AI Native Lang (AINL) Glossary

OpenClaw (MCP skill): ../OPENCLAW_INTEGRATION.md — skills/openclaw/, ainl install-openclaw, ~/.openclaw/openclaw.json (mcp.servers.ainl), stdio ainl-mcp. (Distinct from OpenClaw bridge cron/memory under openclaw/bridge/.)

ZeroClaw: Host integration via ZeroClaw skill, ainl install-zeroclaw, and ainl-mcp — see ../ZEROCLAW_INTEGRATION.md (distinct from OpenClaw bridge/cron docs).

AINL

AI Native Lang. A language designed for AI-first authoring and execution workflows.

Canonical AINL

The strict, line-oriented AINL form used for validation, training targets, and evaluation.

IR (Intermediate Representation)

The compiler output graph/step structure used by runtimes and emitters.

Graph-First Runtime

Execution mode that prioritizes graph semantics and uses step fallback only when needed.

Adapter

A runtime integration surface for external capabilities (HTTP, DB, files, tools, etc.).

LoRA

Low-Rank Adaptation fine-tuning strategy used to adapt base models efficiently.

Strict AINL Rate

Share of prompts where output is non-empty, AINL-like, and passes strict compile checks.

Runtime Compile Rate

Share of outputs that pass runtime (non-strict) compile validation.

Nonempty Rate

Share of prompts producing non-empty output.

Constraint Diagnostics

Telemetry emitted during constrained decoding (allowed/rejected tokens, fallback, EOS gating).

Failure Family

A normalized category of generation failure (timeout, shape mismatch, compile failure, etc.).

Prompt-Length Bucket

Grouping prompts by rendered token length to improve shape stability and analysis granularity.

Checkpoint Sweep

Evaluating checkpoints and ranking them by task metrics instead of raw eval loss.

Trend Gate

Cross-run quality gate enforcing minimum rates and maximum allowed regressions.

Run Health Report

Machine-readable pass/fail summary artifact for automation: corpus/curated/alignment_run_health.json.

Distill Mix

Training-data composition strategy mixing gold examples and repair/check-rewrite supervision.

Failure Boost Dataset

Targeted dataset generated from failing prompt IDs to improve weak families.

OpenClaw

An extension and operator-focused surface area built on top of canonical AINL, used for advanced adapters, orchestration, and multi-agent workflows. MCP skill + bootstrap: ../OPENCLAW_INTEGRATION.md (skills/openclaw/, ainl install-openclaw, ~/.openclaw/openclaw.json, ~/.openclaw/bin/ainl-run). Bridge / cron / daily markdown memory: openclaw/bridge/ and ../operations/UNIFIED_MONITORING_GUIDE.md.

ZeroClaw

A separate MCP-first onboarding path for AINL: ZeroClaw skill in-repo, ainl install-zeroclaw (~/.zeroclaw/mcp.json, ainl-run shim), and stdio ainl-mcp. Documented in ../ZEROCLAW_INTEGRATION.md; does not replace OpenClaw openclaw/bridge/ cron/memory layouts.

Adapter Registry

The catalog of available runtime adapters, their identifiers, and supported capabilities as defined in docs/reference/ADAPTER_REGISTRY.md.

Memory Adapter

An adapter responsible for durable memory operations (read, write, list, delete) over structured memory records, specified in docs/adapters/MEMORY_CONTRACT.md.

Memory Record

A structured, schema-validated unit of persisted memory (with fields such as id, kind, timestamps, and content) managed by the memory adapter.

Conformance Profile

A named set of behavioral and surface-area expectations (e.g., AINL_V0_9_PROFILE) that a runtime, compiler, or adapter must meet to be considered compatible.

Language Lanes and Extensions

The division between canonical AINL and additional extension lanes (including OpenClaw-focused extensions) described in docs/AINL_CANONICAL_CORE.md and docs/language/AINL_EXTENSIONS.md.

Runtime Specification

The formal description of how compiled IR and language constructs must behave at runtime, complementing the human-readable semantics; see docs/ainl_runtime_spec.md and SEMANTICS.md.

Targets Roadmap

The forward-looking map of officially supported runtimes, adapters, and deployment targets, documented in docs/runtime/TARGETS_ROADMAP.md.

State Discipline

The AINL approach to managing workflow state through explicit, tiered adapters (frame, cache, memory/SQLite/FS, coordination) rather than hiding state in prompt history. Described in docs/architecture/STATE_DISCIPLINE.md.

State Tier

One of four levels of state durability in AINL: Tier 1 (frame, ephemeral), Tier 2 (cache, short-lived), Tier 3 (memory/SQLite/FS, persistent), Tier 4 (queue/agent, cross-workflow coordination).

Policy Validator

A pre-execution gate (tooling/policy_validator.py) that checks compiled IR against a declarative policy (forbidden adapters, effects, effect tiers) and returns structured violations. Can be invoked directly or via the optional policy parameter on the runner service /run endpoint.

Capability Discovery

The GET /capabilities endpoint on the runner service, which returns available adapters, their verbs, support tiers, default effects, and the runtime version. Used by external orchestrators to inspect what an AINL runtime instance supports before submitting workflows.

Runner Service

The FastAPI-based execution service (scripts/runtime_runner_service.py) exposing /run, /enqueue, /result/{id}, /capabilities, /health, /ready, and /metrics endpoints for runtime-facing workflow execution.