Architecture & Philosophy · Amplifier Bundle

ZeroVector

State your intent once. Get exactly what you meant.
Zero translation loss through fidelity convergence.

v0.3.0 · Fidelity Convergence
May 2026 · michaeljabbour/amplifier-bundle-zerovector
The Problem

Every handoff
loses signal

# The traditional translation chain — every arrow is lossy compression IntentSketchWireframeMockupHandoffDev InterpretationBuildReviewShip # By the time you ship, what % of the original intent survived?
🔊

Signal Decay

Each step from vision to product is a lossy compression. Requirements get misread, context gets dropped, nuance evaporates.

🔀

Interpretation Drift

Designer interprets PM. Engineer interprets designer. Reviewer interprets engineer. The original intent is nowhere in the chain.

No Feedback Loop

Linear pipelines have no mechanism to measure how much of the original intent survived into the final artifact.

A beautifully crafted artifact that solves the wrong problem has high quality but low fidelity. Quality without fidelity is wasted effort.

— Zero-Vector Design, Fidelity Framework
The Solution

Intent → Crew →
Artifact

Collapse the entire translation chain into three steps. The person with the vision directs a crew of AI agents and moves directly from intent to working artifact.

# OLD: 8 lossy handoffs Intent → Sketch → Wireframe → Mockup → Handoff → Dev → Build → Review → Ship # NEW: Zero translation loss Intent → Crew → Artifact # The artifact IS the thing. Not a picture of the thing.

Fidelity Over Quality

Quality asks: is this well-made? Fidelity asks: does this faithfully represent what was intended? ZeroVector optimizes for fidelity first, because quality without fidelity is wasted effort.

Crew, Not Assistants

Agents are crew members with roles, standards, and contracts — not assistants waiting for literal instructions. Each executes with judgment within their domain.

The Crew

Five agents, one pipeline

Each agent is a quality gate for the next. Together they eliminate translation loss at every stage.

🔍

Intent Analyst

Decode intent, surface constraints, define what success looks like

🏗️

Architect

Translate intent into structure, interfaces, and ordered tasks

🔨

Builder

Implement the artifact from spec, meeting every acceptance criterion

🧐

Critic

Score all five fidelity lenses, identify the priority gap, route action

🚀

Shipper

Package, commit, document, and deliver the finished artifact

An assistant waits for instructions and executes them literally. A crew member has a role, a standard, and a contract — and executes with judgment.

Fidelity Scoring

Five lenses, scored 0.0 – 1.0

Every lens measures translation loss at a specific stage. The Critic scores all five simultaneously. The weakest lens gets routed to first.

1. Intent Clarity
Is the original intent fully understood and unambiguous?
If weak → Return to intent capture with user
2. Specification
Does the spec completely and correctly capture the intent?
If weak → Revise spec, fill gaps, trace requirements
3. Implementation
Does the artifact faithfully implement the spec?
If weak → Continue building, fix deviations from spec
4. Quality
Does the artifact meet quality standards — tests, style, correctness?
If weak → Add tests, fix bugs, improve readability
5. Ship-Readiness
Is the artifact packaged and deliverable?
If weak → Complete docs, clean commits, resolve blockers
Priority rule: When lenses tie, prioritize earlier in the pipeline. Upstream fixes compound downstream.
Core Architecture

The convergence loop

Not a linear pipeline. A convergence engine that routes to the weakest lens and acts until overall fidelity meets the target threshold.

fidelity-convergence.yaml
DECODE-INTENT(GATE 1: approve spec)ASSESS all five lenses simultaneously (score 0.0–1.0)while fidelity_score < target_fidelity: route to weakest lens → act → re-assess → (GATE 2: approve converged artifact)FINISH (merge / pr / keep / discard)

Default Target: 0.85

The convergence loop terminates when the arithmetic mean of all five lens scores meets the domain-specific target.

Max 8 Iterations

Safety bound prevents infinite loops. Persistent failures get Builder the Critic's exact findings for targeted fixes.

2 Human Gates

Approval after spec and after convergence. The human stays in control — the crew does the work.

Domain Tuning

Six crews, one architecture

Same five agents, domain-tuned prompts per crew. Each mode calibrates fidelity targets and lens priorities for its domain.

/crew

General — Any domain. Auto-adapts. The Intent Analyst surfaces the correct domain from your intent.

Target: 0.85

/crew-build

Code — Features, apps, tools, scripts. TDD tasks, test-first build, spec compliance.

Target: 0.85

/crew-product

UX — Flows, specs, validation, strategy. Jobs-to-be-done, user-centricity, decision points.

Target: 0.80

/crew-platform

Infra — Modules, APIs, architecture. Contracts, migration paths, breaking change analysis.

Target: 0.88

/crew-research

Analysis — Investigation, synthesis, papers. Evidence standards, source strategy, actionability.

Target: 0.80

/crew-content

Writing — Docs, curriculum, posts. Audience-fit, narrative arc, clarity.

Target: 0.75
Composability

Universal
Fidelity Layer

The fidelity behavior is an extractable, standalone layer. Any Amplifier bundle gets five-lens scoring, a live ANSI dashboard, and the fidelity state tool — without the full ZeroVector crew.

# Add fidelity diagnostics to ANY bundle — 2 lines includes: - bundle: zerovector:behaviors/fidelity # You get: # • zerovector:critic agent — structured 5-lens scoring # • tool-fidelity-state — Python module for score I/O # • hooks-fidelity-reporter — live ANSI dashboard per turn # • fidelity-framework context — scoring rubric + evidence reqs

3 Python Modules

  • tool-fidelity-state — Read/write fidelity scores to session state (198 lines)
  • hooks-fidelity-reporter — Live ANSI dashboard after each agent turn (388 lines)
  • hooks-crew-gate — Blocks artifact creation without an active crew mode (186 lines)

Evidence Before Claims

A score without evidence is a guess. Every lens score must be accompanied by specific evidence — test results, file paths, cited requirements, concrete metrics. Vague justifications like "looks good" are not evidence.

Philosophy

Seven principles of Zero-Vector Design

Founded by Erika Flowers. This is NOT vibe coding. It is rigorous, intentional, crew-directed work.

1. Work in the Medium

Build real artifacts, not representations. A working module, not a mockup. The medium is the message.

2. Boundaryless by Nature

Disciplinary walls are constraints of the old model. The person with the vision operates across all of them.

3. The Medium Shapes Thought

Working directly in the real artifact changes what you see, what you ask, and what you build. Representations lie.

4. Intentional Impermanence

Don't patch — regenerate. The spec is the truth; the artifact is an instance of it.

5. POSIWID

Judge systems by outputs, not intentions. If the process produces translation loss, change the system.

6. Compound Your Leverage

A precise intent produces a sharper spec. A sharp spec produces a faithful build. Early precision compounds.

7. Venture Beyond “Possible”

The old ceiling was set by the translation chain. When translation is eliminated, the ceiling rises.

By the Numbers

Built in 4 days

63
commits
5
agents
6
domain modes
6
recipes
25
test files
5k
lines of tests
772
lines Python modules
130
total files
March 6–9, 2026 · Sole contributor: Michael J. Jabbour · 51 commits on March 8 alone
Sources & Methodology

How this deck was built

Primary source: michaeljabbour/amplifier-bundle-zerovector on GitHub

Files examined:

Line counts (non-git, non-investigation):

Philosophy source: zerovector.design by Erika Flowers

No metrics were fabricated. All numbers come from git log, git shortlog, wc -l, and find against the cloned repository.

Get Started

State your intent

Install the bundle. Activate a crew. Say what you want. The crew converges to a high-fidelity artifact.

# Install amplifier bundle add git+https://github.com/michaeljabbour/amplifier-bundle-zerovector@main --app # Activate a crew and state intent /crew-build > Build a CLI tool that validates bundle.md frontmatter
github.com/michaeljabbour/amplifier-bundle-zerovector
v0.3.0 · 5 agents · 6 crews · 5-lens fidelity convergence

Built on Zero-Vector Design by Erika Flowers
More Amplifier Stories