Model Intelligence / Creatidy Intelligence

Intended as a public/shared intelligence layer about the external AI ecosystem: model identity, capability evidence, public pricing/plans/promotions, execution surfaces, provenance, freshness, disagreement and semantic snapshot changes.

At the audited develop revision fb7299810fdc4612d4e0559465faef571662c6ef, M0-01R2 is integrated through PR #10. It is a deterministic synthetic architecture proof, not a service, real-source value proof, published knowledge artifact or verified Router integration. The full frozen A/B/C proof remains undelivered. Issue #12 is completed/closed through merged PR #23. The accepted outcome is ADAPT to a thin upstream-backed public-evidence and selected-history layer, not product GO, full A/B/C proof or installed MI-Router acceptance. Future implementation and receipts belong to the remaining backlog, not to reopening #12.

The later owner-accepted #15 contract adds a bounded upstream-backed public source producer and explicit native evidence grammar /2, preserving /1. This is not installed Router adoption, a live intelligence service, private intake or product GO.

Documentation

• Architecture: shared ownership, delivered semantics, evidence limits, integration requirements, security/reuse and decision history.
• Roadmap and gap matrix: G01-G13, actual Forgejo tasks, priorities, cross-product dependencies and measurable acceptance sequence.
• Product-value analysis: issue #12 upstream/Router comparison, real cases and accepted narrow public-evidence ADAPT outcome; not product GO.
• Minimum publication contract: completed/closed #13, merged PR #26, accepted payload/reference/critical semantics and exact-pin decision-bound conformance; not a permanent ABI or installed consumer acceptance.
• Native producer extension: owner-accepted #15 /2 representation, selected models.dev/Z.ai source adapter, atomic retained publication and source-health facts. No consumer admission, private calibration, installed adoption or product GO.
• Coding benchmark evidence comparison: a separately scoped, pinned models.dev and Terminal-Bench 4.0 evidence publication for selected coding benchmarks, including Apache-licensed submission provenance, a real source-to-publication receipt and explicit configuration/comparability limits.
• Capability boundary record: completed/closed #14, merged PR #31, versioned ownership research and guarded offline mapping fixtures; not production mapping approval or installed acceptance.
• Optional operational sharing packet: #17 explicit owner REJECT current operational intake, rationale and preserved reconsideration criteria with original synthetic examples; no receiving format, collection, sharing authorization or automatic calibration.
• Offline operator contract: #16 producer-owned first definition, read-only local CLI/status/inspection/export and independently implemented portable consumer conformance. Console requirements are inputs, not an existing-receiver prerequisite; live/browser/installed gates remain open.
• Installation and maintenance: #18 offline source-build/wheel distribution, installed compatibility and recovery. This resource is included in the package; no registry release or deployment.
• Installed explicit-input acceptance: #19 frozen installed MI/Router offline receipt, retained utilization binding and selected-source adoption limitations; no default activation, live/provider acceptance or product GO.
• Agent guidance: owner-selected work, explicit /loop and native bounded PR review.

The target is a local-first, open and observable system optimizing accepted work, including cost, quota, time, remediation, review and owner attention, not cheapest tokens. MI supplies public evidence; Router chooses/adopts knowledge, Kernel owns task authority and outcomes, and Console presents owner-served state. This is a target architecture, not a claim that these integrations already work.

Delivered Proof

evidence.py retains explicit observations, typed claims and supplied revision links. knowledge() exposes retained/latest-known/future versions and stale observations within its supplied evidence set, not a historical acquisition-time cutoff. effective_at() independently selects versions whose explicit effective_from has arrived, then excludes replaced ancestors. Revision boundaries cannot precede their predecessors; equal boundaries permit explicit corrections. Baselines are durable, without finite intervals. Overrides have separate version boundaries and bounded active periods, conditions and zoned wall-time windows; withdrawal is an explicit term-less revision. Future versions do not suppress current predecessors. An effective override revision replaces its predecessor permanently, even after its terms expire.

Effective offers inherit omitted fields and retain current conflicts without choosing authority/recency/order winners or multiplying quotas. Relevant campaign-boundary diagnostics can cite inactive participants without applying their terms; irrelevant expired/future campaigns do not clutter current results. Model benchmarks retain context; execution evidence keeps advertised/observed/selectable/enforceable separate. Unknown is not false and provider-managed execution need not assert a physical model. Versions of one campaign are alternatives, including omitted-field inheritance; different campaigns compose. A guaranteed independent replacement can mask another campaign's uncertain participation without hiding its boundary diagnostic.

evidence_delta() returns new observation/claim identities, including every supplied revision/withdrawal announcement in a batch. Retrieval-only refresh is silent and cannot renew freshness or rewrite semantics; missing history is rejected, not cessation. projection_delta() compares only typed effective endpoint views, exposing before/after terms, participants and conflicts. It does not reconstruct evidence events or invisible causes. A future baseline revision is new knowledge now, but changes current projection only at its effective boundary. Same-term version changes retain participant identities without inventing a price change.

Reuse candidates remain models.dev, AI Model Watch, Model Price Watch, genai-prices and codingplan, with producer/benchmark sources including Artificial Analysis: upstream catalogs/pricing are not recreated. The later bounded benchmark adapter uses only four MIT-licensed models.dev TOML records at an explicit pin; it retains source attribution and does not store raw source pages. No AA data or unlicensed SWE-Marathon leaderboard dataset is redistributed. No upstream license endorsement, official-source accuracy or product GO is asserted by the synthetic architecture proof.

This project does not own credentials, private subscriptions or remaining quota, held reset credits/cards, authenticated runtime state, private workspaces, local GPUs, preferences or task-specific final routing decisions. It is not a downstream decision engine. Reuse existing open components only where semantics and licenses permit; do not construct another generic model catalog. Public module dependencies are allowed; private creatidy-onprem is not required. Completed #16 supplies an offline read-only CLI and full-snapshot export, not a live Console feed, endpoint, event replay or effectful command API.

Forgejo is canonical for source, issues, PRs, reviews and integration history. GitHub is a read-only public mirror.

Public Producer

The narrow source adapters are model_intelligence.zai and the separately scoped model_intelligence.benchmarks, not a catalog fork or service. SOURCES fixes inspected public URI/pin, rights route and measured byte bounds; SCOPE binds its configuration. Every declared source must parse before activation. The caller supplies time, timeout and artifact byte policy; no global TTL, daemon or private credentials are needed.

from datetime import UTC, datetime
from pathlib import Path
from model_intelligence.zai import refresh

directory = Path("public-artifacts")  # Operator-owned local output, one mutator.
directory.mkdir(exist_ok=True)
result = refresh(directory, at=datetime.now(UTC), max_artifact_bytes=65536, timeout=30)

The sample artifact allowance is an example for the measured first selected cut, not a global/ecosystem limit. History can grow; inspect and explicitly adjust policy rather than dropping observations. Failure retains complete bytes, not freshness. Raw official pages are not distributed. Read NOTICE and the source limits before redistribution. Public evidence does not establish entitlement, account quota, local execution support, universal quality or a routing decision.

Local Development

Offline operator views (no source fetching, publication writes or commands):

mi-evidence status public-artifacts --instance public-demo --at 2026-10-08T20:00:00Z --max-bytes 65536
mi-evidence inspect public-artifacts --instance public-demo --at 2026-10-08T20:00:00Z --max-bytes 65536 --json
mi-evidence export public-artifacts --instance public-demo --at 2026-10-08T20:00:00Z --max-bytes 65536

All JSON forms use the same validated snapshot facts. Time and per-input byte policy are explicit operator inputs, not a global TTL/limit. Optional --checkpoint is a retained cut digest, not an event offset: missing history yields a gap, never a fabricated timeline. Source health and active publication may disagree because their files are separate; mismatch/unavailable is exposed rather than called current. Snapshots retain unknown/stale/future/conflicting/withdrawn/expired classifications without source-policy or admission decisions. Human inspection prints the same projection, not new facts. Reads never refresh or modify an artifact.

The portable consumer is original code implemented independently from the producer in a separate context, with no model_intelligence imports. Its strict shape/hash/ reference/version/namespace checks and immutable last-known state exercise malformed, unsupported, duplicate/reordered, reconnect/resync and explicit missing-history cases. This is not actual Console adoption, browser authorization, transport security, installed integration, source authenticity or product acceptance. Effectful commands are empty; #17 REJECT is unchanged. See the contract for compatibility limits.

Requires Python 3.12+, uv and Make.

uv sync --locked --group dev
make check
uv build

Repeatable outside-checkout installed lifecycle validation (prepared offline backend cache required; use a new staging directory):

uv run --no-sync python tests/verify_distribution.py /tmp/kilo/mi-distribution-check --uv-cache /tmp/kilo/mi15-validation/cache/uv

The verifier uses an allowlisted synthetic environment, current and pinned #13/#15 wheels, exact wheel digests, sdist rebuild, real installed CLI, original installed source/update/outage/legacy/rollback/refusal tests and packaged support/NOTICE. It performs no live acquisition, publishing or consumer admission. See the installed support resource for local installation, artifact copying, recovery and limitations.

make check owns locked synchronization, Ruff lint/format, basedpyright, local standard-library tests and Git whitespace checks. Tests use no live services. Standalone development starts from an explicitly Adrian-selected Forgejo issue; see AGENTS.md. Only an explicit owner /loop delegates successive canonical issue selection, exact independently approved PR merge to develop and verified completed-issue closure. Standalone finishing stops at READY_TO_MERGE; main promotion, releases and deployment are not authorized by these commands. Licensed under Apache-2.0.