AUREN is a provider-neutral AI engineering system that turns software work into a bounded, auditable learning loop: understand the repository, plan, execute, verify, review, record evidence, evaluate what actually transfers, and promote or roll back learned strategies only when the gates pass.
AUREN (Autonomous Unified Reasoning & Engineering Network) is a provider-neutral AI engineering and orchestration system. It coordinates software work through explicit intent, repository context, bounded planning and execution, verification, review, durable evidence, and guarded learning. Before implementation, its orchestration records a review for common requirement conflicts, weak assumptions, and missing success measures, offering constructive options instead of silently choosing user-owned trade-offs. It is designed to improve engineering reliability—not to treat model output, stored memory, or learning signals as authority.
World model, self model, reasoning, and goals.
Planning, StateGraph execution, agent teams, and resource routing.
Episodic, semantic, engineering, and evidence-graph persistence with provenance.
Experimentation, causal attribution, capability invention, and curriculum evolution—without granting authority.
Independent evaluation, sealed holdouts, transfer testing, and attestation. External independence still requires external evidence.
Security, authority, verification, canary, promotion, and rollback controls.
Read the full About page. The runtime registry and receipt improve traceability; neither alone proves that every control is fully enforced in every deployment. See the current architecture and AI-era foundation principles.
Execution authority and learning are deliberately separated. Active engineering stays policy-gated and bounded. Learning becomes advisory only after independent evaluation, regression protection and promotion/rollback evidence.
Deterministic repository maps, symbol/import/call relationships, bounded graph expansion, token budgets and explicit unknowns keep context grounded.
Dependency-aware agent teams, serialized mutation, bounded retries and optional local resource offload prevent learning from silently becoming execution authority.
Experiences can be replayed and evaluated. Independent benchmarks, sealed runs, transfer, randomized attribution and fresh holdouts gate promotion.
AUREN is designed to catch blind spots, weak assumptions, missing success measures, and contradictory requirements before implementation starts. It explains why a concern matters, asks a focused question, and offers options with trade-offs rather than enforcing a one-size-fits-all answer.
Detect common conflicts such as preserving an API while requesting a breaking change. Unresolved explicit conflicts pause dependent implementation before side effects.
Turn claims like “fast”, “secure”, “robust”, or “scalable” into measurable questions and tests. Planner/reviewer reasoning supplements deterministic checks.
Recommend a path with trade-offs, offer a narrower or reversible alternative, and continue safe independent discovery while a user-owned decision is unresolved.
The review is recorded in the existing run evidence and trajectory. Heuristic checks cannot prove the absence of every semantic contradiction. Read the engineering design policy.
AUREN extends established engineering practices with explicit controls for autonomous execution, context, trust, evidence, recovery, and guarded learning. The principles are a shared contract across design review and runtime orchestration, not a replacement for existing security or promotion gates.
IPACT — Intent Before Action
INVAR — Invariant Preservation
DELTA — Defensible Engineering with Limited, Targeted Alterations
CONTEXT — Context Is an Engineering Dependency
VERA — Verification Before Acceptance
TRACE — Traceable Results And Claims Engineering
INDEP — Independent Evaluation Protocol
OWN — Ownership With Named Accountability
MEASURE — Metrics, Evidence, And Results Evaluated
PACT — Permission And Control of Tools
BLAST — Bounded Limits And Scoped Trust
TRUST — Treat Retrieved Untrusted Sources Thoughtfully
CIG — Context Integrity Guard
R3 — Recover, Resume, Roll Back
RAIL — Resource-bounded Autonomous Iteration Limits
HANDOFF — Handoff And Negotiated Definition Of Function
DEGRADE — Designed, Efficient, Adaptive Graceful Recovery Under Degraded Environments
LEARN — Learning Evaluated Against Regression and Novelty
PROMOTE — Proof-Required Operational Maturation Of Tested Enhancements
ECONOMY — Engineering Cost Optimization Measured Over the Full Lifecycle, Not Yield Alone
DETERM — Deterministic Decisions and Artifacts
MINCORE — Minimal Trusted Execution Core
OUTCOME — Explicit Operation Outcomes
MIGSAFE — Compatible and Reversible State Evolution
PROVEN — Verified Supply Chain and Artifacts
REVERS — Reversibility Before Abstraction
HUMANE — Honest and Actionable Operator Feedback
ACCESS — Accessible Operational Interfaces
LIFECYCLE — Named Ownership and Retirement
PRIORITY — Correctness and Authority Over Optimization
Source version: 1.0.6 (not yet released). Latest published release: 1.0.5.
Implementation: machine-readable registry · full principle definitions and control boundaries · root engineering contract.
The orchestrator records a digest-bound principle receipt before graph execution. Existing authority, security, verification, recovery, memory/context, and promotion gates remain authoritative. Receipt presence alone does not prove complete enforcement or independent external validation.
A simple engineering request can be given directly to AUREN. The important part is the requested evidence, not a magic prompt.
Fix the failing login test.
Inspect repository instructions and the existing authentication flow first.
Identify the root cause with evidence.
Make the smallest compatible change.
Add or update regression coverage.
Verify it, review the result, and report:
- what changed
- what was verified
- remaining risks
- evidence used for the decisionFor a feature:
Add retry handling to the outbound payment client.
Preserve current API behavior, inspect existing retry and timeout patterns,
implement the smallest safe change, add regression coverage, verify it,
and report open risks.python -m portable.repo_intelligence . \
--for="Fix the authentication timeout regression" \
--token-budget=4000python ~/.auren/current/auren_cli.py dashboard \
--project-root /path/to/repositoryOpen http://127.0.0.1:8765.
The repository includes a concrete independent-transfer test fixture. These numbers are calculated from the checked-in test data; they are not a claim that AUREN universally beats another agent.
| Measure | Checked-in example | What it demonstrates |
|---|---|---|
| Source-domain score | 0.70 | Baseline capability in the domain that produced the learning signal. |
| Target control | 0.575 | Independent holdout baseline across math and science. |
| Target treatment | 0.690 | Transfer performance after the candidate intervention. |
| Cross-domain transfer lift | +0.115 | Target treatment minus target control; above the fixture's +0.03 gate. |
| Randomized causal sample | 10 control / 10 treatment | Independent treatment/control observations used by the causal estimator. |
| Estimated treatment effect | +0.120 | 0.725 treatment mean minus 0.605 control mean. |
| Target domains | 2 | Transfer is measured beyond the source domain, not only in-domain. |
| Promotion condition | Pass all gates | Sealed receipt + randomized attribution + fresh holdout + transfer validity + evidence. |
Verification snapshot recorded on 2026-10-21 14:20 UTC at main SHA 3c873ddacc018c529ea799c9ce552e2ead47b011. At this snapshot, 17 check runs succeeded, 3 remained in progress, and 4 were skipped by change filtering; the result was not final. Skipped checks are not passes. Check live Actions and main history for newer evidence. This dated snapshot does not claim that its SHA is the current main head. Product version remains v1.0.5; this is not a release.
PR #764 merged an append-only task event journal, monotonic per-task sequencing, cross-instance/process writer locking, heartbeats, stale-task detection, terminal states, secret-like value redaction, and concurrent-writer regression coverage. Five exact-head workflows passed before merge: Runtime Resilience Matrix, Production AI Harness Hardening, AUREN Portable Bundle, Architecture Integrity, and AI Harness Check.
Implementation: PR #766.
The execution controller now records queue, dispatch, preflight, handler, heartbeat, quota-wait, recovery, and terminal events in the durable journal. The dashboard API and console render those snapshots, including stale/unknown state and next action. Terminal completion is shown only after the durable work-unit receipt commits. The read-only CLI command `python auren_cli.py progress --json`, merged in PR #767, exposes the same journal. PR #770 adds a bounded-wait provider executor, provider lifecycle callbacks, a cap on orphaned workers, and a timeout checkpoint that requires reconciliation before retry. PR #771 bounds in-process and cross-process journal lock waits. PR #773 applies the shared bounded-call helper to ProviderBackedCrewAdapter and the resumable executor, with a global eight-worker cap. It bounds caller wait but cannot kill a timed-out Python thread. The callback bridge merged in PR #775. PR #776 merged canonical-journal callback fan-out and explicit supervisor task-ID propagation. Hosts still must pass the controller-owned journal and supervisor-issued task ID; the alternate graph path must receive supervisor_task_id. PR #777 adds startup reconciliation: store disagreements become blocked, expired leases become recovering, and stale terminal projections are corrected by an append-only event. Expanded commit-boundary fault injection, cancellation/deadline/hard-termination testing, and ambiguous side-effect reconciliation remain open.
AUREN has bounded subprocess termination, deadline propagation into selected evidence checks, opt-in process-isolated crew dispatch, durable supervisor state, checkpoint conflict detection, and Windows Job Object descendant containment. Coverage is path-specific: the legacy thread dispatcher cannot kill a stuck Python thread, and these controls do not establish a sandbox or undo an external side effect.
Core, Engine, Memory, Learning, Arena, and Guard remain the canonical ownership model. Repository CI and receipts are not substitutes for independent evaluation, sealed holdout custody, external isolation review, or an operator-run migration/rollback rehearsal.
| Gate | Status | Evidence boundary / next action |
|---|---|---|
| PR #764 exact-head checks | Passed before merge | Five workflows passed on the reviewed PR head. See PR #764; do not treat this as a post-merge run on the current main SHA. |
| Post-merge runtime and Pages verification | Verified main snapshot | All ten post-merge workflows passed on be46356284981376d8aa7471d20cedd9f4a230e7, including Pages deployment, AI Harness Check, Runtime Resilience Matrix, and Architecture Integrity. These are post-merge results for that exact SHA; PR #777's exact-head results are recorded in live Actions. |
| Canonical progress journal | Implemented and regression-tested | Durable journal, monotonic sequencing, cross-instance/process locking and concurrent-writer tests are merged. |
| Supervisor-to-dashboard progress projection | Connected in this integration slice | ExecutionController lifecycle events feed the journal; the dashboard API and console expose snapshots. The adapter-to-journal callback bridge merged in PR #775; callback fan-out and task-ID propagation merged in PR #776. Host bootstrap must still pass the controller-owned journal and explicit supervisor task ID. Startup reconciliation is added in PR #777; expanded commit-boundary crash recovery remains open. The read-only CLI view is available. |
| Provider timeout and progress integration | Registration wiring merged; crash reconciliation in PR #777 | PR #773 applies the shared bounded-call helper with a global eight-worker cap. PR #775 adds allowlisted journal projection; PR #776 adds canonical-journal callback fan-out and supervisor-task-ID propagation. Host bootstrap must pass the controller-owned journal and task ID. PR #777 reconciles task/journal/receipt mismatches at startup but does not replay external effects. Timed-out calls may still finish late; reconcile before retry and use native transport deadlines/process isolation for hard termination. |
| Progress-journal lock waits | Verified on implementation snapshot | In-process and cross-process lock acquisition now has a finite deadline and contention tests. All ten post-merge workflows passed on main SHA 87831dc3e08029b82f51fc81bc8f79f569c9be6d, including Pages deployment; See PR #776's exact-head results in live Actions. |
| Required GitHub checks | Open governance gap | Latest inspected branch metadata reported protection enabled but enforcement off and no required status checks. An owner/admin must configure and verify effective rules. |
| Independent evaluator and sealed corpus | Pending | Requires independent custody, signed campaign evidence, and fresh holdout results. Checked-in fixtures are not external evaluation. |
| Runtime / sandbox isolation | Partial; external audit pending | Some authority bindings remain opt-in at host integration points. Repository CI cannot prove OS-level isolation. |
| Migration and rollback rehearsal | Pending | Run an operator-controlled staging rehearsal with backup, interruption, rollback, and retained-state evidence before closing issue #534. |
| Consequential external side effects | Integration-dependent | Use task-scoped idempotency keys, durable receipts, and ambiguity reconciliation at each effect boundary. |
The alignment review and gap register and unified engineering completion roadmap define ownership and exit criteria. Do not describe an open gate as complete without matching evidence.
The benchmark generator is a distinct principal from the runtime and oracle. Cases are opaque and content-addressed.
A run receipt must match the benchmark corpus and manifest digests. Holdout cases are checked against the sealed set.
At least two independent target domains are measured, while randomized control/treatment evidence estimates the treatment effect.
AUREN already has a portable distribution. The website itself is a static GitHub Pages site and does not require a server runtime.
curl -fsSL https://raw.githubusercontent.com/sourabhJainR/AUREN/main/install.sh | bashWindows PowerShell:
irm https://raw.githubusercontent.com/sourabhJainR/AUREN/main/install.ps1 | iexgit clone https://github.com/sourabhJainR/AUREN.git
cd AUREN
python auren_cli.py build --output auren-portable.zip
python auren_cli.py verify auren-portable.zip
python auren_cli.py install auren-portable.zip --skill autopython ~/.auren/current/auren_cli.py check-update
python ~/.auren/current/auren_cli.py update
python ~/.auren/current/auren_cli.py rollbackBefore deleting ~/.auren, uninstall any AUREN maintenance service using the matching user or system scope.
python ~/.auren/current/portable/maintenance_service.py --scope user uninstall
rm -rf ~/.aurenpython ~/.auren/current/auren_cli.py dashboard \
--project-root /path/to/your/repositoryGitHub Pages can publish a repository directly or through a GitHub Actions workflow. This site uses the Actions path so the static site/ directory is deployed on every push to main.
Repository Settings
→ Pages
→ Build and deployment
→ Source: GitHub Actions
The workflow:
.github/workflows/deploy-pages.yml
publishes:
site/index.htmlOnce enabled, the project site is expected at https://sourabhJainR.github.io/AUREN/.
GitHub may take a few minutes to publish the first deployment. See the official GitHub Pages documentation linked below.
AUREN is open source. Community support helps fund continued development, reliability testing, documentation, contributor onboarding, and work toward independent evaluation.
Improve the runtime, agent orchestration, and evidence-driven engineering workflows.
Expand regression coverage and test recovery, security boundaries, and rollback behavior.
Improve documentation, examples, onboarding, and the contributor experience.
Support is voluntary and does not guarantee specific features or outcomes. Independent evaluation and external validation remain ongoing goals; results will be reported only when evidence is available.