Build with AI speed.
Ship with human-grade evidence.

ASEWAVE (Augmented Solo Engineering With Adversarial Verification + Evidence) is a human-controlled AI software delivery methodology. It compresses solo development cycles for solo developers and small teams building production-grade software, by pairing one human gatekeeper with separate AI roles that do not trust each other by default — an executor that implements in its own session, and a reviewer that writes the prompts and independently re-verifies the results. Every phase must produce evidence before it closes.

ASEWAVE structures every unit of work into three phases — Investigate, Implement, Verify — with a human gate after each; evidence is captured inside Implement and Verify, not as a phase of its own. The human sets direction, owns every gate and commits the work. An Executor AI agent builds the work in its own session. Verification runs in a separate session that receives the results, not the process, re-runs the checks itself and challenges every claim. No phase closes until evidence — terminal output, file hashes, test results — exists on disk and survives independent review.

ASEWAVE is for solo developers, tech leads, and CTOs building production systems with AI assistance where reliability, auditability, or compliance matters. It fits domains like security tooling, regulated industries, and infrastructure — where AI-generated progress cannot be accepted without independent verification.

Vibe-coding accepts AI output if it looks correct. ASEWAVE requires that every AI-generated claim is verified against primary evidence before a phase closes. The human never delegates judgment — only execution. Adversarial separation means the AI that builds the work is never the same AI that approves it.

A phase gate in ASEWAVE is a mandatory checkpoint between work phases. It only opens when specific evidence exists: file existence confirmed, output hashes unique, tests passed, terminal logs captured, and a structured walkthrough completed. The gate prevents AI-fabricated progress from passing as real delivery.

Adversarial verification means the AI role that builds code is structurally separated from the AI role that reviews and challenges it. In ASEWAVE, the Executor Agent implements; the Verifier Agent independently checks claims against disk evidence, git history, and documentation. Neither AI approves its own work.

Yes. The ASEWAVE methodology is model-agnostic. The public method can be applied with any AI coding assistant. The private ASEWAVE Engine — which implements the full prompt contracts, verifier logic, and phase templates — is currently closed and not publicly released.

To get started with your own setup: copy the whitepaper at asewave.org/whitepaper.html and paste it into any AI assistant — ChatGPT, Claude, Gemini, Copilot, or similar — and ask it to design a concrete phase prompt structure for your specific tools. Tell it which AI you'll use as reviewer and which as executor. The method works with any pairing: a browser or desktop AI as the brain, any VSCode-based coding agent as the worker.

The ASEWAVE methodology is publicly documented and free to use with attribution. The operational engine — prompt contracts, verifier logic, memory structure, phase templates, and red-flag checks — is private and not currently released. See asewave.org for the full public method and whitepaper.

Phase A (Investigation) typically takes 10–30 minutes of AI time and 5–10 minutes of human review. Phase B (Implementation) scales with task complexity — a focused feature or fix may take 20–60 minutes. Phase C (Verification) usually takes 15–30 minutes of AI time plus a structured human walkthrough review. The human is not waiting for AI to finish — they are reviewing, deciding, and gating. The discipline cost is in the review time, not the elapsed clock. All times are typical and vary with scope.

The executor agent must never expand scope beyond the approved phase prompt, approve its own output, close a phase unilaterally, or delete or overwrite evidence files. It must also not proceed to Phase B before Phase A has been reviewed and approved by the human, and must not declare a task complete without running and capturing real test output. Any action that bypasses the human gate is a protocol violation — regardless of whether the result looks correct.

Language models in long sessions can subtly reinterpret their original instructions — optimising for what seems reasonable rather than what was specified. The Phase Boundary Re-Read Protocol counters this: the method requires each phase prompt to restate its constraints at the boundary, and every hand-over to be checked against the original prompt — not against accumulated conversation context, which may have drifted through clarifications, digressions, or extended tool use.