Zero Translation Loss

measuring how much intent survives to ship

The Question Quality Never Asks

Your pipeline is lossy compression from intent to ship

ZeroVector is an Amplifier bundle that treats the road from intent to artifact as a chain of handoffs — and reframes the whole problem as translation loss, not craftsmanship.

“Fidelity asks: does this faithfully represent what was intended?” Quality asks whether it’s well-made. These are not the same question.

So the first job is to define fidelity precisely enough to measure it.

Fidelity has a formula and a test suite behind it

The overall score is the arithmetic mean of five lens scores — sum(lens_scores) / len(lens_scores) — and the bundle ships instrumented, not aspirational.

If it’s this concrete, what exactly was broken before?

626
test-function definitions
17
test files
148
tests in the largest file
5
fidelity lenses averaged

Nine handoffs, and every arrow silently lost intent

Each step reinterprets the last — designer reads PM, engineer reads designer. The linear pipeline had no way to measure how much of the original intent survived.

The problem isn’t any one stage — it’s the chain itself.

Intent Sketch Wireframe Mockup Handoff Dev Interpretation Build Review Ship

“Every handoff loses information. The final artifact is a degraded version of the original intent.”

Nine stages collapse into one loop: Intent → Crew → Artifact

ZeroVector replaces the linear discover→specify→build→review→ship pipeline with a single convergence engine driven by a crew of five agents.

Collapsing the chain is what makes measured fidelity possible — here’s how the loop runs.

Intent Crew Artifact
5
crew agents
6
recipes
6
modes

It always routes to the single weakest lens, not everything at once

Five lenses score the artifact; the loop finds the lowest-scoring lens and sends it to that lens’s dedicated agent to close the gap.

Targeting the weakest lens is only half the loop — when does it stop?

Intent Clarityintent-analyst
Specificationarchitect
Implementationbuilder
Qualitycritic
Ship-Readinessshipper

The loop runs until fidelity clears a per-domain target, capped at 8 passes

A while-loop continues while fidelity_score < target_fidelity and breaks when it clears the target — a stopping condition calibrated per domain so “done” is provable, not declared.

A measured stopping point is what turns fidelity into a dial.

Content0.75
Product / Research0.80
Default / Build / General0.85
Platform0.88
Max iterations8

Fidelity becomes a measured dial, not a hope

Each lens score is clamped to [0,1]; the artifact reports CONVERGED only when overall fidelity reaches the target — otherwise NEEDS_WORK. “Is the intent intact?” now has a number.

Optimizing translation loss — not just quality — is the whole point.

0.80 – 1.00Strong
0.60 – 0.79Moderate
0.30 – 0.59Significant gap
0.00 – 0.29Critical gap
What To Keep

Measure the loss, not just the craft

If every handoff degrades intent, the win isn’t polishing each stage — it’s making the degradation visible and driving it toward zero.

Fidelity is translation loss between intent and artifact. Quality asks “is this well-made?”; fidelity asks “does this faithfully represent what was intended?”
ZeroVector v0.3.0 · Active
Sources

Sources & Research Methodology

Data as of: HEAD commit 2812b2c, dated 2026-07-01 (ZeroVector v0.3.0).

Repository: michaeljabbour/amplifier-bundle-zerovector (cloned to /tmp/zerovector-src). Thin Amplifier bundle depending on github.com/microsoft/amplifier-foundation@main; credits Zero-Vector Design by Erika Flowers.

Feature status: Active — v0.3.0 fidelity convergence architecture rewrite.

Commands run:

Gaps: README self-describes “two Python modules” while three module dirs exist on disk (tool-fidelity-state 198, hooks-fidelity-reporter 388, hooks-crew-gate 186 LOC) — an internal repo doc inconsistency, reported not resolved. Not every mode/recipe file’s full contents were read.

Primary contributor: Michael J. Jabbour — 64 of 64 commits (100%).

More Amplifier Stories