This is not a mockup.
Everything below is quoted from a single run of the shipping engine on 8 August 2026 — the import response, the model's proposals, the gate's refusal, and the delivered document. The candidate is a synthetic test identity (an example.com address and a 555 number), because a demo should not need anyone's real resume. The run is real.
Import locks the facts first
A one-page resume goes in. Before anything is tailored, the importer proposes what must never change - and those lines are locked before any rewriting machinery can see them.
- Jane Rivera
- [email protected] | (555) 201-4488 | Portland, OR
- Cascade Analytics — Senior Data Engineer, 2021 - Present
- B.S. Computer Science — State University of Fairhaven
“No highlighting found. Locks below are policy defaults only. To choose your own, highlight the wording that must never change (yellow works well) and re-import.”
The warning is the engine's own, verbatim: it tells you it made policy defaults, not decisions on your behalf.
Then a job description
“We are hiring a Senior Data Engineer to build reliable analytics pipelines in Python for a regulated industry. You will own Airflow orchestration, Snowflake data modeling, pipeline reliability and alerting, and data quality across the platform.”
The AI proposed. The gate decided.
AI wording assist is opt-in per build and was switched ON for this run. The model proposed a rewrite for each bullet. The gate accepted one and refused the other - and the refusal came with its arithmetic.
Before
“Built Airflow pipelines in Python loading inventory data into Snowflake.”
After
“Developed Airflow pipelines in Python for loading regulated industry inventory data into Snowflake.”
Within the bullet's 'moderate' rewording bound, every word supported by the pack's evidence. It ships.
The model proposed
“Optimized pipeline reliability and alerting logic, reducing failure rate, for a regulated industry.”
Rewrite diverges too far from the source bullet for this role's 'moderate' mutation level (similarity 0.43, needs >= 0.70).
What ships instead
“Reduced pipeline failure rate by redesigning the retry and alerting logic.”
The original wording ships instead. A refusal never costs you the bullet.
The model ran locally (NVIDIA Nemotron on Ollama) at temperature 0, so this run is reproducible. The reason string is the engine's, unedited.
Then the document is checked, not trusted
These checks run against the rendered file - the thing you would actually send - not against what the engine intended to write.
- immutable_violationsnoneEvery locked line survived, character for character.
- unsupported_bulletsnoneNo claim without evidence behind it.
- geometry_violationsnoneLayout measured against the template, not eyeballed.
- deliverabletrueOnly true because all three above are empty.
The delivered document, in full
The complete visible text of the DOCX this run returned. Note the accepted rewrite in the first bullet, the refused rewrite absent from the second, and every locked line intact.
Jane Rivera [email protected] | (555) 201-4488 | Portland, OR PROFESSIONAL SUMMARY Data engineer with 6 years of experience building analytics pipelines for regulated industries. EXPERIENCE Cascade Analytics — Senior Data Engineer, 2021 - Present Developed Airflow pipelines in Python for loading regulated industry inventory data into Snowflake. Reduced pipeline failure rate by redesigning the retry and alerting logic. SKILLS Imported: Python, Airflow, Snowflake, Data modeling EDUCATION B.S. Computer Science — State University of Fairhaven
What this page is for
The hosted app is in private preview while the public deployment is finished. This page exists so the claim on the front of this site - it will not write anything you cannot prove - is something you can inspect rather than take on faith.