SENAR

Supervised Engineering & Normative AI Regulation

The methodology for producing verified software when AI writes the code and humans supervise.

v1.3 · 25.03.2026
Core: 8 rules · 2 gates · 2 metrics | Full: 15 rules · 5 gates · 10 metrics
CC BY-SA 4.0
EN + RU

The Problem

AI coding tools are everywhere. Governance for them is nowhere.

What happens without a standard

  • Plausible wrong code. You give AI a vague prompt. It produces code that compiles, runs, and does the wrong thing. You don't notice because it looks right.
  • Self-validating bugs. AI writes code AND the tests. The tests pass because they test the AI's misunderstanding, not your requirement.
  • Knowledge evaporates. You spend 40 minutes discovering that approach X doesn't work. Next session, the AI tries approach X again. Nobody wrote it down.
  • No way to measure. Is AI making you faster? Or just making bugs faster? You have no data either way.

What SENAR changes

  • Requirements before code. AI receives structured context: goal, acceptance criteria, scope boundaries. Not vibes.
  • Human verifies against criteria. Not "looks right" but each acceptance criterion checked independently against the requirement, not the AI's output.
  • Dead ends documented. One sentence: "Approach X fails because Y." Prevents hours of repeated failure across sessions.
  • Ten metrics tell the truth. First-Pass Success Rate, Defect Escape Rate, Throughput, Lead Time, Adversarial Detection Rate, and five more. You know exactly what AI is doing for you.
"AI does not ask clarifying questions — it picks an interpretation silently."

The resulting code compiles, passes the tests the AI wrote (which test the wrong behavior), and looks correct on review. This is the self-consistent artifact problem: AI generates code, tests, and documentation that are perfectly consistent with each other — and all wrong.

Every methodology before SENAR — Scrum, SAFe, Kanban — was designed for teams of humans who can ask each other "did you mean X or Y?" AI doesn't ask.

SENAR is the missing layer between "I use AI" and "I produce verified software with AI."

What Makes SENAR Different

This is not "write good prompts in ISO language." These are structural problems that require structural solutions.

Problem #1

Self-Consistent Artifact Problem

AI generates code, tests, and docs that agree with each other — but not with your requirement. Standard QA catches nothing because the tests validate the code against the AI's interpretation, not yours. SENAR breaks this loop: the Supervisor verifies against acceptance criteria written before the AI saw the task.

Principle #1

Quality at Input

A defect in a requirement cascades to code, tests, and documentation. AI amplifies this because it doesn't push back, doesn't ask "did you mean X or Y?" and doesn't flag ambiguity. SENAR's cascade principle: fix inputs, or every output is contaminated. The 3-level requirement hierarchy (BR → SR → TR) is not bureaucracy — it's defect prevention.

Principle #2

Documentation as Active Context

Code docs are not post-hoc bureaucracy. They are working context that the AI reads on every task. Good docs reduce per-task overhead because the AI starts with correct assumptions instead of guessing. Dead ends, architecture decisions, scope boundaries — they exist for the AI's benefit, not a filing cabinet.

Governance

AI Model as Supplier

When your model version changes, your production unit changes. This is equivalent to a compiler version change: baselines must be recalibrated, quality gates re-validated, and efficiency metrics re-baselined. SENAR is the first methodology to treat AI model governance as a supply chain management problem, not just "pick the best model."

Quick Start: 6 Habits, 5 Minutes

No meetings. No certifications. No tools to buy. Start today.

Habit 1 · Before

Write the WHAT

One sentence: what is done when this task is done? If you can't write it, you don't know what you're building.

Action: Write a goal sentence before you open the AI tool.

Habit 2 · Before

Write the DONE

Numbered acceptance criteria. Testable. "User can X" not "system should be good." Set scope boundaries — tell the AI what NOT to touch.

Action: Write 2-5 acceptance criteria before code starts.

Habit 3 · During

Verify each criterion

Check each AC independently. Not "looks right" but "criterion #3 is satisfied because I tested X." The AI's tests don't count — they test the AI's interpretation.

Action: Go through AC one by one. Check, don't glance.

Habit 4 · During

Record dead ends

When an approach fails, write one sentence: "X fails because Y." Costs 10 seconds. Saves 40 minutes next session.

Action: Keep a dead-ends file. One line per failure.

Habit 5 · After

Tests pass + AC met = done

Not "I think it works." CI green, types clean, every acceptance criterion checked. The task is done or it isn't.

Action: Don't close a task until CI passes and AC are checked off.

Habit 6 · After

Capture knowledge

Write down anything non-obvious you learned. Architecture decisions, gotchas, workarounds. The AI reads these next time.

Action: Spend 60 seconds writing what you learned.

Full Quick Start Guide

Start Here

SENAR Core

8 rules. 2 gates. 2 metrics. Adopt in 1 hour.

The minimal, self-contained subset of SENAR. No tooling required. No meetings. No roles to assign. Just the essential discipline for AI-assisted development.

Choose Your Level

Start with Core. Scale when you need to.

Feature Core Foundation Team Enterprise
Rules 8 11 15 15
Quality Gates 2 (Start/Done) 3 (QG-0..QG-2) 5 (QG-0..QG-4) 5 + custom
Metrics 2 (FPSR, Dead End Rate) 4 10 10+
Roles 1 3 (combined) 5 (dedicated) 5 + portfolio
Ceremonies 0 3 7 7 + portfolio
Tooling required None Recommended Required Required

Before and After

Real data from 552 tasks across 38 sessions. Single-team project (5 services, 2 frontends). Honest caveats included.

Without SENAR

  • Tasks without acceptance criteria → AI produces plausible wrong code that passes review
  • No dead ends → same failed approaches retried session after session
  • Marathon sessions (6+ hours) → 2.7x efficiency drop vs. focused sessions
  • No metrics → no way to tell if AI is helping or hurting
  • Cost: ~$105 per escaped defect in rework time

With SENAR

  • Every task has goal + acceptance criteria → 85%+ First-Pass Success Rate
  • Dead ends documented → 100% knowledge reuse across sessions
  • Sessions capped at natural breakpoints → consistent, predictable output
  • Ten metrics tracked → you know exactly where AI helps and where it doesn't
  • Cost: ~5 min per session + 1-3 min per task overhead

Honest caveats

  • Data is from one team, one stack, one AI tool (Claude Code). Your results will vary.
  • The 85% FPSR was achieved at maturity. Early sessions were lower (~60-70%) as practices were being established.
  • SENAR doesn't make bad requirements good. It makes bad requirements visible before they become bad code.
  • Overhead numbers assume you're already writing some form of task description. If you currently use no process at all, initial overhead will feel higher.

Who Is It For

Same methodology. Four levels of ceremony.

Core

1 developer

Solo developer using Claude Code, Cursor, Copilot, or any AI coding tool. Start with the Quick Start guide. No meetings required.

  • 2 quality gates
  • 2 metrics (FPSR, Dead End Rate)
  • 8 rules
  • ~5 min session overhead
  • ~1-3 min per task overhead
Foundation

1-3 developers

Combined roles, session management, monthly quality sweep. The bridge between solo practice and team coordination.

  • 11 rules
  • 4 metrics
  • 3 ceremonies
  • QG-0 + QG-2
  • 3 combined roles
Recommended
Team

3-10 developers

Full requirement hierarchy (BR → SR → TR). Dedicated Supervisor and Reviewer roles. Requirement library. Federation across projects.

  • 5 quality gates
  • 10 metrics
  • All 7 ceremonies
  • Cross-project coordination
  • Knowledge base with reuse tracking
Enterprise

10+ developers

Requirements-as-code with CI validation. Audit trails for compliance. Portfolio-level AI cost tracking and budget governance.

  • Full governance model
  • AI model supplier management
  • ISO 9001 compatible
  • SAFe integration patterns
  • Efficiency model with cost formulas
Reference Implementation

TAUSIK

Task Agent Unified Supervision, Inspection & Knowledge

Open-source framework that enforces SENAR rules automatically. Quality gates that physically block agents from skipping steps — not recommendations, but enforcement. Works with Claude Code, Cursor, and Windsurf.

  • 15 automated checks — pytest, ruff, tsc, eslint, cargo, go vet
  • 33 structured skills — /plan, /ship, /review, /audit, /debug
  • 80 MCP tools — programmatic access to project memory
  • 6 automatic metrics — FPSR, DER, throughput, lead time
  • Zero core dependencies — Python 3.11+ standard library only
# Install in 3 commands
$ git submodule add \
github.com/Kibertum/tausik-core \
.tausik-lib
$ python .tausik-lib/bootstrap/bootstrap.py \
--smart --init my-project
$ # restart IDE — done

AI writes the code.
You set the standard.

Start with Core. Scale to full governance. Open standard, free forever.

SENAR v1.3 · 25.03.2026 · CC BY-SA 4.0 · Normative language per RFC 2119