A governed agent runtime: clarify the input, gate the outcome, compound the skill

1. Contribution / Thesis

Claimed. Two sovereign gates wrapped around agent execution are sufficient to address two interaction failures (not generation failures): a Clarification Runtime on the input (clarify under-defined goals, asking only decision-critical questions) and an Outcome Runtime on the stop (completion expressed as the predicate Target State Reached AND No Remaining Mandatory Actions). Both follow the discipline "a model may propose; a sovereign predicate disposes" over an event-sourced belief log. Successful execution crystallizes resolved decisions back into that log, so a repeat task pre-binds prior learning — a workflow is the residue of a completed task, not an up-front script. A single end-to-end prototype wires Goal -> Clarify -> Define-Outcome -> Execute -> Verify -> Crystallize and produces the thesis metric directly: questions asked fall with experience (n1=3 -> n2=0), with a one-slot re-ask under retraction.

NOT claimed. This is an architecture paper backed by a single prototype run, not a benchmark-scale empirical result. It does not claim: that the model-elevated "ceiling" (free-text goal -> decision slots, risk inference, optimal question wording) is solved — that needs a capable model; that the loop is production-hardened beyond one end-to-end test; domain-general "always correct" completion; or any generation-quality / throughput / latency result (none is measured). The altitude is honest: this is a design contribution with a prototype existence-proof, distinguishing a product problem (stop gracefully under uncertainty) from an AGI problem (always be correct), and asserting only the former is solvable now.

2. Design

Two gates surround execution; both read/write an event-sourced belief log whose events (assert / update / supersede / retract) reduce to a current state.

Floor vs ceiling (honest split).

Crystallization (procedural memory). Successful execution writes resolved decisions into the belief log; a repeat task in the same class pre-binds them, so experience becomes a governed, reusable capability rather than dying at session end.

3. Evidence / Status

This is prototype-level evidence, not a measured benchmark. A single end-to-end prototype facade composes the clarify / state / skill / epistemic-guard modules, and one test exercises the full loop.

The n1=3 -> n2=0 drop and the one-slot re-ask under retraction together are the existence-proof for the thesis (experience monotonically reduces questions; retraction is correctly localized).

4. Limitations (harsh, honest)

• Sample size is a smoke test, not an evaluation. The entire empirical claim rests on one end-to-end prototype test over one task class with a single retraction negative control. n1=3 -> n2=0 is two integers from one run — no variance, no repeats, no distribution of task classes. It demonstrates the mechanism fires; it does not measure how often it fires correctly.
• No model in the loop. Only the model-free floor is exercised. The ceiling (parsing free-text goals into decision slots, inferring risk, wording questions) is entirely unproven and explicitly model-bound. The "decision-critical slots" are, in the prototype, presumably pre-specified rather than inferred from natural language — the hard part is deferred.
• Predicate correctness is assumed, not validated at scale. "Target State Reached AND No Remaining Mandatory Actions" is only as good as the predicate definitions; the prototype does not show the predicate is correct across adversarial or ambiguous terminal states.
• Not production-hardened. No concurrency, failure-injection, persistence-durability, or long-horizon belief-log compaction is evaluated.
• evidenceLevel = experimental may flatter the work. A single prototype run is closer to a proof-of-concept than to repeated experimentation; the label is retained but readers should weight it as prototype-grade. A sovereignty/parity check says nothing about outcome quality.
• Scope. Applies to tasks where a good outcome is recognizable but the procedure is hard to specify. It does NOT apply to domains requiring always-correct completion, nor to realizing the elegant ceiling without a capable model behind the gates.

5. Provenance / Relation

• Components: a prototype facade composing the clarify / state / skill / epistemic-guard modules; an event-sourced belief log whose events (assert / update / supersede / retract) reduce to a current state.
• Build / run: compiled in release mode on Apple Silicon (M-series) hardware. Exact commit hash and version pins were not recorded in the source notes.
• Validation discipline: sovereignty checked by confirming no third-party ML framework participates in the loop; a single end-to-end test asserts n1=3, no-execute-before-answers, execute+verify+crystallize, n2=0, and one-slot re-ask under retraction.
• Relation: positions a product problem (graceful stop under uncertainty) as solvable now versus an AGI problem (always correct) as out of scope.