Documentation

Voyager vs Fathom

The product boundary between empirical scenario evaluation and predictive fleet analysis.

🧭 Voyager vs Fathom

One of the easiest ways to confuse Skill Harbor's evaluation story is to blur Voyager and Fathom together.

They are related, but they do different jobs.

Short version

  • Voyager measures what happens in a scenario.
  • Fathom analyzes what is likely to help or hurt before or across those scenarios.

That distinction should stay intact even when Voyager adds deterministic benchmark packs.

Voyager: empirical scenario evaluation

Voyager is the place for questions like:

  • Did the skill-enabled path outperform the no-skills path?
  • Did the agent invoke the expected tools?
  • Did the scenario pass or fail?
  • What trace did the run produce?
  • Can this evaluation be reproduced in CI?

So Voyager owns:

  • scenario execution
  • branch comparison
  • traces
  • assertions
  • pass/fail outcomes
  • benchmark-pack results

Fathom: predictive and audit analysis

Fathom is the place for questions like:

  • Is this fleet too large or too noisy?
  • Are skills colliding semantically?
  • Is the routing surface too vague?
  • How much context bloat is this adding?
  • Which skills look risky or promising before deeper evaluation?

So Fathom owns:

  • heuristics
  • token/context analysis
  • contract validation
  • routing-risk signals
  • fleet-wide reports
  • predictive usefulness guidance

Why benchmark packs still belong in Voyager

It may feel surprising that offline deterministic packs are still a Voyager feature.

But the important boundary is not:

  • online = Voyager
  • offline = Fathom

The real boundary is:

  • scenario outcomes = Voyager
  • predictive audit = Fathom

A deterministic benchmark pack still asks a Voyager question:

what happened in this scenario, and how did the with-skills branch compare to the without-skills branch?

That is still empirical evaluation, even when the evaluation is fixture-driven and reproducible in CI.

How they should work together

The intended flow is:

  1. Govern the fleet with Skill Harbor
  2. Inspect and predict with Fathom
  3. Measure scenario outcomes with Voyager
  4. Feed what you learn back into fleet refinement

In other words:

  • Fathom helps you decide what looks worth evaluating or trimming.
  • Voyager tells you what the scenario evidence actually says.

A practical rule

If a command produces:

  • scenario traces
  • branch outcomes
  • assertion results
  • uplift/regression evidence

it belongs in Voyager.

If a command produces:

  • heuristics
  • token/collision reports
  • probabilistic routing confidence
  • fleet-level recommendations

it belongs in Fathom.

Non-goal of Voyager benchmark packs

Voyager benchmark packs should not become:

  • fleet health scores
  • heuristic usefulness scores
  • predictive deployment recommendations

Those belong in Fathom.