vajelski/georgian-booking-safety-examples

Synthetic Georgian booking-safety examples

Twelve hand-annotated design examples. Not an ASR/NLU benchmark, a training corpus, or a measurement of any deployed product.

This package makes the OMO AI educational examples available as a small, flat JSONL table with a complete, lossless copy of each original case. The source material and this packaging were prepared with AI assistance. All utterances, identities, times, state and destination responses are fictional. There are no customer recordings, personal records or production call logs.

Intended use

Use the examples to discuss an application's action boundaries: which current offer was confirmed, whether an availability observation is stale, whether the speaker's words are quoted rather than direct consent, and what to do when a write outcome is unknown.

These are examples of manually specified expectations, not empirical evidence. They are not representative of Georgian speakers or businesses, and do not establish language coverage, fairness, recognition accuracy or production safety. The English text was prepared for explanation; it has not been independently validated as a translation benchmark.

Data structure

One configuration (default), one examples split, twelve rows:

Nested evidence is retained inside source_case_json, not inferred from the utterance. Parse it with json.loads to restore the complete original object. The original top-level policy and provenance remain in booking-safety-cases.json.

The labels are: 9 no_write, 1 create_once, 1 reconcile, and 1 handoff. create_once is permission for a hypothetical guarded attempt, not proof that a booking was made. handoff does not prove delivery to an operator.

Read without extra dependencies

import json
from pathlib import Path

rows = [
    json.loads(line)
    for line in Path("booking-safety-examples.jsonl").read_text(
        encoding="utf-8"
    ).splitlines()
]
first_case = json.loads(rows[0]["source_case_json"])
print(first_case["case_id"], first_case["expected"]["mode"])

No API key, model, audio processing, network call, or booking integration is used.

Verification and limitations

The local packaging check verified twelve rows and lossless reconstruction of all original case objects. The source repository's 70 unit/mutation tests and twelve fixture checks validate structural and hand-authored expectation consistency only. They do not evaluate Georgian meaning, speaker identity, live calendar availability, concurrency, idempotency implementations, or OMO runtime.

The illustrative 120-second freshness limit is not an OMO setting. Dates in 2030 are fixture anchors, not claims about current events or actual appointments. Do not use these twelve examples as an unattended production decision policy or infer a statistical product-performance score from them.

Source, attribution and license

Source commit: 101fe8b983cde9c8edeeb8e5c2a6fda2bd77c228.

Based on the Georgian Voice AI Readiness Checklist by OMO AI, under its existing CC BY 4.0 notice. Packaging changes: a flat JSONL projection, a dataset card, and a lossless round-trip check. No original case, utterance, annotation or expectation was changed. No license terms were changed.

• —Original case data
• —Source validator, tests and verification scope
• —ქართული გზამკვლევი
• —English explanation
• —OMO knowledge center

ქართული შენიშვნა

ეს არის 12 გამოგონილი ქართულ-ინგლისური სასწავლო მაგალითი. მასალა არ შეიცავს რეალურ ზარებს ან მომხმარებლის მონაცემებს. ტექსტის ხელით მინიჭებული ანოტაციები არ არის ქართული ენის ამოცნობის ხარისხის მტკიცებულება. ლოკალური შემოწმება ადგენს ფაილების თანმიმდევრულობას და არა OMO-ს რეალური სისტემის მუშაობას.