Safety Gates¶

Traigent includes several fail-closed or explicit-opt-in gates around real optimization, CI runs, dataset paths, backend validation, and persisted example content.

Cost Coverage Preflight¶

Real runs check model cost coverage before trials start. Unknown or unpriced models are detected from the configuration space or current/default config.

With TRAIGENT_STRICT_COST_ACCOUNTING=true, unpriced models always raise UnknownModelError before any trial and no prompt is shown.

When the variable is unset, runs whose objectives include cost are strict at runtime only: this preflight still follows the approval rules below, because it runs before any response exists and cannot see provider-reported costs such as OpenRouter's. An approved run then stops at the first call that has neither a table price nor a reported cost, instead of recording $0 for it. Set TRAIGENT_STRICT_COST_ACCOUNTING=false to keep recording $0 with a warning.

To proceed after acknowledging that affected results will report $0 cost while your provider may still bill you, use one explicit approval:

pass cost_approved=True as a real Python boolean; strings such as "true" are ignored,
set TRAIGENT_COST_APPROVED=true; 1, yes, and on do not approve this gate,
or create the XDG cost approval token (approved or approved:<limit>).

Approved runs still emit the unpriced-model warning. Mock LLM quickstart mode does not skip this model-pricing preflight (raw provider calls can still be real under the mock flag); pass cost_approved=True or set TRAIGENT_COST_APPROVED=true for a keyless demo that uses unpriced model names.

To fail before any trial starts, enable strict cost accounting:

export TRAIGENT_STRICT_COST_ACCOUNTING=true

Provide explicit pricing for private or unsupported models with:

export TRAIGENT_CUSTOM_MODEL_PRICING_JSON='{"my-model":{"input_cost_per_token":0.000001,"output_cost_per_token":0.000002}}'
export TRAIGENT_CUSTOM_MODEL_PRICING_FILE=.traigent/model-pricing.json

CI Run Approval¶

Optimization in CI requires explicit approval. The recommended environment approval is:

export TRAIGENT_RUN_APPROVED=1
export TRAIGENT_APPROVED_BY="your_name"

TRAIGENT_MOCK_LLM does not bypass this CI approval gate.

Mock Mode and the Cost Preflight¶

Mock mode alone does not skip the pre-run cost approval or the model-price coverage preflight, so an agent that mixes real and mock calls is still checked. Keyless walkthroughs and tests that are genuinely fully mocked opt out with TRAIGENT_COST_PREFLIGHT_MOCK_EXEMPT=true. The exemption applies only when mock LLM mode is on and the environment (ENVIRONMENT, TRAIGENT_ENV or TRAIGENT_ENVIRONMENT) is explicitly non-production, e.g. ENVIRONMENT=development. An unset or unknown environment counts as production and never gets the exemption. It is test/walkthrough-only and must not be set for real-LLM runs. Runtime cost accounting stays active.

Dataset Path Containment¶

Evaluation datasets must resolve under the trusted dataset root. Set the root explicitly when a process should only read datasets from one directory:

export TRAIGENT_DATASET_ROOT="$PWD/evals"

When TRAIGENT_DATASET_ROOT is not set, the current working directory is the trusted root.

Backend Validation Loopback Gate¶

Loopback backend validation is default-closed. For local backend development, use an explicit non-production environment:

export ENVIRONMENT=development

Do not use this override for production.

Content Logging Opt-Out¶

Per-example prompt, response, and expected-output content is persisted by default for local developer experience. To keep IDs and metrics while omitting content-bearing fields from disk:

export TRAIGENT_LOG_EXAMPLE_CONTENT=false

Copy-Paste Example¶

export TRAIGENT_STRICT_COST_ACCOUNTING=true
export TRAIGENT_DATASET_ROOT="$PWD/evals"
export TRAIGENT_LOG_EXAMPLE_CONTENT=false

Honesty note: Traigent cost limits and estimates are local guardrails, not provider billing caps. Set provider-side billing limits in your LLM or cloud account for hard spend protection.