model-routing-as-code explores how to represent model/provider/local/cloud selection rules, correctness gates, cost routing, fallback policy, and eval thresholds as version-controlled, reviewable, testable, auditable, and agent-operable source material.
Model Routing-as-Code means treating model/provider/local/cloud selection rules, correctness gates, cost routing, fallback policy, and eval thresholds as canonical operational state expressed through files, schemas, examples, tests, and reviewable change history.
- Define the domain clearly.
- Collect prior art and examples.
- Provide reusable schemas and templates.
- Support human review and agent execution.
- Preserve auditability, provenance, and governance boundaries.
- This repo is not a universal standard.
- This repo is not legal, security, compliance, or operational advice.
- This repo does not canonize HUMMBL/BaseN/Ownward concepts unless explicitly marked and audited.
seed->v0.1-draft
- Boundary:
docs/v0.1-boundary.md - Schema:
schemas/model-routing-as-code-v0.1.json - Example:
examples/route-manifest-v0.1.example.json - Fixtures:
fixtures/valid/route-manifest-v0.1.valid.json,fixtures/invalid/route-manifest-v0.1.invalid.json
Per scope decision #7 (Option B), the policy packet uses a layered runtime-envelope + model-route architecture:
- Documentation:
docs/policy-packet-v0.1.md - Schema:
schemas/policy_packet.v0.1.schema.json - Validator:
tools/validate_policy_packet.py - Valid fixture:
fixtures/valid/policy_packet.v0.1.valid.json - Invalid fixtures (9 adversarial cases):
fixtures/invalid/policy_packet/ - Example route decisions (5, placeholder outcomes):
examples/policy-packet/ - Tests:
tests/test_policy_packet.py
Runtime, privacy, and authority gates resolve before model/provider/cost selection. A cheaper model cannot bypass an unsafe runtime, privacy boundary, authority gate, or required human review.
- GPT-5.6 Sol Ultra evidence packet (search-bounded, checked 2026-07-11):
docs/evaluations/2026-07-11-gpt-5.6-sol-ultra-evidence-packet.md - Provisional reasoning-effort route policy:
docs/route-policy/reasoning-effort-route-policy.v0.1.md - Controlled xhigh–Max–Ultra protocol:
docs/evaluations/ultra-evaluation-protocol.v0.1.md - Machine benchmark inventory and sanitized chart provenance:
data/gpt-5.6-sol-ultra/benchmark_dataset.v0.1.json - Governed execution-receipt contract:
schemas/ultra_evaluation_receipt.v0.1.schema.json - Mandatory deterministic receipt validator:
tools/validate_ultra_evaluation_receipt.py - Sanitized publication rebuild receipt:
receipts/2026-07-11-gpt-5.6-sol-ultra-publication-rebuild.json - Promotion and review-closure receipt:
receipts/2026-07-11-gpt-5.6-sol-ultra-promotion.json
These artifacts are evidence and experiment scaffolding, not adopted universal routing policy.
Full receipt conformance requires the deterministic validator. Draft 2020-12 schema validation covers the portable structural contract; the validator also enforces cross-field route, lane, telemetry, timestamp, and observation semantics.
Public seed repository. Initial executable packet in progress as seed -> v0.1-draft.
This repo is part of the HUMMBL AI governance architecture.
| Repo | Purpose |
|---|---|
| hummbl-governance | Governance runtime — kill switch, circuit breaker, cost governor, 34 primitives |
| mcp-server | MCP server for Base120 mental models — 120 reasoning operators for Claude/Cursor |
| base120 | 120 mental models for structured reasoning — stdlib-only Python library |
| crab | CRAB protocol — Check, Reason, Act, Bus for multi-agent turn execution |
| docs | Canonical public documentation — Mintlify-powered |
| hummbl-toolkit | HUMMBL utility toolkit |
Learn more at hummbl.io.
Reuben Bowlby — Founder, HUMMBL
- GitHub: @hummbl-dev
- X/Twitter: @ReubenBowlby
- LinkedIn: reuben-bowlby
- Website: hummbl.io