Technical review

Inspect the code, reproduce a failure and its repair, and assess what the evidence supports.

Delta Atlas is part of Jake Tiller's independent, AI-assisted research and engineering work: browser tools, explicit decision rules, and reproducible checks around uncertain AI work.

Try Gap Check Inspect the repository

Three things to inspect

1. Shared implementation
Gap Check and Continuity Audit use the same lexicon engine. Framework Audit and Continuity Audit share a scoring engine. The browser and Node checks reuse those implementations. The tools require no model call at runtime.
2. A failure with a concrete repair
The earlier bridge detector counted “Human review was never implemented” as oversight. The corrected detector applies cancellation checks to ordinary-language bridges as well as aliases. The corpus retains positive and negative examples. Finite phrase checks still have limits; the tool does not establish real authorization.
3. Research beyond the interface
Clutch specifies approved task handoffs and bounded recovery; its published preview remains PREPARE_ONLY and establishes no general savings claim. The STP candidate packet separates evidence, effects and accepted state. The Stable's findings preserve experiments and failed hypotheses.

Reproduce a small check

From a repository checkout with Node.js installed, run:

node gapcheck-corpus.js
node coherence-audit-corpus.js
node continuity-audit-corpus.js

These checks need no package installation or API key. They test specified behavior on labeled cases. For broader validation, inspect the repository workflow and its runtime requirements. Record the commit you checked; a passing historical build does not verify later edits.

The theory and its current boundary

The research question is practical: what must be checked before an AI-assisted proposal changes a record or action we rely on? The work separates proposals, permission, execution, observations and acceptance, then retains failures for bounded recovery. The project guide explains the Atlas methods and their limits. From Model Output to Accepted State, v4 is a separate owner-review paper candidate.

My contribution is the problem framing, functional synthesis, task and evidence design, and direction of the implementation and corrections. AI assistance is part of the drafting and development process. Evaluate that work through the source, tests, decisions and preserved failures.

Dataset status

The candidate source inventory has 439 records; the current Ask and Explore snapshot has 435. All source records carry the repository label reviewed, but complete reviewer, date, source-locator and digest receipts are not consistently present. That label is not proof of semantic correctness. The data-sync baseline records the older projections and their differences.

Atlas checks text against explicit rules. It does not enforce policy or certify safety. Source-support indicators describe recorded sourcing; they do not prove a definition true. General productivity, security and economic benefits remain questions for appropriately scoped evaluation.

Useful feedback

A reproducible counterexample is especially useful: include the page or commit, input, expected behavior and observed result. Open an issue or browse the other projects.

Updated 5 September 2026. Independent educational research, provided as-is. Not legal, compliance or professional advice; no endorsement by the organizations cited. Content license: CC BY 4.0.