Problem
Most autonomous coding systems blur reasoning, execution and proof. VAJRA explores a different model: an engineering run must survive worker, model, process and machine failure without allowing the reasoning layer to become the authority.
Constraints
- Local-first and model-agnostic.
- ₹0.00 operating-cost target.
- No paid inference or infrastructure dependency.
- Canonical state must survive disposable workers and attempts.
- Humans remain the final authority.
Architecture
The central boundary is:
Objective → Context → Reasoning → Intent → Policy → Execution Broker → Sandbox / Workspace → Artifact → Independent Verification → Evidence
A worker or model proposes work. Policy decides what is permitted. The broker is the execution authority. The sandbox and workspace contain effects. Independent verification decides whether evidence supports progress.
The durable distinction is fundamental:
Run = durable · Step = durable · Attempt = disposable · Worker = disposable · Model = disposable
Engineering Evolution
The repository has progressed through durable execution and recovery, safety/control boundaries, context and workspace isolation, independent verification and anti-gaming, bounded objective-to-evidence autonomy, chaos and long-run durability, capability-aware routing, an always-on control plane, provenance-bound engineering memory, production hardening, advanced autonomy, and controlled self-improvement.
The current repository reports Phase 17 — Controlled Self-Improvement: complete / local gate passed.
Self-Improvement Boundary
Self-improvement is deliberately not unrestricted self-modification. A bounded proposal moves through:
proposal → isolated branch → tests → independent verification → security verification → human approval → explicit promotion
Authority and security surfaces remain outside the improvement scope. The model cannot authorize, promote, modify protected authority or restart VAJRA itself.
Failure as a Design Input
VAJRA treats failures, leases, retries, recovery, provenance and verification evidence as first-class state rather than incidental logs. Recovery is therefore a controlled engineering operation instead of simply asking a model to try again.
Verification
The project has explicit phase gates and a substantial validation trail. Independent verification is structurally separated from worker/model output so that producing an artifact and proving it correct are different responsibilities.
Why It Matters
VAJRA is less interesting as an “AI agent” than as an experiment in trustworthy autonomy: how far autonomous engineering can go while deterministic policy, evidence and human authority remain above model reasoning.
Current State
The repository is the authoritative source for current phase status, implementation details and gate evidence. The frozen v0.1.0 baseline remains preserved while later phases continue on main.