PaperlayerView exact revision on arXiv

Source-linked AI summary

Governance Records as Supervision: Verifier-Selected Self-Training for Structured Workflow Repair

Jesus Salas

arXiv:2608.18324v1cs.AI

TL;DR

The paper asks whether runtime verification records can supervise bounded models rather than merely reject invalid outputs. It trains same-checkpoint models on independently verified plans and finds verifier-selected supervision substantially improves fresh-case execution, while remaining limited to one synthetic planning domain.

  • Problem

    The paper asks whether runtime records from mechanically verifiable workflows can become post-training supervision instead of serving only for inference-time rejection.

  • Method

    The study trains bounded models on independently verified outputs whose governance records preserve contract, verifier decision, target origin, and lineage.

  • Results

    102/160 accepted plans followed VAL selection versus 69/160 with model self-selection, a paired net advantage of +33 across both difficulty strata.

  • Takeaways & Limitations

    Verifier-selected model outputs can consolidate occasional or expensive capability into more consistent one-shot execution within machine-verifiable task contracts.

  • Takeaways & Limitations

    The evidence comes from one public synthetic PlanBench Blocksworld domain with 5- or 6-block replanning cases, not arbitrary planning or enterprise work.

Abstract

from arXiv · show

Machine-verifiable workflows produce governance records linking a task contract, model attempt, verifier decision, accepted output, and target origin. We test whether these records can supervise bounded models, consolidating occasional or expensive capability into reliable one-shot execution. On fresh, structure-disjoint PlanBench replanning cases, Qwen3-14B thinking generated 24 plans admitted by the independently authored VAL verifier. Those plans trained the same checkpoint for non-thinking execution, without oracle targets or a stronger teacher. On 80 unopened cases, VAL-accepted plans increased from 1 to 57, with 56 paired gains and zero regressions; thinking reached 30. The adapter was schema-valid on all cases and used approximately 1/56 of thinking's mean latency. The separate paired interface-cure gate did not pass. A matched ablation fixed the source cases, 52-candidate pool, 24-target count, model, recipe, and seed while changing target selection. On 160 new cases, base, schema-selected, model-self-selected, and VAL-selected execution reached 1, 55, 69, and 102 accepted plans. VAL exceeded self-selection by paired net +33 (p=0.0000019647), with gains in both difficulty strata. Independent semantic selection is therefore load-bearing relative to matched alternatives within this band. A complementary Phi stronger-teacher arm raised base Phi-4 from 2 to 51 accepted plans and from 35 to 80 schema-valid outputs. Earlier synthetic experiments establish teachability, cumulative learning, construction robustness, and stopping boundaries. The results support verifier-selected supervision for bounded, machine-checkable capabilities, not arbitrary planning, enterprise validity, or unrestricted self-improvement.

1 Introduction

The paper asks whether verifier-admitted workflow outputs can become governance-record supervision, consolidating occasionally demonstrated or expensive capability into reliable one-shot execution. External evidence supports this mechanism for bounded, machine-verifiable planning, while remaining limited to one synthetic domain, difficulty band, and training setup.

  • External self-distillation: 24 verifier-admitted Qwen3-14B thinking plans trained the same checkpoint for non-thinking execution without oracle targets or a stronger teacher.Fresh, structurally disjoint cases used an independently pinned VAL verifier, with harvest and evaluation partitions kept disjoint.
  • External self-distillation: 57 verifier-accepted plans versus 1 for base non-thinking and 30 for base thinking showed 56 paired gains with zero regressions on 80 unopened cases.The adapter was schema-valid on all 80 cases, improved both difficulty strata, and used approximately 1/56 of thinking latency.
  • Target-selection ablation: 102 VAL-accepted plans versus 69 for model self-selection and 55 for schema validity alone showed that independent semantic verification was load-bearing in the matched ablation.The case IDs, candidate pool, target count, checkpoint, recipe, and seed were fixed across 160 new cases.
  • Stronger-teacher arm: 51 verifier-accepted plans versus 2 for base Phi-4 demonstrated the complementary stronger-teacher route, while schema-valid outputs rose from 35 to 80.Phi-4-reasoning generated the 24 VAL-selected plans; this arm is stronger-teacher distillation rather than teacher-free self-training.
  • Core proposal: Governance records preserve the contract, model attempt, verifier decision, accepted output, producer, student lineage, and gate outcome for reusable supervision.The external verifier remains the acceptance control while low-rank adaptation consolidates recurring verified behavior.
  • Scope and boundaries: The evidence supports verifier-selected supervision within machine-verifiable task contracts, not arbitrary planning, enterprise validity, or unrestricted self-improvement.Earlier synthetic experiments establish teachability, cumulative learning, construction robustness, selective transfer, frontier movement, and measurement-specific stopping.

2 Problem Formulation

The paper formulates structured repair as one-shot generation judged by a deterministic verifier under a public contract. It distinguishes supervision surfaces and target origins while defining research questions around teacher-free consolidation, selection, provenance, synthetic mechanisms, and claim boundaries.

  • 2.1 Verifiable structured repair: A public contract C specifies tasks, actions, dependencies, state transitions, resource constraints, or completion conditions for candidate plan y.The deterministic verifier V(C, y) returns transport and semantic verdicts together with failed checks F.
  • 2.1 Verifiable structured repair: A repair succeeds when one response emits ŷ that is transport-valid and has no failed checks, measuring adaptation rather than inference-time search depth.Correctness is verifier-confirmed within the declared contract, not objective correctness for institutional work.
  • 2.2 Three supervision surfaces: The supervision surfaces are a verdict, a localized governance trace, and a complete verified target that passes V(C, y+).The primary LoRA treatment contains localized feedback and correct targets, so a matched ablation is needed to attribute their roles.
  • 2.3 Target origin: Target validity and target origin are separate variables, including procedural oracle, verifier-selected, same-checkpoint self-distillation, stronger-teacher, diagnosis-steered, and human-adjudicated targets.The taxonomy distinguishes how valid targets are produced and used across experiments.
  • 2.4 Research questions: The research questions test external teacher-free consolidation, independent semantic selection, stronger-teacher provenance, controlled synthetic mechanisms, and boundaries on broader claims.The boundary question includes iteration, search reachability, cross-construction transfer, interface reliability, and frozen continuation gates.
  • 2.4 Research questions: The reported conclusions are positive for external semantic consolidation, matched selection alternatives, stronger-teacher classification, and controlled synthetic findings, but bounded by stated transfer and evaluation limitations.The paper notes selective transfer, saturated measurements, non-material later rounds, one external planning band, one seed per adapter, and no institutional work.

3 Governance Records as Supervision

Governance records connect contracts, model outputs, verifier decisions, accepted targets, and provenance so only admissible passing outputs enter training. The protocol separates search yield, one-shot execution, and cumulative training through case-disjoint, predeclared gates, while treating the verifier as an admission function rather than an oracle or author.

  • Record construction: Five linked objects—contract, model output, verifier result, target origin, and experimental gate—define each governance record.A passing output becomes a training target only when its origin is admissible; failed outputs remain in provenance but cannot enter training as accepted targets.
  • Causal scope: The verifier admits candidates semantically, but training performs consolidation and the ablation isolates selection value rather than the full training effect.The verifier neither authors candidates nor establishes that the contract captures institutional truth.
  • Provenance: Target-hidden records retain prompt hashes, model and adapter identity, serving regime, sampling configuration, verdicts, token counts, and attempt indices.The first verifier-passing candidate within a frozen per-case budget becomes admissible, while rejected predecessors remain retained.
  • Evaluation gates: Three effects are distinguished: search yield, one-shot evaluation on fresh cases, and cumulative training that supports another adapter.Further rounds require a predeclared gate on a case-disjoint partition, and unresolved rows cannot be backfilled with procedural targets under a self-sourcing label.
  • Evaluation criteria: External evaluation requires parseability under the frozen output contract and zero failed checks from the deterministic workflow verifier.Transport validity is reported separately so semantic gains are not conflated with formatting repair.
  • Limitations: Fixed-budget results measure bounded empirical search rather than absolute model reachability, and frozen benchmark tests do not establish population-level certainty.Repeated seeds and broader task samples remain necessary for broader statistical claims.

4 Core Evidence

On fresh PlanBench replanning cases, verifier-selected self-training converted occasional Qwen and Phi reasoning capability into substantially more reliable one-shot execution. Matched ablation evidence shows VAL-based target selection contributes independently, while latency and synthetic results define operational and scope boundaries.

  • Qwen external result: 57/80 VAL-accepted plans followed Qwen self-training, versus 1/80 for base non-thinking and 30/80 for base thinking.The adapter produced 56 adapter-only successes and zero base-only regressions, with gains in both block-count strata.
  • Matched target-selection ablation: 102/160 VAL-accepted plans followed VAL-selected training, versus 69/160 for model-self-selected and 55/160 for schema-selected adapters.Against model self-selection, verifier selection produced paired net +33, with exact p = 1.9647 × 10−6 and positive gains in both strata.
  • Phi stronger-teacher arm: 51/80 verifier-accepted plans followed Phi distillation, versus 2/80 for base Phi-4 and 30/80 for Phi-4-reasoning.Schema validity also rose from 35/80 for base to 80/80 for the distilled adapter, with zero paired losses against the teacher.
  • Operational consequence: Qwen execution used approximately 1/56 of thinking’s mean latency, while Phi used approximately 1/153 of the reasoning teacher’s latency.The corresponding completion-token ratios were approximately 1/86 for Qwen and 1/215 for Phi.
  • Scope and boundaries: Synthetic experiments established cumulative gains, construction-separated replication, and measured stopping and transfer boundaries rather than unrestricted improvement.Generator A supported two material cumulative gains, Generator B two one-response gains, and Generator C reached its measurement ceiling after one round.

5 Discussion

The discussion argues that verifier-selected records consolidate occasional, expensive planning behavior into reliable one-shot execution on fresh cases. It also limits the claim to bounded, machine-checkable capabilities, distinguishing same-checkpoint self-distillation from stronger-teacher distillation and identifying clear mechanism boundaries.

  • Core result: Qwen training produced 56 paired gains, zero regressions, both-stratum coverage, and 80/80 schema validity on fresh external cases.The result combines expensive thinking-generated accepted plans with cheaper non-thinking execution by the same checkpoint.
  • External validity: PlanBench and VAL show the mechanism survives an externally authored task specification and verifier, leaving the remaining externality gap institutional.Earlier self-authored generators established the mechanism and its iteration boundaries.
  • Matched ablation: 102/160 for VAL-selected execution versus 69/160 for model-self-selected execution shows independent semantic admission matters under matched targets, cases, candidates, model, recipe, and exposure.The model selector chose semantically valid plans in 20 of 24 cases and still beat schema-only selection, so it was not a straw baseline.
  • Evidence distinction: Qwen supports teacher-free same-checkpoint self-distillation, whereas Phi supports governance-selected distillation from a separately trained reasoning teacher to a base student.Combining the arms would overstate the Phi evidence and understate the Qwen evidence.
  • Mechanism: The narrower mechanism is filtered consolidation: verifiers select occasional successes from noisy reasoning samples, and supervised training shifts one-response probability toward accepted behaviors.This can make a narrow specialist more consistent on the tested contract without showing general superiority over its source regime.
  • Boundaries and deployment: Verifier-selected training requires admissible fuel in the sampling distribution, does not create targets absent from search, and does not guarantee uniform transfer across constructions.The measured architecture separates offline acquisition, governance records, cheap routine execution, and verifier-controlled retry, rerouting, or escalation.

6 Limitations and Evidence Boundaries

The evidence is bounded to synthetic, structure-limited planning studies and isolated training runs, so it does not establish broad planning, institutional, or production claims. Several comparisons and interface results support narrower operational conclusions rather than superiority, universal laws, or deployment readiness.

  • External scope: External validation covers only fresh PlanBench Blocksworld replanning with 5 or 6 blocks, not harder bands, arbitrary planning, or independently authored enterprise work.PlanBench and VAL provide task and verifier authorship, while the study controls case construction, contamination checks, prompting, target selection, training, and analysis.
  • External scope: Each external adapter is one training run with one optimizer seed and one frozen recipe, so paired results describe this corpus rather than training-seed variance or a broader planning population.Broader claims require multiple seeds, additional external domains, larger samples, and general-capability regression batteries; Qwen and Phi share the same 80 evaluation cases.
  • Interface boundary: Qwen’s semantic-only adapter produced 80/80 schema-valid responses and zero endpoint errors, but the preregistered paired interface-cure endpoint over thinking did not pass.The small three-case schema gain and p = 0.25 are not positive interface-cure evidence or evidence of an adapter defect.
  • Comparison boundaries: Student-versus-source comparisons were descriptive, while Qwen-versus-Phi results are confounded and cannot isolate the contribution of a stronger teacher.The matched Qwen ablation supports independent semantic verification as load-bearing relative to self-selection and schema-only selection within one candidate pool and seed, not a universal target-purity law.
  • Synthetic boundaries: Synthetic results use self-authored generators, clustered generated cases, unequal Qwen and Phi testing, and fixed repeated-sampling budgets, making them controlled results rather than prevalence estimates or model ceilings.Generator A transfer is family-selective, while Generator C starts with a high base rate and saturates after one round.
  • Deployment boundary: Verifier compliance establishes contract adherence, not policy legitimacy or real-world completion, and deployment properties such as escalation, routing, replacement, and specialist size were not tested.Deployment would require authority over contract changes, versioned applicability, grounded facts, external completion evidence, appeal, and human escalation.
  • Deployment boundary: Runtime measurements come from research paths on one A100, so latency and token ratios do not characterize production throughput, concurrent serving, queueing, or total oversight cost.They characterize the frozen experimental arms only.

7 Related Work

This work distinguishes verifier-selected provenance supervision from governed execution, self-training, distillation, RLVR, process supervision, and inference-time self-correction. Its contribution is the explicit separation of workflow-record roles and matched tests of admission, learning, iteration, cost, and stopping on a declared PlanBench/VAL surface.

  • Governed execution: The studied record narrows governed execution’s auditable lifecycle to a task contract, model attempt, verifier decision, accepted output, and target origin, then tests whether admitted records train a bounded specialist.The paper does not infer that provenance alone improves models or extend its evidence to decision or change governance.
  • Self-training and distillation: The synthetic programs are filtered self-training: models sample candidates, a deterministic verifier admits passing outputs, and LoRA consolidates accepted targets on case-disjoint evaluations.The external Qwen arm is same-checkpoint self-distillation, with thinking producing targets for non-thinking execution; the Phi arm uses an independently executed teacher.
  • RLVR and process supervision: Unlike RLVR, the method uses binary verifier admission followed by ordinary LoRA SFT on complete accepted targets, with no policy-gradient update or reward optimization.The verifier is shared infrastructure with RLVR, not the novelty claim.
  • RLVR and process supervision: Unlike process supervision and code-feedback systems, the verifier is not learned, the output is a structured plan rather than source code, and acceptance is defined exactly by the declared surface.This restricts claims to what that machine-checkable surface expresses.
  • Benchmark and contribution: Using PlanBench T6 replanning and pinned VAL, the study retains VAL-admitted runtime plans as provenance-explicit supervision and evaluates trained specialists on fresh cases rather than frozen models or retry-only systems.The distinguishing combination separates contract, attempt, decision, target, producer, and recipient, then tests admission, learning, iteration, cost, and stopping.

8 Conclusion … A.3 Qwen formal protocol

Verifier-selected governance records supervise bounded, machine-checkable workflow repair, converting occasional reasoning capability into faster, reliable one-shot execution. Fresh PlanBench results support Qwen self-training and stronger-teacher Phi distillation, while matched ablation and protocol limits constrain the claim.

  • 8 Conclusion: 57/80 verified plans followed Qwen self-training from 24 VAL-admitted targets, rising from 1/80 with 56 paired gains and zero regressions.The adapter was schema-valid on 80/80 cases and ran at approximately 1/56 of thinking’s mean latency.
  • 8 Conclusion: 51/80 verifier-accepted plans and 80/80 schema-valid outputs resulted when a separate reasoning teacher supplied targets for base Phi-4.This arm is stronger-teacher distillation rather than self-training; earlier synthetic work supplied teachability, cumulative-learning, replication, transfer, and stopping evidence.
  • 8 Conclusion: 102/160 accepted plans came from VAL selection, versus 69/160 for model self-selection and 55/160 for schema selection.VAL exceeded self-selection by paired net +33, with significance under the frozen gate and gains in both difficulty strata.
  • 8 Conclusion: The conclusion does not establish general planning, enterprise validity, unrestricted self-improvement, or model-size compression.Further domains, repeated seeds, institutional evidence, and a direct specialist-size study are required.
  • A External PlanBench Study: 80 evaluation cases were structure-disjoint from harvest and external published and prior experimental problems, with VAL as the only semantic admission authority.The corpus contained 32 harvest cases and 80 evaluation cases, while the frozen interface only canonicalized declared action and object sequences.
  • A.2 Qualification history and protocol separation: NO_GO_VALIDITY preceded the positive formal study: non-thinking Qwen3-14B and base Phi-4 found no VAL-accepted plan on 24 cases within eight attempts.A later smoke study authorized a new frozen protocol but was retained as directional rather than publication evidence.
  • A.3 Qwen formal protocol: The formal Qwen protocol selected the first VAL-accepted plan per successful case and trained the same Qwen3-14B checkpoint with thinking disabled.Each record preserved case, attempt, seed, hashes, model revision, serving regime, verifier commit, and verdict; training used fixed LoRA and optimization settings.

A.4 Qwen harvest and evaluation

Qwen3-14B produced 24 independently VAL-admitted training targets, yielding a 56/80 net gain over base non-thinking on fresh cases. The semantic endpoint passed, while the paired interface-cure gate failed despite perfect schema validity on the measured set.

  • Harvest: 24 admitted targets included 11 first-attempt-only successes, covered both block counts, and had 24 distinct canonical plan hashes.Training completed all 72 optimizer steps in 63.6 seconds.
  • External evaluation: 56/80 net VAL successes were adapter-only against base non-thinking, with zero base-only successes and exact p < 10^-10.The comparison also recorded one both-pass and 23 both-fail cases.
  • External evaluation: 34 adapter-only versus 7 thinking-only cases produced a descriptive p = 0.0000253208 against thinking.The comparison included 23 both-pass and 16 both-fail cases; superiority over thinking was not a predeclared endpoint.
  • Endpoint gates: The semantic endpoint passed in both strata, but the interface endpoint failed with 3 paired schema gains, 0 losses, observed p = 0.25, and a gate requiring 5 gains and exact p < 0.05.The combined reading was SEMANTIC_ONLY.
  • Endpoint gates: 80/80 responses were schema-valid with zero endpoint errors on the measured deployment interface.This perfect absolute interface result did not make the paired interface-cure gate pass.

A.5 Matched verifier-selection ablation

A matched ablation found that VAL-based semantic target selection outperformed model self-selection and schema-only selection on 160 structurally excluded replanning cases. The result supports verifier-selection advantage within the tested model, domain, harvest pool, seed, and difficulty band.

  • Experimental design: The ablation reused 24 successful harvest cases and 52 schema-valid candidates, while selection was isolated from candidate generation on 11 contested cases.Qwen thinking selected opaque, shuffled candidate IDs without rewriting, repairing, or combining plans.
  • Primary comparison: Paired net +33 favored VAL selection over model self-selection across all 160 cases, with exact two-sided p = 1.9647 × 10−6.There were 41 VAL-only successes and eight self-only successes.
  • Primary comparison: Paired net +10 in the 5-block stratum and +23 in the 6-block stratum favored VAL selection over model self-selection.The corresponding outcomes were 54/80 versus 44/80 and 48/80 versus 25/80.
  • Secondary comparisons: Paired net +47 favored VAL selection over schema-only selection, with p = 4 × 10−10.The comparison used the same frozen candidate pool, target count, model, training recipe, and seed.
  • Scope and limitations: The causal reading is limited to semantic selection relative to two matched alternatives within one model family, domain, 24-case pool, adapter seed, and difficulty band.It does not establish verifier-only causation for the full training effect or generalization to open-world correctness.

A.6 Phi protocol, audit, and evaluation · B Synthetic Development Program · B.1 Study map and controls

The Phi arm tested governance-selected stronger-teacher distillation under a frozen protocol, while synthetic studies mapped mechanisms, controls, and evaluation boundaries. Phi distillation substantially improved verifier acceptance and schema validity, with audits preserving sealed partitions and frozen settings.

  • A.6 Phi protocol, audit, and evaluation: Phi-4-reasoning served as the governance-selected teacher for base Phi-4 under the same corpus, prompts, verifier, recipe, and gates.The protocol was frozen before formal Qwen response partitions were opened.
  • A.6 Phi protocol, audit, and evaluation: 24/32 distinct targets were admitted, including 10 first-attempt successes and 24 successes within eight attempts.All 256 requests completed; acquisition used 1,725,255 completion tokens and 2.791 A100 hours, and training took 72 optimizer steps in 56.8 seconds.
  • A.6 Phi protocol, audit, and evaluation: 49/80 was the net adapter gain over base Phi-4, with zero base-only successes and exact p < 10−10.The comparison recorded 49 adapter-only successes, two both-pass cases, and 29 both-fail cases; both semantic strata passed their frozen gates.
  • A.6 Phi protocol, audit, and evaluation: 39 paired cases improved schema validity over the reasoning teacher, with zero losses and exact p < 10−10.The reported reading was FULL_BAND_DISTILLATION.
  • A.6 Phi protocol, audit, and evaluation: 51/80 distilled-student acceptance versus 30/80 teacher acceptance yielded p = 0.0010650196.The paired comparison had 30 student-only, nine teacher-only, 21 both-pass, and 20 both-fail cases, and was treated as a compression diagnostic rather than a preregistered superiority test.
  • B Synthetic Development Program: Synthetic studies provided mechanism, replication, and boundary evidence before the external experiment, serving as supporting rather than primary results.Generators A, B, and C differed in construction code, vocabularies, and mutation logic while sharing the public schema, deterministic verifier, base model families, and researcher.
  • B.1 Study map and controls: Self-sourced stages admitted targets only when sampled model responses were parseable and passed the frozen verifier, without procedural answers replacing failures.The oracle precursor used procedural schedules as targets; cumulative adapters were reinitialized from the original base.
  • B.1 Study map and controls: One-response verifier acceptance on an unopened partition was the main endpoint, while search studies additionally reported pass@32.Paired changes used exact two-sided McNemar tests, and frozen gates checked transport validity, minimum effect size, and family-level regressions.

B.2 Oracle teachability and learning-signal attribution

Oracle-target supervision substantially improves accepted repairs for both Phi-4 and Qwen3-14B, including compound cases, but the result is in-distribution oracle-target SFT rather than oracle-free learning. Within a matched Qwen comparison, complete accepted targets provide only a modest advantage over verdict-only traces, while localized residuals can aid separate inference diagnostics.

  • Oracle teachability: Phi-4 rose from 33/144 to 137/144 accepted repairs, while Qwen3-14B rose from 31/144 to 139/144.The precursor held the corpus, verifier, target plans, training recipe, and one-response evaluation fixed across both model families; adapter-only passes numbered 104 for Phi-4 and 108 for Qwen3-14B, with no base-only passes.
  • Oracle teachability: Compound-subset acceptance increased from 6/48 to 46/48 for Phi-4 and from 5/48 to 45/48 for Qwen3-14B.Every compound family appeared in training, so this establishes in-distribution rather than structure-disjoint teachability.
  • Target provenance: The training result is oracle-target SFT with deterministic verification, using 48 unique case-level targets reused across 192 compound rows.The verifier confirmed each target but neither searched for nor generated it.
  • Learning-signal attribution: 139/144 versus 133/144 accepted plans favored localized-trace training over verdict-only training, a 4.2 percentage point difference within the frozen five-point boundary.Both conditions used the same localized inference interface, and verdict-only training remained above the 90% sufficiency threshold.
  • Inference localization: 3/15 verified completions came from exact residuals versus none from generic retry in a separate Phi diagnostic.All three completions were in one co-location family, and the narrow diagnostic did not change the training attribution.

B.3 Verifier-selected learning, iteration, and stopping · B.4 Transfer and explanatory boundaries

Verifier-selected self-training yields cumulative, construction- and model-family-specific gains in fresh repair and fixed-budget reachability, with verifier-defined stopping points. Transfer is selective rather than universal: destination-native training resolves failures that transfer misses, and saturated evaluations constrain claims about unrestricted improvement or enterprise validity.

  • B.3 Verifier-selected learning, iteration, and stopping: 16 accepted targets from 32 cases enabled Qwen repair to rise from 10/48 to 21/48, with 14 paired gains and three losses.The targets came from model sampling followed by verifier admission and LoRA training.
  • B.3 Verifier-selected learning, iteration, and stopping: 22/32 fixed-budget reachability followed 13/32 for Generator A’s Default30 adapter, with nine paired gains and no losses; a 52-target continuation reached 35/48 versus 33/48.The continuation’s net two gains were below the frozen continuation threshold, so Round 4 was blocked.
  • B.3 Verifier-selected learning, iteration, and stopping: 12/80 to 44/80 to 62/80 one-response repair across Generator B added timing plus separation only in the second round, while other family trajectories also improved.The second round added a capability family not improved by the first round.
  • B.3 Verifier-selected learning, iteration, and stopping: 14/32 pass@32 for both base and Self-16 blocked training on an earlier null branch, despite first-attempt success rising from 3 to 10 and calls falling from 697 to 622.On shared unresolved rows, mean unique plans fell from 13.71 to 5.14 while duplicate rate rose from 57.1% to 83.9%.
  • B.3 Verifier-selected learning, iteration, and stopping: 128/128 Generator C Round-1 performance exceeded base’s 57/128, while Round 2 reached 126/128 and lost two rows relative to Round 1.The preregistered Round-2 versus base gain was 53.9 points, with a row-bootstrap interval of [45.3, 62.5].
  • B.3 Verifier-selected learning, iteration, and stopping: 77/80 fresh one-response repair and 26/32 fixed-budget reachability followed Phi training from 10/80 and 9/32, respectively, with paired net 17.This was fixed-benchmark cross-model replication, not a controlled model ranking.
  • B.4 Transfer and explanatory boundaries: 55/160 Generator B transfer cases passed with Round3, compared with 28/160 for base, 29/160 for Self-16, and 47/160 for Default30; Default30 reduced precedence plus control from 10/40 to 6/40.Default30’s paired net versus base was 19, but it crossed the family-regression guard.
  • B.4 Transfer and explanatory boundaries: Native B training resolved failures that transfer did not, including timing plus separation; selective transfer and saturated evaluations therefore rule out unrestricted improvement or general enterprise validity.Round3’s net eight over Default30 was not significant (p = 0.1153), and no scalar separator emerged in a post-hoc repair-graph audit.
Loading 2608.18324v1…