Research Agenda

Research question

Can an autonomous system change what it believes and does without losing the path that warrants the change?

Here, warrant means what the available evidence actually justifies. Palimpsestus asks whether provenance, warrant, inherited state, uncertainty, and accountable transition can be built into adaptive systems as architectural constraints rather than reconstructed after a decision is made.

Central research bet: provenance-preserving constraints may produce behavioral regularities that survive changes of implementation and substrate. That is a hypothesis, not an established result.

Frozen foundation

Current ARC experiments are pinned to the Palimpsestus Open Spec v1.1.3 under its C₁ constitution. The architecture separates specification, conformance, and implementation so implementations can be tested against a stable public reference. ARC is the current controlled testbed, not the intended scope of the architecture.

Experiments follow a No-Ad-Hoc-Rescue rule: a failed result cannot be rescued by adding machinery to the frozen architecture after the fact.

Evidence so far

ARC-AGI-2

A frozen implementation satisfied 192/192 semantic obligations and 160/160 derivation-ancestry edges. Confirmatory experiments reproduced a preregistered structural pattern associated with Fold compression. Controller endpoint efficacy was null, and preservation of a useful operational horizon under finite resource budgets remains unresolved.

ARC-AGI-3

On a preregistered paired Crucible using official ls20, provenance-faithful recurrence produced 0 repeated disproven actions in 250 decisions, compared with 35/250 (14%) for a matched learner. A su15 control showed no measurable behavioral separation, and neither architecture achieved level progress.

An official ARC-AGI-3 Kaggle submission also completed with publicScore = 0.08. The sealed record does not yet establish that the score came from Palimpsestus agent execution on hidden tasks, so no stronger performance claim is made.

What remains unresolved

• Do independent implementations converge on a common substrate-independent kernel?
• When two agents reach apparently identical present beliefs through different evidence histories, should they have identical future affordances?
• How much history can be compressed without losing future warrant?
• Can improved epistemic behavior coexist with strong goal-directed utility?
• Does Palimpsestus produce behavior that existing approaches do not already explain?
• Are apparent recurrences across domains genuine structural invariants or post-hoc analogies?

Next experiments

The immediate priority is a prospective comparison between Palimpsestus, Active Inference, NARS/AERA, and a matched conventional learner under identical observation, action, memory, and resource constraints.

Tests should target cases where the architectures ought to make different predictions: conflicting evidence, different evidentiary histories, compression, recurrence of disproven actions, revision, uncertainty, and resource starvation.

A second track will test the same frozen architecture in independent external domains such as autonomous synthetic worlds, aerospace autonomy, and provenance-intensive distributed systems.

The domain must consume Palimpsestus. Palimpsestus must not be rebuilt to fit the domain.

Collaborators sought

Especially useful collaborators include researchers in formal methods and mathematical systems; autonomous agents, Active Inference and cognitive architectures; distributed systems and provenance; experimental methodology and independent evaluation; aerospace, robotics, simulation, games, and persistent synthetic worlds.

Belief in Palimpsestus is unnecessary. Competent attempts to break it are particularly useful.