qlm-measure

Open measurement SDK for education AI — evidence events, record verification, and dataset clients.

Apache-2.0 · Documentation · JS package

Limitations (read first)

• This SDK is the public interface layer, not the measurement engine. The estimation service (Bayesian posteriors, calibration) is QLM's hosted engine.
• The verifier checks bookkeeping integrity, not estimation correctness. It can tell you if a record was tampered with. It cannot tell you if the posteriors are accurate. That distinction is the design.
• Evidence event schemas are v0.2 and may change in minor versions. Pin your dependency.
• verify makes no network calls. Records never leave your machine.

What the verifier deliberately cannot check

• Whether posteriors are correctly computed (that requires the engine's update rule; see ESTIMATE.REPRODUCE, planned)
• Whether evidential weights are appropriate (those are calibrated server-side)
• Whether triage classifications are accurate (those use rules not in this package)

Install

pip install qlm-measure

Quickstart — produce a record from your own tool

See docs/recorder-quickstart.md — ten lines, no QLM account, no network.

Verify a file

qlm-measure verify record.json
qlm-measure verify records.jsonl --format json --report report.json
qlm-measure explain HASH.ENTRY_RECOMPUTE
qlm-measure version

Exit codes: 0 all valid, 1 at least one invalid, 2 usage or parse error.

Report formats

Text (default):

record syn-8f2c-0042  (schema 0.2, 6 entries)
  Schema
    ✓ Required fields present
    ✓ Field types valid
    ✓ Timestamp is ISO 8601
  ...
  VERDICT: CLEAN — record is verifiable

JSON (--format json or --report FILE):

{
  "tool": {"name": "qlm-measure", "version": "0.2.2", "language": "python"},
  "summary": {"records": 1, "valid": 1, "invalid": 0, "checks_run": 9, "checks_passed": 9}
}

Verify in code

from qlm_measure import verify_record

record = {"studentScopeId": "...", "entries": [...]}
result = verify_record(record)
if not result.valid:
    for v in result.violations:
        print(f"[{v.check_id}] v{v.version}: {v.message}")

Check catalog

10 shipped checks, run by verify:

5 planned checks (listed by explain, not yet run): ESTIMATE.REPRODUCE, TEMPORAL.T_DESIGN, TEMPORAL.T_RESULT, CONSENT.SCOPE, CONSENT.WITHDRAWAL

Use qlm-measure explain CHECK_ID for details on any check.

File formats

• .json — one record object, or an array of records
• .jsonl / .ndjson — one record per line
• - — stdin (format auto-detected)
• UTF-8, tolerates BOM. Parse errors report line number.

Privacy

• studentScopeId is pseudonymous by contract. Never include PII.
• --redact replaces scope IDs with their SHA-256 prefix in output.
• The SDK does not store data, make network calls, or set cookies.

Links

• JS package: npm install qlm-measure
• Developer portal
• Open Measurement Manifesto

License

Apache-2.0. See LICENSE.