Clarify Runtime: the model-free floor of the input gate

Contribution

Claimed. A model-free input-gate prototype, built only on a state runtime plus an epistemic guard with no language model loaded, passes an 8-test probe covering four behavioral gates: (G1) when a decision-critical slot is unbound it clarifies rather than silently proceeding, with a negative control showing a fully-specified task proceeds (so the gate is not "always ask"); (G2) it asks only decision-critical questions, reducing a naive question set of four to three (cosmetic defaults auto-bind; questions are capped at the top three by a risk ordering of irreversible > policy > impactful); (G3) it emits an outcome predicate (target state reached AND no remaining mandatory actions) that distinguishes approval-pending (not done) from the same terminal state reached by a different path (done); and (G4) it demonstrates the thesis metric — decision-critical questions decreasing with experience, from three to zero after crystallization (partial crystallization yields one; a retract step makes the gate re-ask).

NOT claimed. This is a correctness/behavioral probe of a floor, not a novel algorithm and not a product. It does NOT claim free-text goal understanding, optimal question wording, or any model-bound "ceiling" behavior — those explicitly require a capable model. It makes no throughput, latency, generation-quality, or benchmark-scale claim. The eight tests are a behavioral smoke test of predicate logic over hand-constructed slot scenarios, not an evaluation over a realistic task distribution. No always-correct completion guarantee is implied.

Method

This is a predicate/behavioral probe, not a measured-performance study; most Methods fields standard to empirical notes are not applicable. The table records what the source specifies.

Results

Only outcomes stated in the source are reported. All four gates pass (8/8 tests).

Negative controls

Threats to validity / Limitations

• Sample size. Eight hand-constructed predicate tests over slot scenarios constitute a behavioral smoke test, not a benchmark-scale evaluation over a realistic task distribution. Coverage of slot/risk combinations beyond those eight is unverified.
• Predicate-level, not end-to-end. The probe validates gate logic on pre-formed decision-critical slots. It does NOT exercise free-text goal -> slot extraction, which is explicitly out of scope and model-dependent.
• Floor, not ceiling. Results bound only the model-free floor (safety disposition via sovereign predicates). They say nothing about question-wording quality or optimality, which need a capable model.
• Parity != quality. Passing the gates demonstrates correct disposition, not that resulting clarifications are well-phrased or that downstream task completion is correct. No always-correct completion guarantee.
• No measured performance. No hardware detail, stack versions, timing, or oracle/baseline are recorded; there are no performance numbers and thus no variance/single-run concerns to report, but also no quantitative external validation.
• EvidenceLevel. Tagged "experimental"; given the 8-test predicate scope this sits at the low end of that level — a correctness probe rather than a controlled experiment over data.

Provenance / Relation

• Artifact: model-free input gate = state runtime + epistemic guard, with no language model loaded.
• Relation: front-half probe of the broader clarify-then-act loop; the complementary model-bound "ceiling" (free-text goal parsing, optimal wording) is explicitly deferred and out of scope here.
• Build/run: compiled and tested as a release build via the standard toolchain (details omitted).
• Sovereignty: no third-party machine-learning framework is linked; only the platform's native numerical library is used.
• Version pins: not recorded in source notes.