🪦 Harness Path — The Universe Builder Retired

10 · universe-builder — where the harness was applied and should not have been — an entry in the origin fleet catalogued in The Harness Paths (group: The negative result). The catalogue is the operator's work WITH the harness, not part of the system you adopt; none of its code lives in the harness repo. This is the map entry, so you can see what a filled config looks like at this weight.

(in this repo, at universe/retired/. Kept on purpose.)

The most useful entry in this catalogue, because it is the only one that
failed. Its target was a map of a corpus, and it was given the full
apparatus: two blind lenses, a Fiat-Shamir Gap, a held-out gate, an advancing
frontier. Two rounds, roughly 1.4 million subagent tokens, and not one fact
learned about the map.

• u1 died of infrastructure — and paid for itself by exposing that the
  engine counted dead agents as dry rounds, reporting an outage as an
  exhausted search. Fixed, and pinned by engine/loop.test.mjs.
• u2 ran perfectly: 9/9 agents, seeds re-derived, verdicts on disk. Both
  candidates beat the metric, and both scored 0/8 — because the Gap was
  drawing arbitrary facts (a licence string, a JSON version field) from
  612 KB of sources that the 7 KB map was never responsible for carrying. The
  objective demanded compression; the gate demanded a transcript. The
  feasible set was empty. The verdicts measured the config, not the artifact.

Then a nine-line auditor found, on its first run, in milliseconds, for zero
tokens, the very defect u2's by-container lens had found and been unable to
validate: the map described five corpus layers where the frontier enumerates
six.

Why it failed. The Gap earns its cost only where the claim space is too
large to check. The map makes ~100 enumerable claims. Check them all, and a
mirage becomes impossible — there is nothing to tune to. What that domain
wanted was an integrity gate, not an adversary, exactly as this catalogue
already classifies the substrate's audit triad and the publishing loop. The
lesson was on the page the whole time: the harness transfers to any pipeline
where generated artifacts must survive adversarial scrutiny.

What it gave the core. The mis-gated critic classification — the only one
that accuses the config rather than the proposer, because structural,
probe-limited, and noise leave a critic nowhere else to look. And
objective.canary: name an artifact that passes your gate by construction,
or you will not be able to tell a bad candidate from an impossible gate.

A harness is for adversaries. An auditor is for facts. Knowing which you have
is the first design decision, and the cheapest one to get wrong.

At a glance —
objective (historical): words of a corpus map, ↓ · gate: 8 comprehension claims hashed from the map — drawn from 612 KB of sources a 7 KB map was never responsible for carrying; the feasible set was empty · hard constraint: every claim traces
the Gap: a sampled draw it never needed — the map makes ~100 claims and every one is enumerable · lenses: by-container ⊥ by-content — both refused at 0/8 while both beat the metric · canary: MISSING, and that was the diagnosis: without one you cannot tell a bad candidate from an impossible gate. objective.canary was born here.
weight: RETIRED — replaced by universe/audit.mjs, exhaustive and agent-free, which found on its first run what two rounds and ~1.4M tokens could not · the lesson it carries: a harness is for adversaries; an auditor is for facts.

Navigation

← Welcome Visitors · The Harness Paths · Adopting The Harness · Workflows And Runtimes

the wiki is a projection; ~/dual-agent-harness (git) is the source of truth · (⚔️⊥⿻⊥🧙)😊 = neg ⊕ bnot → succ