C3R

Robust Calibrated Compute Control for Machine-Native Intelligence

C3R is an open-core control plane that decides which computation is worth performing next. It evaluates tools, retrieval, local and frontier models, verification, placement, and stopping as typed candidates under one conservative value-of-computation policy.

The v0.1 vertical slice includes bounded hierarchical candidate compilation, a calibration-gated System-One seam, and a reproducible DecisionMix v1 schema preview. CLM is the new default System-One provider in code; Laya remains optional. It is a research alpha: the controller is runnable and tested. Fine-tuned weights and empirical calibration remain gates for the empirical model/data release, not for the separate stateless recommendation API described below.

Launch status: Neither the standalone decision service nor governed trace collection is enabled for public use. The independent PR #2 review requests changes. A real cloud storage audit probe and the first natural empty-bucket purge run are recorded for reviewer inspection, but they do not establish deletion of aged traces or backups, trained weights, or calibration for the research release. The separate stateless API still needs its own security, live provider, and canary evidence. See the reviewer packet.

Separate stateless API path

The C3R Core API v1 contract separates a recommendation-only, non-persistent inference service from the governed trace collection and empirical-release program above. The current branch implements the typed CLM API, direct ranking, a text-only Responses subset, a production host, and a fail-closed production_inference mode. A private, loopback-only candidate has returned actual CLM rankings and local DeepSeek text; it is not a publicly launched or production-qualified API. Upstream CLM provides advisory System-One ranking, not generative text or calibrated task-success probabilities. /v1/c3r/execute remains disabled. /v1/responses invokes DeepSeek through an admitted, independently checked text-only controller fallback; it does not claim positive learned CVoC or expose private reasoning. The first public hostname is a dedicated C3R endpoint, not an MC-1 integration. See the scope-specific release policy.

The developer API guide covers SDK examples, scoped keys, tenancy, streaming, limits and privacy for the integration candidate. The branch reconciliation record explains how the production and readiness histories were joined without losing reviewed source.

Default language model

C3R's default deliberative language model is deepseek-ai/DeepSeek-V4.1-Flash (MIT), self-hosted on the existing eight-H100 node alongside Qwen3-8B/CLM. The default provider uses the private loopback /model alias; no OpenRouter or hosted inference provider is part of the production target. CLM is the default separate System-One decision engine; Laya remains optional. DeepSeek recommendations remain subject to the same verifier and trusted commit boundary as every other candidate.

The official checkpoint has passed a private 8×H100 GCP serving smoke test and is backed up in a private GCS bucket. Its vLLM endpoint is bound to localhost; this is not a public C3R service or end-to-end production qualification. The checkpoint's active parameter count does not imply it fits on one GPU.

The recovered older serving image passed all 13 private C3R checks at 85% GPU reservation with Qwen still running. The clean, pinned newer image is a separate qualification target: its first startup rejected an obsolete flag. The corrected launch configuration is tracked in deploy/deepseek-v41. Repeated cold starts, sustained mixed load, outage/rollback, final-head builds, TLS and canary remain required; recovery alone is not production qualification.

Why C3R

Most agent stacks decide what to say. C3R decides what computation should happen next—and keeps that recommendation separate from authority to act.

flowchart LR
    S[Versioned state] --> C[Hierarchical candidate compiler]
    C --> L[CLM default / Laya optional]
    C --> D[Deliberative envelope]
    L --> V[Robust CVoC]
    D --> V
    V --> F[Verifier Firewall]
    F --> G[Trusted Commit Gateway]
    G --> O[Outcome + trace + cost twin]

Loading

Hard masks run before candidate expansion. Missing calibration or an uncertain typed prediction abstains. Learned components cannot grant permissions, select their authoritative verifier, turn failed verification into success, or directly commit an external effect.

What ships in this slice

Surface	Included now
Candidate Compiler	family → subgroup → operation → arguments → placement → verifier; hard masks, budget pruning, caps, progressive widening
System-One fast path	CLM loopback rank adapter with strict response checks and calibration-gated abstention; optional revision-verified Laya
DecisionMix v1	validated records, immutable deterministic splits, source/license provenance, SHA-256 manifest, 144-record synthetic preview
Authority boundary	action-bound verifier attestations, expiring single-use approvals, atomic nonce claims
Runtime control	conservative CVoC selection, deterministic STOP, cost twin, deliberative contracts, trace schema and optional transactional SQLite hash chain
Operational controls	fail-closed feature flags, tested frontier/open-weight HTTP contracts, Colibri shadow recommendations, canonical trace hash chain
Standalone controller boundary	tested state → candidates → optional System-One → CVoC → independent verifier → read-only recommendation or deterministic fallback → redacted trace composition; the standalone controller rejects external executors until effects and durable trace commits can be made atomic. A separate private Cloud Run staging host is fixed-disabled; it is not the decision service or a public launch.

This compiler is the reviewed vertical slice, not the directive's full Candidate Compiler Definition of Done. Rich typed value constraints, per-argument provenance, dominated-branch pruning, and mandatory production placement policy remain explicit roadmap gates.

Quick start

Core tests require only Python 3.11+:

python -m unittest discover -s tests -v

Install the optional pinned Laya integration:

pip install -e ".[laya]"

Build the identical DecisionMix schema preview and verify its hashes:

python scripts/build_decisionmix_preview.py

Minimal conservative selection:

from c3r.cvoc import RobustCvocController
from c3r.state_schema import ActionCandidate, ActionFamily, RiskClass, ValueEstimate

candidate = ActionCandidate(
    id="retrieve",
    family=ActionFamily.RETRIEVAL,
    risk_class=RiskClass.READ_ONLY,
    optimistic_utility=0.7,
)
decision = RobustCvocController().select(
    (candidate,),
    {"retrieve": ValueEstimate(0.8, 0.2, 0.0, 0.1)},
)

If the conservative lower bound is not positive, decision.selected is None and the fallback is STOP.

Launch artifacts

• C3R v0.1 Hugging Face collection — the model preview, DecisionMix preview, and attributed upstream Laya checkpoint in one release collection.
• C3R DecisionMix v1 Preview — published synthetic schema dataset with immutable splits and content hashes; its source is mirrored in data/decisionmix-v1-preview.
• C3R Decision Laya 421M v0.1 — integration-preview model card, exact base revision, both research papers, and machine-readable calibration/training/evaluation gates; no derived weights are claimed.
• docs/architecture.md — trust boundaries and component contracts.
• docs/roadmap.md — what remains before production qualification.
• docs/directive-compliance.md — evidence-backed audit against every Definition of Done item in Execution Directive v2.
• docs/empirical-release-plan.md — gated path from synthetic preview to trained, calibrated, independently reproducible release.
• docs/standalone-launch.md — current production exit gates, evidence status, and operator inputs, with MC-1 excluded from standalone scope only.
• docs/launch-announcement.md — canonical launch copy plus LinkedIn, X, and Hacker News variants with a publication checklist.
• docs/prior-art.md — explicit attribution links and the canonical novelty boundary required by the execution directive.

CLM System-One integration

C3R_SYSTEM_ONE_PROVIDER=clm is the configuration default, while C3R_ENABLED and C3R_SYSTEM_ONE still default to off. The adapter targets the upstream Contrastive-LM/CLM /v1/rank API on a loopback-only origin. Its source is pinned at bb42c6c5bf914fd449bed2f6ca65be80602cb1f7 (Apache-2.0). The running encoder and CLM head need their own immutable artifact revision, passed as C3R_CLM_ARTIFACT_REVISION; the current adapter validates the declaration's format but does not yet attest the live server's artifact hash. This repository does not bundle or claim trained C3R-specific CLM weights. The host supplies C3R_CLM_URL (default http://127.0.0.1:8700), an optional C3R_CLM_API_KEY, and a fitted TemperatureCalibrator to build_default_clm_fast_path. The initial C3R_CLM_TIMEOUT_MS=500 bounds the complete System-One decision; production latency thresholds still require live measurement.

CLM ranks bounded, policy-surviving candidate labels and fixed typed questions. Its raw ranking is advisory, not an action selection. Missing calibration, malformed responses, outages, or timeouts lead to abstention or deterministic fallback. CVoC, independent verification, and the commit boundary retain authority. Private live CLM ranking and co-resident DeepSeek recovery have been observed; neither sustained GPU coexistence qualification nor held-out C3R calibration is claimed by this code change.

Laya integration

Laya remains an explicitly selected comparison/compatibility provider; it is not the default System-One engine.

The adapter pins convaiinnovations/laya at 1c5edc17a7acd8701df6fc341c0d179f1c62c982. Before loading, the backend resolves the Hub metadata, verifies that exact SHA and the Apache-2.0 license, downloads that revision, and passes the local snapshot to laya==0.3.5.

The fast path answers fixed typed questions only. Every decision slice is keyed by question type, action family, option-count bucket, language, and consequence class. A missing temperature, insufficient top probability, or insufficient top-two margin returns an abstention signal that the host orchestration must route to its deliberative envelope.

DecisionMix v1

The published preview is intentionally synthetic. It validates the entire publication contract without presenting generated fixtures as real training evidence. The empirical corpus will ship only when provenance, licensing, held-out integrity, and calibration support are independently auditable.

For the first empirical source, ColomboAI approved only C3R-authored internal tasks under the interim trace policy. It sets a 30-day private retention limit and requires independent review before any de-identified row is published. Collection remains off until the technical controls are verified; this approval does not make the preview empirical.

Required empirical metrics include accuracy, Brier score, ECE, maximum calibration error, NLL, selective risk versus coverage, abstention, escalation, p50/p95 latency, throughput, calls avoided, and cost per completed task.

Feature flags

Production hosts must preserve independent control of:

C3R_ENABLED
C3R_SYSTEM_ONE
C3R_SYSTEM_ONE_PROVIDER=clm
C3R_DELIBERATIVE
C3R_ROUTING
C3R_SPECULATION
C3R_MOE_CONTROL
C3R_ONLINE_LEARNING=false

Disabling the learned fast path must leave the host's normal deterministic fallback operational. FeatureFlags.from_mapping enforces those switches fail-closed and rejects autonomous online learning. The reference provider and Colibri shadow contracts are documented in docs/runtime-integrations.md.

Release truth

Not yet claimed: C3R-trained CLM heads or Laya weights, empirical DecisionMix training data, live provider qualification, MC-1 product integration, a Colibri shadow deployment, or measured production calibration/latency/cost results. Tested adapter and shadow-control contracts are included, but they are not represented as production runs. These remain documented gates.

The upstream base is Laya by Convai Innovations, licensed Apache-2.0. C3R is a broader runtime architecture, not a fork or rebranding of Laya. See NOTICE for the attribution boundary.

Security and license

Do not report vulnerabilities in a public issue; follow SECURITY.md. Production effects must be completely mediated by an independently configured commit gateway.

Apache License 2.0. See LICENSE. If you use C3R, cite CITATION.cff.