Superalignment

Abstract

AI-assisted workflows, from "vibe coding" to tool-using agents, now turn lossy natural language into executable behavior. Yet behavior that looks correct under a demo, visible review, or task-completion score can still fail under hidden world states, scenario-dependent requirements, or operational constraints. We call this failure false convergence: the trajectory appears ready according to visible proxies while hidden requirements, critical blockers, or untested world outcomes remain unsafe, unsatisfied, or unmeasured.

We introduce Convergence Programming as a response to false convergence. Rather than treating AI-assisted creation as a prompt or final artifact, Convergence Programming treats it as a world-grounded trajectory: prompts, generated procedures, actions, observations, feedback, repairs, and act/ship/block decisions over time. The central claim is that AI-generated behavior fails not only because generation is weak, but because readiness is information-limited: a system cannot certify distinctions its trajectory has not extracted from the world. We recast a set of information-theoretic and probabilistic limits in a common convergence notation, organizing them into one account of when behavior can be certified as ready. As a scoped agentic extension, we compose autonomy accounting with a common-mode verifier model to derive a conditional action-window bound, implying that persistent autonomy must be receding-horizon rather than one indefinitely valid approval under stale evidence. We specify a preregisterable evaluation design centered on hidden-constraint discovery and calibrated permission to act, instantiated in a worked clinic-intake world.

The figures on this site illustrate the arguments. They do not report measured results or establish that this approach outperforms other development workflows.

Contributions

  1. A definition and theory of false convergence.The failure mode in which visible interaction, user acceptance, model confidence, known-ledger scores, or other observable proxies suggest readiness while hidden requirements, critical blockers, or untested world states remain unsafe or unmeasured, formalized as information-limited under bounded evidence.
  2. Convergence Programming as a response.AI-assisted creation defined as a world-grounded trajectory from lossy intent signals to executable behavior; as a scoped agentic extension, autonomy accounting and verifier correlation combine into a conditional action-window budget rather than a standalone certificate of long-running operation.
  3. Design criteria for convergence evaluation.Convergence Worlds: partially observable executable environments with hidden constraints, action surfaces, evidence envelopes, autonomy budgets, discoverability invariants and preregisterable metrics, instantiated in the clinic-intake world.

The named results, in plain statements

Formal statements and proofs are in the manuscript. Each result below is conditional on the paper's boundedness assumptions: scoped executable worlds, finite observation histories, evaluator ledgers, defined action surfaces, and evidence envelopes.

Latent operational target and intent ledger
The hidden target behavior is supposed to satisfy is represented for evaluation by a ledger of anchors with weights, severities, visibility status and satisfaction scores. Three residual gaps follow: full (all anchors), known (anchors revealed so far), unrevealed (the rest). The full ledger is an evaluator instrument, not available to the acting system.
Specification debt
A prompt is an information bottleneck. Constraints neither encoded in it nor revealed by later interaction are absent from the evidence, and no model capability turns unobserved world structure into evidence.
No free readiness
If two worlds produce the same observation history but demand different safety decisions, no rule that sees only that history can certify both. The distinguishing observation must be made, elicited, monitored, or carried as open risk.
Proxy improvement is not full convergence
A known-ledger or visible score can improve while full-ledger behavior does not. A large difference between known and full gaps is a false-convergence signature.
Critical blockers are not averages
One unresolved critical blocker changes the type of decision. Under any average-gap threshold, enough routine successes hide a single critical failure; readiness requires a lexicographic blocker gate.
Autonomy spends evidence
Every accepted action draws down a finite budget of residual uncertainty. Given an independently provenanced critical-risk bound and a common-mode verifier model, the number of actions a body of evidence licenses is bounded: the conditional action window. Persistent autonomy must be receding-horizon.
Repair can be anti-convergent
A repair that fixes the latest visible issue can corrupt prior anchors or optimize a proxy. The positive counterpart is conditional: if repair is contractive with respect to the latent residual gap, convergence is reachable at bounded oversight cost, with interactions logarithmic in the desired precision.
Evidence must be independent, scoped, and horizon-bounded
Verifier multiplicity is not independence: shared families, prompts, data and frames impose a common-mode risk floor. Self-reported evidence not bound to independent observation collapses evidence-gated readiness into self-report. Evidence certifies systems under conditions, not globally.
True convergence may surface bad news
Discovering a hidden anchor can reduce epistemic uncertainty while making the measured known-ledger gap temporarily worse. Non-monotonic known scores are not necessarily failures.
The architecture rule
If a design cannot distinguish critical hidden scenarios, preserve discovered anchors across repair, bind readiness claims to independent evidence, veto known critical blockers, or scope evidence to the deployment envelope, then there is a bounded world consistent with its observations in which it falsely converges.

Relation to existing approaches

Requirement discovery, iterative refinement, external feedback, and regression testing have established roles in software development. Convergence Programming proposes evaluating their connection across one recorded trajectory, including the evidence behind permission to act. The approaches below overlap; this comparison describes their focus and the proposed framing, not measured differences in performance.

ApproachFocusFeedback already availableConvergence Programming's proposed emphasis
Traditional developmentSoftware built from an evolving understanding of the workUser collaboration, prototypes, tests, review, and operationEvaluate the history of requirement discovery and preservation alongside the artifact and its tests.
End-user programmingSoftware created by the person who knows the workDirect use, logic inspection, examples, and colleague feedbackRetain the requirements and evidence discovered during use, with their conditions of validity.
Programming by demonstrationProcedures inferred from demonstrations or examplesAdditional examples, corrections, and interactive refinementConnect demonstrations to a broader record of observations and the actions they support.
Program synthesisA program satisfying formal, partial, or example-based requirementsCandidate checking and counterexample-guided refinementRecord how operational requirements enter the specification and where its guarantees apply.
Prompt to codeApplications created and revised through natural languageFollow-up questions, execution, testing, and reviewEvaluate the whole build trajectory, including requirements discovered after the initial prompt.
Tool-using agentsPlanning and acting through tools toward a goalTool results, environment observations, execution checks, and human feedbackConnect accumulated evidence and preserved requirements to scoped permission for subsequent actions.
Agent and code benchmarksPerformance measured under a defined evaluation protocolTests, held-out scenarios, environment outcomes, and trajectory checksMeasure hidden-requirement discovery, evidence sufficiency, critical blockers, and decisions to act or seek information.
Convergence ProgrammingThe recorded path from partial intent to executable behavior in a scoped worldActive observation, requirement confirmation, repair, and regression checksA proposed framework connecting discovery, preservation, evidence, and permission. Discovery may remain incomplete; comparative outcomes require measurement.

The Agile principles explicitly welcome changing requirements and collaboration. Syntax-Guided Synthesis describes counterexample-guided refinement against a stated specification. Anthropic's Building effective agents describes grounding progress in tool results and code execution. These are foundations and neighbors of the proposed framework.

The epistemic contract

Convergence Programming is bounded not because it is small, but because bounded worlds make oversight measurable.

BoundaryMeaningWhy it matters
TaskScoped executable world, not all valuesPrevents overclaiming superalignment
WorldEnvironment modeled, simulated, sampled, or operationally scopedMakes grounding possible
LedgerHidden but finite or sampled evaluator targetMakes scoring possible
ObservationClaims limited by observation historyGives the non-certification theorem force
ActionDefined action surfaceLets risk scale with autonomy
BudgetOversight costs questions, scenarios, experts, tools, computeConnects to scalable supervision
EnvelopeEvidence valid only under stated conditionsHandles runtime drift and deployment shift

Citing this work

Until the paper is released, cite the manuscript:

Pantasdo, K. D., & Wu, P. (2026). Convergence Programming: Iterative, World-Grounded
Alignment of Lossy Human Intent with Executable Behavior. Manuscript in preparation.
https://superalignment.inc/research/convergence-programming/