Scaling alone cannot absorb their system responsibilities. The position paper develops the A² taxonomy, three major hypotheses, and the prospective tests that could falsify them.
Unpublished manuscript. Hypotheses and protocols are prospective; no causal result is claimed.
01
RESEARCH AGENDA / H1—H3
From position to testable hypotheses.
Three connected hypotheses ask what remains stable, how agency is constrained, and whether explicit variation boundaries improve self-modification. Each claim keeps its evidence and falsification boundary in view.
Read the paper PDF ↗Manuscript context: §3, “Harness Design Patterns and the A² Taxonomy”; §4, “Three Major Hypotheses”; and §6, “Future Work: Prospective Tests of the Three Hypotheses.” The hypotheses below are an agenda, not completed results.
FIG. R / RESEARCH MAP
H1 identifies the asset to preserve. H2 makes separation operational across A². H3 turns the A3 evolution seam into a prospective comparison—not an observed benefit.
H1
ASSET HYPOTHESIS
Stable assets are design commitments.
Claim
Under rapid prompt/code rewriting, stable assets are inspectable design commitments—not code or prompts: pointcut + contract + permitted variation + test obligation + negative control.
Metrics
Contract retention: protected obligations that remain explicit and satisfied after rewriting.
Contract drift: unauthorized semantic change across successive versions.
Evidence status
Conceptual + exploratory detection/source evidence. Existing support can motivate and locate commitments; it does not establish their utility.
Open / falsification boundary
The causal utility remains open. The hypothesis would be weakened if matched mechanism-only descriptions retained contracts as well and drifted no more under the same rewriting regime.
H2
AGENCY HYPOTHESIS
Separation must be mechanized.
Claim
When agents evolve their own harness, separation of concerns must be mechanized through guards, gates, checkpoints, rollback, and evaluation gates; docs/comments alone are not constraints.
Metrics
Unauthorized-variation escape rate and violation containment.
Checkpoint recovery and rollback success, latency, and evaluation-gate rejection of known violations.
Evidence status
Analytical/source support only.E-family records are candidate mechanisms, not a demonstrated comparative advantage.
Open / falsification boundary
Comparative causal benefit remains open. A controlled comparison that finds no improvement over documentation/comment guidance in containment, recovery, or violation detection would challenge the claim.
H3
CONTROLLED-VARIATION HYPOTHESIS
Bound the change to clarify the outcome.
Claim
Self-modification with explicit variation boundaries should yield lower contract drift, lower regression rate, faster failure recovery, and clearer change attribution than unconstrained self-modification.
Metrics
Contract drift and regression rate per modification round.
Time to successful recovery and completeness of change attribution.
Evidence status
Prospective / causal-open. There is no current experimental evidence for this predicted comparison.
Open / falsification boundary
The entire causal claim is open. Equivalent or worse outcomes for bounded modification would falsify the predicted benefit under the tested conditions; mixed results would narrow, not confirm, its scope.