21:32 +00:00