Update multi-model LLM operating policy and role assignments #48

New issue

Closed

opened 2026-09-07 08:29:44 +00:00 by kilo · 1 comment

kilo commented

2026-09-07 08:29:44 +00:00

Member

Goal

Align Scarcity Router's durable model-routing and execution-governance policy
with the owner's updated LLM Operating Policy for Multi-Model Projects.

Why

The current Scarcity Router policy already implements the core role-based,
least-scarce-capable philosophy, but the owner's operating model has evolved:
GPT-6 Astra Low is now the preferred scientific/methodological specialist and
a deep-technical reasoning option; reasoning effort is an explicit routing
parameter; durable-state, review-independence, runtime-attestation,
incremental-work and completed-work rules need stronger durable representation.

Scope

human-readable LLM operating policy
model-policy.json descriptive/reference metadata
Kilo operating governance
AGENTS guidance where necessary
README/documentation map
policy decision record
focused policy tests
record the Astra/reasoning-effort selector-onboarding gap for later work

Acceptance Criteria

stable role taxonomy remains model-independent
current reference role assignments reflect the new owner policy
Astra Low may appear as descriptive/reference assignment only
Astra is NOT added to the active model catalog
current selector-facing task requirements remain unchanged
reasoning effort is recorded as a routing concern but dynamic selector
support remains explicitly unimplemented
review independence is durable policy
evidence preparation and adjudication are separated
large work requires early durable checkpoints
durable repository/artifact state outranks session UI state
retry/loop handling is finite
completed work is not repeated without source change/regression
runtime model attestation states are documented
local-model generic policy and Scarcity Router's stricter D-017 override are
distinguished
existing M2 selector behavior remains unchanged
make check and basedpyright pass

Constraints

no Astra capability calibration
no Astra capacity binding
no model-catalog change
no task-profile numeric change
no selector/scarcity/policy algorithm change
no provider change
no model execution
no M3b/M3c implementation
human merge gate

Links

docs/capability-model.md
docs/model-calibration.md
docs/selection-policy.md
docs/decisions.md
docs/roadmap.md

## Goal Align Scarcity Router's durable model-routing and execution-governance policy with the owner's updated LLM Operating Policy for Multi-Model Projects. ## Why The current Scarcity Router policy already implements the core role-based, least-scarce-capable philosophy, but the owner's operating model has evolved: GPT-6 Astra Low is now the preferred scientific/methodological specialist and a deep-technical reasoning option; reasoning effort is an explicit routing parameter; durable-state, review-independence, runtime-attestation, incremental-work and completed-work rules need stronger durable representation. ## Scope - human-readable LLM operating policy - model-policy.json descriptive/reference metadata - Kilo operating governance - AGENTS guidance where necessary - README/documentation map - policy decision record - focused policy tests - record the Astra/reasoning-effort selector-onboarding gap for later work ## Acceptance Criteria - stable role taxonomy remains model-independent - current reference role assignments reflect the new owner policy - Astra Low may appear as descriptive/reference assignment only - Astra is NOT added to the active model catalog - current selector-facing task requirements remain unchanged - reasoning effort is recorded as a routing concern but dynamic selector support remains explicitly unimplemented - review independence is durable policy - evidence preparation and adjudication are separated - large work requires early durable checkpoints - durable repository/artifact state outranks session UI state - retry/loop handling is finite - completed work is not repeated without source change/regression - runtime model attestation states are documented - local-model generic policy and Scarcity Router's stricter D-017 override are distinguished - existing M2 selector behavior remains unchanged - make check and basedpyright pass ## Constraints - no Astra capability calibration - no Astra capacity binding - no model-catalog change - no task-profile numeric change - no selector/scarcity/policy algorithm change - no provider change - no model execution - no M3b/M3c implementation - human merge gate ## Links - `docs/capability-model.md` - `docs/model-calibration.md` - `docs/selection-policy.md` - `docs/decisions.md` - `docs/roadmap.md`

kilo referenced this issue

2026-09-07 08:40:41 +00:00

Evaluate GPT-6 Astra capacity scope and reasoning-effort selector onboarding #49

kilo commented

2026-09-07 08:40:51 +00:00

Author

Member

Follow-up issue #49 covers evidence-first GPT-6 Astra capacity scope, identity, capability calibration and reasoning-effort selector onboarding. No onboarding implementation is part of this policy slice.

kilo referenced this issue from a pull request that will close it,

2026-09-07 08:43:44 +00:00

Update multi-model LLM operating policy #50