Local checklists, experimental text checks, and inspectable evidence. Delta Atlas helps you review a work handover and explore fixed checks for AI plans and action logs. No model call or account is needed to use the tools.
Created by Jake Tiller through AI-assisted development. The project combines working JavaScript tools, labeled tests, documented corrections, and proposed research. Its value can be assessed from those artifacts.
Open the tools · Two-minute guide · Guided technical review
Look for handover gaps: answer five short questions about a task, its usual owner, an authorized backup, written instructions and a trial. Missing answers are reported as not provided; arrangements remain not verified, even when you supply an answer. Try the built-in example without preparing a document. The older text audit remains available as an advanced option.
Review the wording of an AI plan (Gap Check) looks for recognized risk and safeguard phrases. Review an AI action log (Priority Tracer) highlights some wording about bypassing or stopping for a rule. Both include examples. Everyday wording, context and contradictions can be missed; supplied text does not establish what actually happened.
Inspect the shared detector, its labeled cases, and the Continuity Audit cases that exercise the same engine in another domain. The coherence scorer is also shared across tools rather than copied into each page.
The detectors use finite word and structural rules. Their output points to text worth reviewing; a clean result does not establish meaning, correctness, or safety. Text-pattern scores in the advanced audit are diagnostics, not measurements of operational resilience. Automated checks and example walkthroughs do not establish beginner comprehension, accuracy across arbitrary inputs, or productivity gains. Independent user testing remains open.
| Question | Artifact |
|---|---|
| Does the detector retain its expected behavior? | Tracer corpus, Gap Check corpus, and calibration history. |
| Can the same detector run in a pipeline? | CLI, with JSON output and explicit exit codes. |
| How are decisions and corrections retained? | Governance ledger, append-only events, and Node/Python projection checks. |
| How are external contract changes noticed? | MCP manifest pin, an offline hash-based comparison tool. |
| What is actually checked in CI? | Repository workflow, with pinned runtime versions and named checks. |
The basic detector checks require Node.js and no package installation.
git clone https://github.com/JakeTOpenSource/Resilience-Ledger.git
cd Resilience-Ledger
node tracer-corpus.js
node gapcheck-corpus.js
node continuity-audit-corpus.js
Run the CLI against a local file:
node delta-atlas-cli.js trace mytrace.txt --json
node delta-atlas-cli.js gapcheck myplan.txt --json
Exit codes are 0 for no flags, 1 for flags, and 2 for a usage or load error. These are detector outcomes, not safety certificates.
For the browser tools, download the repository and open index.html or a tool page. README-Portability.md explains file-access restrictions, optional assets, offline caching, and which views need more. The governance documentation and STP reproduction instructions cover their additional checks and prerequisites.
The declared candidate input, terms.enriched.json, contains 439 terms, all carrying a stored reviewed label. That label is not a receipt proving who reviewed a definition, when, against which exact source, or whether it is semantically correct.
Ask, Explore, and the Curation dashboard retain 435-term projections: 177 reviewed and 258 candidate. Gap Check retains 433 terms. These are different snapshots, not interchangeable counts of verified knowledge.
The data-sync baseline records missing IDs, status mismatches, source differences, and the smaller Canon projection. Its current decision is DEFER. A passing baseline check preserves that recorded disagreement; it does not reconcile or accept the content.
| Project | What to review |
|---|---|
| Clutch | A separate human-approved task-routing protocol, canonical skill, deterministic reducers, adversarial fixtures, and release package. Version 0.4.0-rc.1 is a research preview in PREPARE_ONLY; publication does not activate routing. |
| The Stable | Agent-calibration instruments, recorded plans, replay tools, and retained failed hypotheses. Start with its findings. |
| The Crosswalk | Static decision aids for deciding whether to code, learn, direct AI, or hire. |
| typed-refusal-harness | Aggregate experiment counts and an attribution audit. The original model harness is not published there. |
GroundingHarness and AI-Governance-Ledger are historical project names; this repository is the current home.
The STP v1.2 packet is proposed research containing public sources, schemas, synthetic fixtures, and JavaScript/Python reducers. Agreement on its enumerated cases does not establish live calibration or general reliability.
The paper From Model Output to Accepted State remains in owner review: PR #15 contains the v4 candidate, and PR #14 retains v3. Neither is merged. Passing a release verifier checks the stated artifacts; it is not peer review or acceptance of every claim.
The power-witness studies are an experimental side project. Power telemetry can characterize activity under a defined measurement setup; it does not establish useful work, authorship, or truth.
For the method, read White-Paper.html, Coherence-Ledger-Method.md, and the recorded adversarial review. Sources and borrowed ideas remain attributed in the relevant documents.
Six-Signal-Method.html explains six evidence categories and what each can establish. The pinned, read-only snapshot packet in research/atlas-snapshot-read-only/ preserves a dated observation; it is not a live system assessment.
The analysis tools run locally in the browser and do not send submitted tool text to a model or API. The hosted site still makes ordinary page requests, and Cloudflare may collect visit and page-performance telemetry.
Local overlay filenames are ignored by default. This reduces accidental inclusion of untracked files; it does not prevent forced adds, previously tracked files, or renamed copies. Review the staged diff before publishing. See README-Portability.md for the operating boundaries.
Repository checks do not establish which exact files a live deployment is currently serving. Deployment receipts are dated observations with their own evidence boundaries.
The content, data, tools, and documentation in this repository are CC BY 4.0, as stated in LICENSE.txt. Credit Jake Tiller and Delta Atlas / Resilience Ledger, link to the license, and indicate changes. CITATION.cff supplies citation metadata.
This is independent educational research, with no affiliation with or endorsement by the organizations it cites. AI assistance is part of the development history; the code, sources, tests, limitations, and recorded corrections are the basis for review.