Environment Variables

Overview

AspectDetails
PurposeEnvironment-level toggles for network access, evaluation, snapshots, and docs tooling.
AudienceCLI users and operators tuning runtime behavior.
ScopeCLI commands and programmatic runs; precedence is setting-specific when env and config/CLI both exist.
NetworkOffline by default; network must be explicitly enabled.
Source of truthdocs/reference/env-vars.md, src/invarlock/cli/commands/*, src/invarlock/core/plugins_inventory.py, src/invarlock/runtime_security.py, src/invarlock/core/runner.py.

Quick Start

# Allow model + dataset downloads for a single command
INVARLOCK_ALLOW_NETWORK=1 invarlock evaluate --baseline gpt2 --subject gpt2

# Force evaluation device for a one-off compare/evaluate run
INVARLOCK_EVAL_DEVICE=cpu INVARLOCK_ALLOW_NETWORK=1 \
  invarlock evaluate --baseline gpt2 --subject gpt2 --device cpu

Concepts

  • Offline-first: all network access is opt-in and must be explicitly enabled.
  • Precedence: when a setting exists in both env + config/CLI, the winner is setting-specific (see the matrix below).
  • Auditability: selected env flags are recorded in report.meta.env_flags for traceability.

Precedence (conflict cases)

  1. CLI/config values for assurance-critical policy (strictness, drift/acceptance bands, overhead skip).
  2. Env overrides only for explicitly env-scoped toggles (for example, downloads, calibration materialization, and tiny-relax smoke behavior).
  3. Packaged defaults when no explicit setting exists.

Key override matrix

SettingEnv varConfig/CLIWinner ruleHow to confirm
Calibration materializeINVARLOCK_ALLOW_CALIBRATION_MATERIALIZEcontext.eval.materialize_calibration / context.eval.allow_iterable_calibrationEnv wins.Config shows in report.context; env is not recorded.
Tiny relaxINVARLOCK_TINY_RELAXcontext.run.tiny_relaxEither opt-in enables tiny-relax run/report policy.Recorded in run context and surfaced through report validation.
Network downloadsINVARLOCK_ALLOW_NETWORKEnv-only toggle.Not recorded; rely on env.
Offline datasetsHF_DATASETS_OFFLINEEnv-only toggle.Not recorded; rely on env.

Conflict examples

ScenarioResultFix
context.run.skip_overhead_check: true in --profile releaseOverhead check is skipped and recorded in guard_overhead.source.Set context.run.skip_overhead_check: false for full overhead enforcement.
context.run.tiny_relax: trueTiny-relax gating is enabled from config and recorded in auto.tiny_relax.Remove or set to false for full policy strictness.

Reference

Network & data

VariableDefaultPurpose
INVARLOCK_ALLOW_NETWORKunsetEnable outbound downloads for models/datasets.
HF_DATASETS_OFFLINEunsetForce Hugging Face datasets to use local cache only.

Model loading

VariableDefaultPurpose
INVARLOCK_ALLOW_REMOTE_CODEunsetExplicitly allow remote model code execution.

INVARLOCK_ALLOW_REMOTE_CODE is the public environment gate for remote model code execution. Use INVARLOCK_ALLOW_REMOTE_CODE=1 for invarlock evaluate when remote code is required; --allow-remote-code is exposed on advanced calibration/config-runner commands that load models directly.

Evaluation & pairing

VariableDefaultPurpose
INVARLOCK_TINY_RELAXunsetDev/demo compare-evaluate override for tiny smoke runs; records tiny-relax provenance and applies tiny-relax report-policy semantics. Do not use for production assurance.
INVARLOCK_EVAL_DEVICEunsetForce evaluation device (cpu, cuda, mps).
INVARLOCK_STORE_EVAL_WINDOWS1Store token windows in reports (set 0 to disable).
INVARLOCK_ALLOW_CALIBRATION_MATERIALIZEunsetAllow materializing iterables lacking __len__.

Bootstrap CI method and replicate counts are controlled by runtime profile, tier policy, and report policy. There is no public env var that forces BCa.

Dataset preparation

VariableDefaultPurpose
INVARLOCK_CAPACITY_FASTunsetApproximate capacity estimation for quick runs.
INVARLOCK_DEDUP_TEXTSunsetExact-text dedupe before tokenization.
INVARLOCK_HF_DATASETS_CACHEunsetOverride the writable fallback cache used when HF dataset loads hit a shared-cache lock/permission error.

Determinism & performance

VariableDefaultPurpose
INVARLOCK_OMP_THREADS1Thread caps for determinism preset.
INVARLOCK_DEBUG_TRACEunsetVerbose debug traces for data/eval paths.
INVARLOCK_LIGHT_IMPORTunsetAvoid heavy imports for docs/tests.
PACK_DEFER_REPORT_RENDERINGunset (1 under run_pack.sh --release-review)Evidence-pack wrapper toggle that skips optional markdown/reviewer rendering during evaluation.
PACK_DEFER_OPTIONAL_REPORT_RENDERINGunsetAlias for PACK_DEFER_REPORT_RENDERING.

Evidence-pack evaluation-loop toggles are repo-wrapper controls, not public invarlock evaluate defaults. Required JSON reports and sidecars are still written; PACK_DEFER_REPORT_RENDERING=1 skips optional rendered review files in the evaluation hot path.

Checkpointing & snapshots

VariableDefaultPurpose
INVARLOCK_SNAPSHOT_MODEautoauto, bytes, or chunked snapshot strategy.
INVARLOCK_SNAPSHOT_AUTO_RAM_FRACTION0.4RAM fraction threshold for auto mode; the lower of this and the absolute threshold selects chunked snapshots.
INVARLOCK_SNAPSHOT_THRESHOLD_MB768Absolute model-size threshold for chunked snapshots, used even on high-RAM hosts.

Model export

VariableDefaultPurpose
INVARLOCK_EXPORT_MODELunsetEnable HF export during model-export capable CLI flows.
INVARLOCK_EXPORT_DIRunsetTarget directory for model export.

Guarding & evidence

VariableDefaultPurpose
INVARLOCK_ASSERT_GUARDSunsetEnable guard runtime assertions.
INVARLOCK_EVIDENCE_DEBUGunsetEmit guards_evidence.json for audit.

Primary-metric gate bounds are profile/config settings (primary_metric.acceptance_range and primary_metric.drift_band), not environment overrides. Strictness/tiny-relax/overhead-skip are also config/profile policy: context.eval.strict / context.eval.strict_errors, context.run.strict_guard_prepare, context.run.tiny_relax, context.run.skip_overhead_check.

Config loading

VariableDefaultPurpose
INVARLOCK_CONFIG_ROOTunsetOverride packaged runtime/ data.
INVARLOCK_ALLOW_CONFIG_INCLUDE_OUTSIDEunsetAllow YAML !include outside config dir.

Reporting & telemetry

VariableDefaultPurpose
INVARLOCK_TELEMETRYunsetEmit single-line telemetry summary.

Plugins

VariableDefaultPurpose
INVARLOCK_ALLOW_THIRD_PARTY_PLUGINSunsetEnable third-party plugin discovery.
INVARLOCK_MINIMALunsetShow minimal plugin list in invarlock advanced plugins.

Runtime enforcement

VariableDefaultPurpose
INVARLOCK_ALLOW_HOST_EXECUTIONunsetAdvanced/internal host-execution override. Prefer invarlock evaluate --execution-mode host for the public compare/evaluate path.
INVARLOCK_CONTAINER_EXECUTIONunsetInternal recursion guard marking runtime-container execution.
INVARLOCK_CONTAINER_ENGINEunsetForce the OCI engine used for default runtime-container execution (podman or docker).
INVARLOCK_RUNTIME_IMAGEunsetOverride the OCI image used for containerized model execution.
INVARLOCK_RUNTIME_IMAGE_DIGESTunsetSupply the immutable digest recorded into runtime.manifest.json.
PACK_RUNTIME_IMAGE_FLAVORdefaultRemote evidence-pack setup helper image selector. Use quant on CUDA hosts to build/use invarlock-runtime:cuda-quant for containerized hf_bnb, hf_gptq, hf_awq, hf_torchao, hf_hqq, hf_quanto, and hf_ct evidence. The quant image uses a pinned CUDA devel base so GPTQModel can JIT-compile kernels with nvcc; strict custom-image evidence still requires INVARLOCK_RUNTIME_IMAGE_DIGEST.

Docs build

VariableDefaultPurpose
INVARLOCK_DOCS_MERMAIDunsetEnable Mermaid diagrams in MkDocs.

Troubleshooting

  • Downloads blocked: set INVARLOCK_ALLOW_NETWORK=1 and retry.
  • Multiple container engines installed: set INVARLOCK_CONTAINER_ENGINE=podman or INVARLOCK_CONTAINER_ENGINE=docker.
  • HF dataset cache lock/permission errors on local reruns: set INVARLOCK_HF_DATASETS_CACHE=/path/to/writable/cache or let InvarLock retry under its own writable cache.
  • Calibration iterables fail: use INVARLOCK_ALLOW_CALIBRATION_MATERIALIZE=1.
  • Third-party plugins missing: set INVARLOCK_ALLOW_THIRD_PARTY_PLUGINS=1; advanced plugin/calibration commands also expose --allow-third-party-plugins.

Observability

  • report.meta.env_flags records selected env toggles.
  • reports capture telemetry and policy digests derived from these flags.