Skip to content

PSION_PROGRAM_MAP.md

Latest commit

 

History

History
156 lines (118 loc) · 6.38 KB

PSION_PROGRAM_MAP.md

File metadata and controls

156 lines (118 loc) · 6.38 KB

Psion Program Map

Status: canonical PSION-1 / #357 learned-model program-map contract, written.

Why This Doc Exists

The Psion learned-model lane now has canonical docs for governance, data, training, evaluation, serving, rollback, and bounded decentralized follow-on work. This doc is the umbrella map that ties those contracts together in one dependency-ordered public-safe program.

It exists so the repo can point at one durable, reviewable learned-lane map instead of asking readers to reconstruct the program from scattered issue titles or doc filenames.

This doc is the generic learned Psion family map.

It is not the executor-capable Psion umbrella contract. That split now lives in docs/PSION_EXECUTOR_PROGRAM.md, which keeps generic compact-decoder Psion separate from the bounded executor-capable Tassadar profile.

Claim Boundary

This tranche is for a learned bounded-success systems model with explicit route and refusal behavior.

It is not:

  • CPU-reference exactness
  • compiled exact execution
  • proof-backed verification
  • implicit execution hidden behind language answers

Exact executor closure remains separate work under the Tassadar lane and its canonical docs, including docs/PSION_EXECUTOR_PROGRAM.md.

The generic learned lane now also has one retained operator-facing serve closure in docs/PSION_GENERIC_LOAD_AND_GENERATE.md. That packet proves artifact-backed load-to-serve integrity on the current generic runtime without claiming executor-lane replacement or exactness.

The current reference pilot is also a CPU-reference lane. It remains a valid bounded operator, receipt, checkpoint, and promotion fixture, but it is not a valid GPU training proof target. Future Google GPU audits must name the exact trainer command and execution backend explicitly instead of treating GPU-hosted reference runs as accelerator-backed training proof.

The canonical bounded accelerator-backed reference trainer is still the dedicated psion_accelerated_reference_pilot lane, not the CPU reference pilot or the plugin reference bundle paths.

The canonical default operator entrypoint for the real broader-pretraining lane is now:

  • TRAIN
  • docs/PSION_ACTUAL_PRETRAINING_RUNBOOK.md

The older bounded reference lane remains available explicitly through:

  • ./TRAIN --lane reference_pilot
  • docs/PSION_LOCAL_FIRST_TRAIN_RUNBOOK.md

The generic learned family also has one canonical serve packet on the current artifact-backed runtime. That packet is a family-integrity proof, not a second serve stack and not a claim that the generic family has already closed the full compact-decoder-anchor runtime gap.

The umbrella family now also has one canonical family serve vocabulary in docs/PSION_FAMILY_SERVE_VOCABULARY.md. That packet keeps generic learned, plugin-conditioned, and executor-capable rows visible under one family name without flattening their evidence classes.

Dependency Order

The Psion program is dependency-ordered in four tracks:

  1. governance critical path: PSION-2, PSION-3, PSION-4, PSION-5, PSION-6
  2. data/tokenizer critical path: PSION-7, PSION-8, PSION-9, PSION-10
  3. model/training critical path: PSION-11, PSION-12, PSION-13, PSION-14, PSION-15, PSION-16, PSION-17, PSION-18
  4. eval/route/serve critical path: PSION-19, PSION-20, PSION-21, PSION-22, PSION-23, PSION-24, PSION-25, PSION-26, PSION-27, PSION-28, PSION-29, PSION-30

The ordering matters because later work must bind back to earlier governance, acceptance, capability, and rollback truth instead of silently widening claims.

Canonical Track Map

Program / Governance

  • PSION-2 / #358: docs/PSION_CORPUS_ADMISSION.md
  • PSION-3 / #359: docs/PSION_SOURCE_LIFECYCLE.md
  • PSION-4 / #360: docs/PSION_BENCHMARK_ISOLATION.md
  • PSION-5 / #361: docs/PSION_ACCEPTANCE_MATRIX.md
  • PSION-6 / #362: docs/PSION_CAPABILITY_MATRIX.md

Data / Tokenizer

  • PSION-7 / #363: docs/PSION_RAW_SOURCE_INGESTION.md
  • PSION-8 / #364: docs/PSION_TOKENIZER_TRAINING.md
  • PSION-9 / #365: docs/PSION_TOKENIZED_CORPUS.md
  • PSION-10 / #366: docs/PSION_SAMPLING_POLICY.md

Model / Training

  • PSION-11 / #367: docs/PSION_COMPACT_DECODER.md
  • PSION-12 / #368: docs/PSION_PRETRAIN_STAGE.md
  • PSION-13 / #369: docs/PSION_RUN_OBSERVABILITY.md
  • PSION-14 / #370: docs/PSION_PILOT_PRETRAINING_RUN.md
  • PSION-15 / #371: docs/PSION_CHECKPOINT_RECOVERY.md
  • PSION-16 / #372: docs/PSION_RENTED_CLUSTER_RUNBOOK.md
  • PSION-17 / #373: docs/PSION_TRUSTED_CLUSTER_RUN.md
  • PSION-18 / #374: docs/PSION_REASONING_SFT.md

Eval / Route / Serve

  • PSION-19 / #375: docs/PSION_BENCHMARK_PACKAGES.md
  • PSION-20 / #376: docs/PSION_BENCHMARK_LABEL_GENERATION.md
  • PSION-21 / #377: docs/PSION_ARCHITECTURE_REASONING_BENCHMARK.md
  • PSION-22 / #378: docs/PSION_NORMATIVE_SPEC_READING_BENCHMARK.md
  • PSION-23 / #379: docs/PSION_ENGINEERING_SPEC_INTERPRETATION_BENCHMARK.md
  • PSION-24 / #380: docs/PSION_MEMORIZATION_VS_REASONING_PROBES.md
  • PSION-25 / #381: docs/PSION_ROUTE_CLASS_EVALUATION.md
  • PSION-26 / #382: docs/PSION_REFUSAL_CALIBRATION.md
  • PSION-27 / #383: docs/PSION_SERVED_EVIDENCE.md
  • PSION-28 / #384: docs/PSION_SERVED_OUTPUT_CLAIMS.md
  • PSION-29 / #385: docs/PSION_CAPABILITY_WITHDRAWAL.md
  • PSION-30 / #386: docs/PSION_DECENTRALIZED_CONTRIBUTION.md

Operating Rules

Every later Psion artifact must preserve the following:

  • governance and rights state stay explicit and versioned
  • held-out and contamination boundaries remain reviewable
  • promotion stays bound to docs/PSION_ACCEPTANCE_MATRIX.md
  • publication and refusal posture stay bound to docs/PSION_CAPABILITY_MATRIX.md
  • rollback and downgrade history stay bound to docs/PSION_CAPABILITY_WITHDRAWAL.md
  • CPU-reference lanes may not be used as the primary proof target for Google GPU training claims
  • future GPU audits must state the trainer path and delivered execution backend explicitly

The TRAIN entrypoint preserves that rule by defaulting to the accelerated reference lane and refusing explicitly when the admitted CUDA host is unavailable, busy, or missing required tooling.

No part of this map authorizes widening the learned lane into hidden executor claims or using decentralized participation as a shortcut around evaluation, rollback, or provenance discipline.