Nitelos

Verify the outcome, not the attempt.

Nitelos is a Python SDK that checks whether an agent's success criteria are supported by evidence you supply — normalized observations, screenshots, actions and host-authorized external reads. It never acts on the agent's behalf: no clicks, no network fetches, no retries. A result is a time-scoped judgment over explicit criteria, not a transaction guarantee.

What "verified" means here

Verification is scope-relative. A criterion says exactly what final condition must hold, for which subject, at which checkpoint — and Nitelos reports success/failure/abstained for those criteria against that evidence:

• A deterministic assertion (JSON pointer + equals/in/exists over a normalized observation) is fully checked by the SDK: supported, refuted or unknown, with the exact evidence references that decided it.
• An external criterion additionally requires a host-constructed TrustBinding — imported JSON can never authorize its own external claims.
• A visual criterion (optional, experimental) is decided by a local image model only when you explicitly opt in; every such result is labeled model_inferred and carries the VISUAL_GROUNDING_UNVERIFIED limitation.

Missing evidence abstains. Contradictions become unknown. A tool reporting "completed" never proves an outcome.

Quickstart (deterministic, offline)

import hashlib, json
from nitelos import Verifier, canonical_json, parse_request

observation = {
    "id": "obs-form",
    "kind": "observation",
    "sequence": 0,
    "captured_at": "2026-10-20T10:00:00Z",
    "checkpoint_id": "cp-final",
    "source_id": "my-agent-runner",
    "subject": {"form": "registration-1"},
    "phase": "final",
    "channel": "dom",
    "data": {"validation_error": False, "email_filled": True},
}
observation["sha256"] = hashlib.sha256(canonical_json(observation["data"]).encode("utf-8")).hexdigest()

request = parse_request(
    json.dumps(
        {
            "schema_version": "1",
            "request_id": "quickstart-1",
            "objective": "Complete the registration form",
            "checkpoint_id": "cp-final",
            "as_of": "2026-10-20T10:01:00Z",
            "criteria": [
                {
                    "id": "no-errors",
                    "description": "No validation error remains on the form",
                    "kind": "assertion",
                    "scope": "observed_state",
                    "evidence_ids": ["obs-form"],
                    "subject": {"form": "registration-1"},
                    "pointer": "/validation_error",
                    "operator": "equals",
                    "expected": False,
                }
            ],
            "evidence": [observation],
        }
    )
)

result = Verifier().verify(request)
print(result.decision)  # success | failure | abstained (or None on execution errors)

The default is a plain Verifier() with the strict VerificationPolicy (model inference disallowed); screenshots arrive via Evidence.screenshot(...) plus a raw-bytes mapping, and ArtifactStore/TrustBinding cover the image-store and external-trust cases in the guides below. Field names follow the schema-1 VerificationRequest contract.

See examples/quickstart.py for the runnable script and docs/quickstart.md for the walkthrough. The agent-summary demo — a claim of success refuted by a validation_error observation — is examples/demo_incomplete_form.py; the pending-external abstention is examples/demo_external_pending.py.

Deterministic vs experimental

Everything in the quickstart is deterministic and offline: assertions, eligibility (subject/checkpoint/freshness/phase), trust bindings, aggregation and every error code. The optional local visual backend (SmolVLM) is experimental; its executability gate is under review, and it is disallowed by the default policy. No benchmark numbers are published yet — evaluation runs have not been executed.

Documentation

• Installation — core / eval / local extras
• Quickstart
• What verified success means
• Evidence and provenance limits
• Sync/async lifecycle
• Errors and abstention
• Backends and compatibility
• Custom adapter conformance
• Traces and caller collection
• Evaluation and calibration limitations
• Security and privacy
• Performance measurement
• Contributing · Security policy

Apache-2.0. Python 3.12/3.13.