Policy: Add portable model policy v1 #9

Merged

creatidy merged 2 commits from policy/model-policy-v1 into main

2026-09-03 18:05:35 +00:00

creatidy commented

2026-09-03 15:45:39 +00:00

(Migrated from github.com)

Closes #8

Why

Adds model-policy.json at the repository root as the canonical, strict-JSON policy artifact. Other repositories can fetch it directly without parsing Markdown; reproducible consumers should pin a commit or release.

Policy contents

schema_version: 1; policy_version: 1; updated 2026-09-03.
Five non-exclusive descriptive classes: execution_generalist, orchestration_synthesis, deep_technical_reasoner, scientific_methodological_specialist, and translation_editorial_specialist.
Six capability dimensions: reasoning, coding, scientific_methodological, writing_editorial, tool_use, and first-class translation_multilingual.
Existing formal profiles retained, with conceptual orchestration and translation profiles added.
Class typical_task_profiles references only formal selector-facing profile IDs; non-formal use cases remain descriptive typical_workflow_roles.
Explicit task-level/class relationship examples and non-binding profile/class affinities.
Dated 2026-09-03 current owner workflow mapping for GLM-5.3-Flash Max, GPT-5.6 Luna Max, GLM-5.3 Max, and GPT-5.6 Sol High. No Qwen mapping is added.

Classes are descriptive archetypes, not rankings or selector eligibility shortcuts. Actual capability minima and hard constraints must still be satisfied, while runtime capacity and scarcity/reservation policy remain independent. The least-scarce-sufficient selector sequence remains authoritative. Numeric model ratings and numeric profile minima are deferred to M2 calibration.

Documentation

Updated docs/capability-model.md with the six-dimension vocabulary, new profiles, canonical artifact link, translation policy, deferred orchestration dimension, and formal-profile/workflow-role distinction.
Added minimal links in README.md and docs/selection-policy.md.
Added decision record D-014 in docs/decisions.md.

Validation

make check: PASS
150 tests total, including 13 focused policy contract tests: PASS
uv run basedpyright: 0 errors, 0 warnings, 0 notes
python3 -m json.tool model-policy.json: PASS
No type suppressions.
No selector, provider acquisition, network, credential, model execution, ratings, minima, or scarcity implementation was added.

Closes #8 ## Why Adds `model-policy.json` at the repository root as the canonical, strict-JSON policy artifact. Other repositories can fetch it directly without parsing Markdown; reproducible consumers should pin a commit or release. ## Policy contents - `schema_version`: 1; `policy_version`: 1; updated 2026-09-03. - Five non-exclusive descriptive classes: `execution_generalist`, `orchestration_synthesis`, `deep_technical_reasoner`, `scientific_methodological_specialist`, and `translation_editorial_specialist`. - Six capability dimensions: `reasoning`, `coding`, `scientific_methodological`, `writing_editorial`, `tool_use`, and first-class `translation_multilingual`. - Existing formal profiles retained, with conceptual `orchestration` and `translation` profiles added. - Class `typical_task_profiles` references only formal selector-facing profile IDs; non-formal use cases remain descriptive `typical_workflow_roles`. - Explicit task-level/class relationship examples and non-binding profile/class affinities. - Dated 2026-09-03 current owner workflow mapping for GLM-5.3-Flash Max, GPT-5.6 Luna Max, GLM-5.3 Max, and GPT-5.6 Sol High. No Qwen mapping is added. Classes are descriptive archetypes, not rankings or selector eligibility shortcuts. Actual capability minima and hard constraints must still be satisfied, while runtime capacity and scarcity/reservation policy remain independent. The least-scarce-sufficient selector sequence remains authoritative. Numeric model ratings and numeric profile minima are deferred to M2 calibration. ## Documentation - Updated `docs/capability-model.md` with the six-dimension vocabulary, new profiles, canonical artifact link, translation policy, deferred orchestration dimension, and formal-profile/workflow-role distinction. - Added minimal links in `README.md` and `docs/selection-policy.md`. - Added decision record D-014 in `docs/decisions.md`. ## Validation - `make check`: PASS - 150 tests total, including 13 focused policy contract tests: PASS - `uv run basedpyright`: 0 errors, 0 warnings, 0 notes - `python3 -m json.tool model-policy.json`: PASS - No type suppressions. - No selector, provider acquisition, network, credential, model execution, ratings, minima, or scarcity implementation was added.

Rows
Columns