Hermes Labs

Project Fiction · Experimental apparatus

Fiction-CWR

Inspect the task, the contract, and the evaluator.

Version 0.1.1 · Offline reference implementation. No recorded model runs.

← Back to Project Fiction

A small world, an exact contract

Worked apparatus example · Existing public token fixture · Not a model experiment

Begin with two ore tokens, one flux, and one catalyst. The goal is exactly one widget with the catalyst intact and no other tokens remaining. There are at most four actions and four units of cost; every action costs one.

ActionConsumesProduces
smelt1 ore1 blank
coat1 blank + 1 flux1 coated
rivet1 blank + 1 coated1 widget

Each action requires the catalyst without consuming it. The inventory cap is six; the catalyst must remain at least one and widgets at most one.

A valid four-action solution
  1. smelt: one ore becomes a blank.
  2. smelt: the remaining ore becomes a second blank.
  3. coat: one blank and the flux become a coated token.
  4. rivet: the remaining blank and coated token become the widget.

Only the catalyst and widget remain. The plan meets the exact goal and both budgets.

Why a plausible-looking plan fails

coat → smelt → smelt → rivet fails immediately: there is no blank available to coat in the initial state. The checker evaluates each transition, not the plausibility of the written explanation.

The evaluator checks actions, resources, constraints, budgets, and the final state. A separate search implementation finds reference solutions; it shares fixture validation, which limits its independence. Exact validity in this world does not identify an innovation mechanism.

Inspect the exact fixture

Inside the apparatus

The implemented token workshop defines a finite world, legal actions, resource constraints, and an exact goal. The evaluator checks submitted plans; an independent transition/search implementation finds reference solutions. Shared fixture validation limits that independence.

Evaluator source is provided for inspection. Private witness files and protected execution materials are not included in this presentation. No browser model execution is offered.

Scope and limitations

Offline checks establish behavior within the supported token contract. They do not establish an innovation event, absence of a composition from pretraining, a human-relative model deficit, or reward suitability.

Source worlds retain the 2-versus-13 valid-plan mismatch. Matching and private execution gates are still open.

EXP-0002 · Closed-world recombination