Document model routing and outcome observation

.agents/skills/orchestrator/SKILL.md

@@ -44,25 +44,32 @@ moving meanwhile. State is a record of authority, never a new authority source.
    for dispatch, return routing and branch retirement. Give DV independent spec-derived context;
    do not pass implementation-derived expected answers. Delegate through an
    available, authorized mechanism and record the returned durable handle.
-3. Choose a configured model/effort for uncertainty, impact and context needs.
-   Use a strong reasoning configuration for architecture, ambiguous failures
-   and consequential independent review; routine mechanical work can use an
-   available lower-cost configuration. Respect explicit requested models;
-   capability or credential absence is a blocker to report, not permission to
-   silently substitute. Keep concrete model/provider choices in local state,
-   requested settings separate from observed identity, and UNKNOWN/none/UNKNOWN
-   where runtime identity is not independently exposed. Do not rank models from
-   completion counts or treat their agreement as evidence.
+3. Choose a configured model/effort from the task's ambiguity/complexity,
+   failure impact, verification ease/cost and context needs; record those
+   inputs, the selected setting and rationale in the assignment. Use lighter
+   configured work for bounded routine tasks, and stronger reasoning for
+   architecture, difficult semantics, ambiguous failures and consequential
+   independent review. A role name alone does not choose a configuration.
+   Respect explicit requested models; capability or credential absence is a
+   blocker to report, not permission to silently substitute. Keep concrete
+   model/provider choices in local state, requested settings separate from
+   observed identity, and UNKNOWN/none/UNKNOWN where runtime identity is not
+   independently exposed. Do not rank models from completion counts or treat
+   their agreement as evidence.
 4. Reuse a live assigned worker; dispatch parallel work only when dependencies,
    ownership and resource budgets permit it. Inspect meaningful progress and
    final receipts, not just process liveness. Read the substantive review
    verdict. Route valid findings to the original author, require affected
    reruns and independent re-review, and keep the same work item/PR synchronized.
    Keep a PR Draft if making it Ready would trigger integration outside the
    recorded authorization; workflow readiness is not permission to merge.
-5. Evaluate output and review quality against cited rules and raw evidence.
-   If results are weak, correct context, acceptance, decomposition, skill gaps
-   or model fit. Do not weaken gates or manufacture bugs to score workers.
+5. Evaluate the complete delivery loop against cited rules and raw evidence.
+   Record acceptance/quality, confirmed misses or false alarms, repair cycles,
+   root versus human intervention, measured time/token/cost, and unknowns.
+   Compare only like difficulty, context, skill version and evidence scope. If
+   results are weak, correct context, acceptance, decomposition, skill gaps or
+   model fit. Do not weaken gates, manufacture bugs, or make unsupported model
+   ranking or maturity claims.
    Genuine defects and observed friction can motivate a separately owned,
    independently reviewed skill change; integrator curates accepted additions.
    After such a change, assign prospective application on the next matching
@@ -76,6 +83,14 @@ moving meanwhile. State is a record of authority, never a new authority source.
    evidenced, or checkpoint a precise remaining dependency/authority blocker;
    task counts are not closure.
 
+When configuration must change, use only a runtime-supported switch. Reconcile
+the live worker, owned jobs and candidate first; preserve the original author's
+ownership, a sufficient context handoff and independent review. If an in-place
+switch is unavailable, record that limitation and make an explicit handoff
+instead of blindly restarting or representing a requested setting as observed
+identity. Each assignment also records bounded continuation or escalation
+triggers.
+
 ## Persistence and operating limits
 
 The orchestrator owns the state index; specialists own their output and receipt

.agents/skills/orchestrator/references/operating.md

@@ -13,8 +13,9 @@ handoff file for each assignment:
 | Contract | Work-item/requirement IDs, exact spec/profile revisions, acceptance criteria, prerequisites and unresolved questions |
 | Ownership | One author role, separate independent reviewer, approver and state writer; explicit transfer if replacing a worker |
 | Inputs and scope | Pinned source/dependency identities, allowed reads/writes, forbidden artifacts, tool/access limitations and public/private boundary |
+| Model decision | Task ambiguity/complexity, failure impact, verification ease/cost and context needs; selected configured model/effort and rationale; intended independent reviewer; bounded continuation or escalation trigger |
 | Execution | Branch from current main (or approved dependency), worktree, requested model/effort, observed identity or UNKNOWN, commands, resource bounds and durable worker/job handles |
-| Return | Changed artifacts, exact candidate, planned/selected/executed/skipped/checked scope, raw receipts and verdicts, open bugs/risks, applicable reruns, manifest PR and next action |
+| Return | Changed artifacts, exact candidate, planned/selected/executed/skipped/checked scope, raw receipts and verdicts, acceptance/quality, confirmed misses or false alarms, repair cycles, root versus human intervention, measured time/token/cost and unknowns, manifest PR and next action |
 
 Freeze the independent DV context from the approved spec and golden-model
 sources; do not copy implementation-derived answers. Sparse worktrees or another
@@ -47,6 +48,17 @@ approval. Read raw checker results and the review verdict. Missing evidence is
 unverified; preserve explicit executed failures according to receipt semantics.
 Evaluate actual friction and genuine defects, not planted bugs for agent ranking.
 A skill improvement is its own bounded, independently reviewed framework change.
+Compare outcomes only across comparable difficulty, context, skill version and
+evidence scope; they do not establish a model ranking or maturity claim.
+
+## Change a runtime configuration
+
+Use an in-place switch only when the runtime supports it. Before switching,
+reconcile the live worker, all owned jobs and the candidate; retain the original
+author's ownership, provide the replacement sufficient context, and preserve
+independent review. If the capability is unavailable, record the limitation and
+use an explicit handoff when authorized. Do not blindly restart a worker or
+present a requested configuration as observed runtime identity.
 
 ## Close the work item and retire its workspace
 

docs/templates/ORCHESTRATION_STATE.md

@@ -45,6 +45,17 @@ Task status: OPEN (not started), RUNNING, WAITING (known condition), BLOCKED
 These are workflow states, separate from signoff acceptance. Record assignment
 transfer explicitly; the state index has one designated writer at a time.
 
+For every assignment, retain task ambiguity/complexity, failure impact,
+verification ease/cost and context needs; the selected configured model/effort
+and rationale; intended independent reviewer; and bounded continuation or
+escalation triggers. On return, retain acceptance/quality, confirmed misses or
+false alarms, repair cycles, root versus human intervention, measured
+time/token/cost and unknowns. Compare observations only when difficulty,
+context, skill version and evidence scope are comparable; they do not establish
+a model ranking or maturity claim. If changing configuration, record supported
+switching capability or its limitation, reconcile the live worker/jobs/candidate,
+and preserve author ownership, context handoff and independent review.
+
 Record PR lifecycle separately: Draft / in review / changes requested / approved /
 merged / terminal closed, with exact head and reason. After a terminal decision,
 record cleanup owner, active/dependency/dirty/evidence checks, retained artifacts,