Implementation and validation status
Implemented and exercised
peppa.compiled implements deterministic topological scheduling, immutable input copies, unique artifact producers, dependency and budget checks, revision checks, bounded attempts, atomic output validation/commit, content-addressed caching, hash-linked semantic traces, replay, duplicate-lineage invariance, nonnegative ridge stacking, worst-context generation reward, deterministic failure-category scheduling, and complete-score release checks.
The reference kernel uses a scalar reserved-cost budget and Python validator functions. Worker-specific semantic preconditions, wall-clock timeouts, actual multi-resource accounting, and parallel execution are production-adapter responsibilities. The manuscript states the full scientific protocol. The CPU kernel is a tested reference for its scheduling and mathematical operations.
The tests passed. Eight real SNOOPPI records were processed in run_compiled_example.py; original labels and source IDs were retained. Required PAMPA scores and chemical verification were absent, so the release report correctly contains zero nominated peptides. Cached execution and replay returned identical artifacts. This validates software behavior, not design efficacy.
Existing utilities implement chemical annotations, AF3 PTM input export, masked-label PTM-head fitting, connected-component partitions, conformal radii, conservative margins, diversity selection, grouped bootstrap, and mass-balanced ternary fitting. Existing controller interfaces and the moPPIt/PeptiVerse bridges are retained for integration and comparative baselines.
Integration required before scientific evaluation
- Connect Muse/GPT-6 Astra Max/Claude Opus Max structured outputs to PlanProposal validation and the compiled runner. The existing baseline provider adapters have not been run against live services here.
- Register the native generator and scientific workers, with checkpoint and environment hashes. moPPIt/PeptiVerse bridges require native known-input tests. Other scientific tools have contracts and integration instructions.
- Implement complete graph-based chemical standardization, noncanonical monomer mapping, formula/mass calculation, supplier/route retrieval, and manufacturing annotations. The CPU release routine consumes verified identity records from these adapters.
- Fit ensemble weights and calibration on the declared disjoint data; curate structural predicate thresholds and target/species domain rules.
- Compile the two-cycle scientific graph and its fixed diagnostic mapping. The robust objective and category scheduler are executable; native generator conditioning is an integration task.
- Implement timeout enforcement, multiple resource budgets, scientific validators, full report export, and held-out benchmark evaluation around the reference kernel.
No new therapeutic peptide generation, GPU structural prediction, prospective assay, animal study, or clinical evaluation was executed while authoring this archive. The tables and empirical plots are pending. The Hugging Face Space is specified, not deployed.
Trace interpretation
compiled_trace.jsonl is the current deterministic source-data example. real_data_trace.jsonl belongs to the earlier generic controller example. Neither contains newly measured biology. Synthetic unit tests are separate from these real-data examples. A candidate's source interaction label must never be copied into an unmeasured property endpoint.
The v4 revision also checks Astra maximum reasoning and Opus maximum effort/adaptive thinking request settings using local provider mocks. Live provider calls and scientific model benchmarking remain pending.