Context Layer

Version Roadmap

Overview

Context Layer contracts are versioned. Modifications to behavioral guarantees require explicit version increments and will be documented here.

v1.0 Runtime Foundations

Execution Authority

Admission, constraint enforcement, and project-scoped execution isolation.

Flow and Pulse Modes

  • Flow: Structured workflow execution. Sequential step ordering. Explicit termination. Authority Report on termination.
  • Pulse: Conversational execution. Constraint enforcement. No Authority Report.

Authority Reports

Flow workflows produce Authority Reports upon session termination. Session-scoped metadata and execution trace. Retrievable via runtime API.

BYOK Provider Model

Provider credentials are configured in the console and stored securely. Provider configuration is applied at runtime. Context Layer does not persist or log provider keys. Developers retain full control of credentials.

v1.1 Verification Authority

Verification

Context Layer v1.1 introduces output verification as a runtime guarantee. Every execution is verified against owned context. Violations are reported authoritatively by the runtime, not inferred from model output.

Verification Dimensions

Five dimensions are evaluated on every execution:

  • Instruction coverage
  • Constraint compliance
  • Cross-turn consistency
  • Hallucination and leakage detection
  • Explanation reliability

Verification Report

Every response includes a structured verification report with per-dimension results, a tier decision, and a trust contract status.

See Verification for full details.

Pulse Memory Contract

Pulse sessions enforce a deterministic conversational memory contract. Session state is stable and reproducible across executions.

Benchmark

Context Layer v1.1 achieves 81.7% success on strict multistep agentic workflows. Baseline without Context Layer: 7%. Same model. Same task. Same verifier. 120 runs.

v1.1.1 State Bridge

State Bridge

Context Layer v1.1.1 ships cl-state-bridge.js as an optional Flow utility. State Bridge converts structured output from any external system into a natural language string for invokeCL(). It is not required for standard integrations.

buildCLMessage

Converts structured data into a natural language string. Accepts booleans, strings, numbers, objects, arrays, and nested structures. Flattens nested data with dot notation keys.

assertCLResult

Validates that the Context Layer response contains expected outcome strings. Throws a descriptive error if any expected outcome is missing.

Availability

Flow mode only.

See State Bridge for full documentation.

v1.2 Reasoning Contamination Control

Authority Claim

Context Layer v1.2 detects and corrects contaminated reasoning within a session.

Reasoning Contamination Control

v1.2 introduces runtime detection of contaminated reasoning: outputs that contradict owned context due to interference from prior session history. Contamination is detected deterministically by the runtime.

Entanglement Skepticism

Context Layer evaluates trigger conditions after every Flow step. When contamination signals are detected, the execution is flagged as potentially entangled and an entanglement report is included in the response.

Fresh-Slate Verification

On every Flow step, Context Layer re-evaluates the output against owned constraints and canonical context, with prior session history excluded. This produces a divergence signal used to detect contamination.

Verification Receipt Upgrades

Every execution receipt now surfaces:

  • model_self_diagnosis_conflict: whether the model's self-explanation conflicted with system verification
  • authoritative_failure_causes: system-derived failure attribution, independent of model explanation

Benchmark

Context Layer v1.2 achieves 81.7% success on strict multistep agentic workflows. Baseline without Context Layer: 7%. Same model. Same task. Same verifier. 120 runs.

v2.0 Active constraint enforcement

Authority Claim

At each workflow step, Context Layer enforces the constraints that apply to that step and keeps later steps accountable to constraints introduced earlier, without treating them as inert context.

Enforcement across steps

v2.0 enforces constraints accumulated across workflow steps at each step. Constraints are not held as passive context. When output violates an active constraint, the runtime detects and corrects it before the output is returned.

Interaction and Adversarial Risk Evaluation

Conversational sessions are evaluated for adversarial intent and interaction risk before execution proceeds.

Stale Context Detection

The runtime identifies when canonical context has drifted from its source constraints before it reaches the model.

Constraint-Specific Repair

When a constraint is violated, the runtime identifies the exact constraint that was violated and what the output needs to contain.

Benchmark

Context Layer v2.0 was benchmarked across two realistic multi-step engineering workflows across 280 total runs. Baseline without Context Layer: 16.4%. With Context Layer v2.0: 97.1%. Same model. Same prompts. Same verifier.

v2.1 Content Freshness Enforcement

Authority Claim

Context Layer v2.1 enforces context freshness at runtime.

Content Freshness Enforcement

v2.1 completes the content freshness story started in v2.0. When canonical context has drifted from its source constraints, the runtime enforces on it. The output is still returned but the trust signal is degraded. Staleness is recorded in the execution receipt.

What Stale Context Means

A canonical summary becomes stale when the constraint it was written against has been updated to a newer version. The summary reflects an older constraint state. The runtime detects this before the model is invoked and flags the execution.

Benchmark

Context Layer v2.1 benchmark numbers are identical to v2.0. Content Freshness Enforcement affects trust signal quality, not output generation. The 97.1% pass rate across 280 runs remains the current benchmark.

v2.2 Durable Constraint State, Unified Execution, and MCP

Authority Claim

Durable constraint state is reliable across the runtime, and execution is unified and externally accessible.

Durable Constraint State

Constraint activation state is now durable and consistent across the runtime. Constraints established during a session persist reliably through the full workflow lifecycle and are cleared on session termination.

Unified Execution Interface

All execution flows are routed through a single endpoint: POST /api/execute. Flow and Pulse requests are mode-routed internally. This replaces the separate /api/flow and /api/pulse endpoints.

Agent-Compatible Flow Execution

Flow now supports continuous multi-step execution without explicit termination. Sessions persist across calls and accumulate state. Agent sessions are finalized automatically after inactivity. See Usage Type for configuration.

Automatic Session Finalization

Agent sessions that go inactive for approximately two hours are finalized automatically. The runtime injects workflowEnd, generates the Authority Report, and terminates the session.

MCP Interface

Context Layer is accessible via MCP at https://mcp.cl.kaisek.com/sse. External agents can invoke Flow execution through the send_message tool without a wrapper script. See MCP documentation.

Benchmark

Context Layer v2.2 benchmark numbers are identical to v2.0 and v2.1. The unified execution path and durable constraint state do not affect output generation. The 97.1% pass rate across 280 runs remains the current benchmark.