1 Introduction↩︎

Financial institutions in Vietnam operate under data-residency and sector-specific rules that restrict where customer data may be processed [1], [2]. For a regulated tenant, sending account records or underwriting files to a foreign frontier-model API is not a procurement decision but a compliance violation. A model that runs inside the institution’s perimeter therefore becomes a deployment requirement, and the frontier model functions as a quality ceiling rather than a deployment alternative. This framing reverses the usual cost–quality trade-off. The question is not whether a smaller model is cheaper to run, but whether a locally deployable model can reach the grounding quality that regulated work demands.

Earlier work established that supplying a domain ontology to a frontier model improves its domain fidelity [3]–[5]. That work measured the effect on a frontier model. It left open the question that matters for sovereign deployment: whether the same ontology grounding can lift a smaller, locally deployable student to the frontier’s level on regulated-domain tasks. A distilled student carries thinner parametric coverage of a specialized domain than a frontier model does, so ontology grounding may matter more for the student—a larger relative gain from a lower base—without necessarily closing the absolute gap on the hardest tasks. The companion neurosymbolic study named this pattern the Inverse Parametric Knowledge Effect (Inverse PKE): grounding value rises as the model’s parametric coverage of a domain falls [3].

This paper studies that question through ontology-amplified distillation, a two-stage adaptation. A student model is first supervised-fine-tuned on frontier-teacher trajectories that carry an injected ontology slice, then tuned with ontology-grounded direct preference optimization [6], [7]. The preference signal pairs an ontology-grounded frontier answer, marked preferred, against an ungrounded base-model answer, marked dispreferred, so the student learns to reproduce ontology-grounded behavior rather than to imitate the frontier’s full parametric knowledge. The student is a Qwen3.6-27B model [8]; training and evaluation run locally in Apple’s MLX framework on a single M5 Max [9], and the trained student is fused to a four-bit checkpoint of roughly 14 GB for deployment-consistent inference.

On a held-out set of 40 Vietnamese financial-domain tasks, the distilled student grounds 36 of 40 tasks—a grounded rate of 0.90, or mean term-coverage \(r_{\mathrm{onto}} = 0.95\) on a metric floored at 0.50—equal to the GPT-5 frontier baseline, which also grounds 36 of 40. The \(n=40\) binary outcome cannot establish equivalence: the paired-difference 95% confidence interval spans \(\pm 4\) tasks. Training-side measurements are consistent with the intended preference shift: preference accuracy moves from 0 to 1.0 and the reward margin from 0 to 0.307 over a single epoch, on a five-pair validation split. The result is scoped as a reduced-power proof of mechanism, not a deployment claim, and not the pre-registered amplification—which predicted the student exceeding, not equalling, the frontier. One binding gate was run—ontology compliance—on a metric that records whether the answer surfaces at least one of the task’s ontology terms, a degenerate proxy for the registered four-component composite, not whether the answer is complete or correct. The preference data are synthetic, English-only, and drawn from non-target domains; the frontier ran at minimal reasoning effort; the student matches rather than exceeds the frontier; and the questions of cost, abstention safety, model scale, and the Vietnamese-versus-English grounding contrast are held for a full-power evaluation.

The study makes four contributions. First, it specifies ontology-amplified distillation as a training recipe for sovereign enterprise models, with a preference construction that separates ontology-grounded behavior from general frontier imitation. Second, it defines an ontology-compliance metric scored locally and identically for the student and the frontier over a shared ontology slice, which removes judge-model variance from the comparison. Third, it reports the proof honestly, with an explicit map of which claims the present evidence supports and which await full-power confirmation. Fourth, it integrates a contextuality-audit method for enterprise-agent routing: when the same regulated task changes under role frame, prompt order, or construct wording, the audit distinguishes direct influence and construct coupling from residual contextuality before a platform routes the task to solo response, prompt standardization, debate, synthesis, or human review.

2 positions the work against distillation, preference optimization, and ontology-grounded generation. 3 describes the student model, the ontology slicer, the distillation pipeline, the evaluation gate system, and the compliance metric. 4 reports the held-out result and the training-side mechanism. 5 reports the consolidated contextuality-audit method and its negative result. 6 discusses scope, threats to validity, and the full-power evaluation that follows.

Combined-submission note↩︎

This manuscript consolidates two related arXiv-ready FAOS research notes into one article. The first was the ontology-amplified-distillation proof of mechanism reported in the main empirical sections. The second was a contextuality-auditor method note for enterprise LLM-agent routing. They share the same applied setting—ontology-grounded enterprise agents under regulated deployment constraints—and are combined here as one model-building-plus-governance article rather than submitted as separate variations on the same theme.

Dual-use disclosure↩︎

This study reports results from a dual-use experimental pipeline that serves two outputs: an internal minimum-viable sovereign model for the FAOS platform, and the academic contribution reported here. The principal investigator is a co-founder of the platform vendor and holds a commercial interest in the deployable artifact. 6 states this researcher-as-practitioner position and the mitigations applied, which include a locked held-out set, deterministic scoring, and a compliance metric computed without a judge model.

The method also engages a stated counter-position. Microsoft AI’s MAI-Thinking-1 report argues that capabilities should be learned, not inherited, holding that intelligence acquired through distillation lacks the steerability and robustness of capability trained from scratch [10]. The objection applies, on its face, to ontology-amplified distillation, so it is worth stating plainly. Three points bound the tension. The objection targets dependence on a third-party teacher, not distillation as such; the same report uses self-distillation. The objective also differs: the MAI program builds a frontier model, whereas this work ships a compliant model that must run inside a data-residency perimeter, which makes distillation a consequence of the deployment constraint rather than a shortcut. The steerability concern is nonetheless real, and the abstention gate that would test it—whether the student declines out-of-distribution prompts at a governed rate—belongs to the full-power evaluation deferred here, not to this proof of mechanism.

2 Related Work↩︎

2.1 Distillation and efficient student models↩︎

Knowledge distillation trains a compact student to reproduce a larger teacher’s behavior [7]. The technique scaled from classification to language models with DistilBERT [11] and, for instruction following, to preference-based distillation such as Zephyr, which distills alignment from a teacher’s ranked outputs [12]. Industrial pipelines now publish distilled open students in the Qwen family [13], and low-rank adapters make student adaptation cheap on commodity hardware [14]. This body of work concentrates on general-purpose capability: the student should track the teacher across broad benchmarks. The present study asks a narrower question—whether distillation can transfer one specific behavior, adherence to a domain ontology, into a student that must run inside a regulated perimeter.

2.2 Preference optimization↩︎

Learning from preferences began with reinforcement learning from human feedback [15] and its alignment variants [16]. Direct preference optimization (DPO) removed the separate reward model, training the policy directly on pairs of preferred and dispreferred responses [6]. The preference target here is neither human-likeness nor general helpfulness but ontology adherence. The preferred response is an ontology-grounded frontier answer; the dispreferred response is the same base model answering without the ontology slice. The signal isolates the grounded behavior from the teacher’s broader parametric knowledge, which is the property a sovereign student needs to acquire.

2.3 Ontology-grounded generation↩︎

A growing line of work couples language models to structured knowledge. Surveys map the integration of large language models with knowledge graphs [17], and the neurosymbolic program frames the broader aim of combining learned and symbolic representation [18], [19]. Recent systems ground generation at inference through ontology-anchored retrieval [4] or align a model to an ontology through self-training [5], and language models have been used to build ontologies in turn [20]. The question reaches back to the symbol grounding problem [21]. The companion neurosymbolic study in this program measured ontology grounding on a frontier model and reported the Inverse PKE: grounding helped most where the model’s parametric coverage of a domain was thinnest [3]. That result motivates the present question. If grounding compensates for thin parametric coverage, a distilled student, which carries thinner coverage than the frontier, should gain at least as much from the same ontology—the mechanism this study sets out to test.

2.4 Sovereign deployment under data residency↩︎

Vietnamese financial regulation constrains where customer data may be processed. The 2025 Law on Artificial Intelligence introduces a tiered risk regime with sector grace periods [1], the banking sandbox decree structures AI-enabled financial services [2], and the anti-money-laundering and insurance rules impose due-diligence and solvency obligations that reach model-mediated decisions [22], [23]. Under these constraints a tenant cannot route customer data to a foreign frontier API, so the deployable model must run on hardware the institution controls. On-device frameworks make local inference of mid-sized models practical on commodity accelerators [9]. The governance frameworks that regulated buyers map against—the EU AI Act, the NIST AI Risk Management Framework, and ISO/IEC 42001—treat traceability and control as first-order requirements [24]–[26], which a tenant-owned model satisfies more directly than a remote service. The deployment constraint, not a cost preference, makes the sovereign student the object of study.

3 Method↩︎

3.1 Student model and deployment target↩︎

The student is Qwen3.6-27B, an open model in the Qwen family [8] built on the qwen3_5 architecture. Training and evaluation run locally in Apple’s MLX framework on a single M5 Max with 128 GB of unified memory [9]; no cloud compute is used. After training, the student is fused for local inference; its deployment target is a four-bit (nvfp4) checkpoint of roughly 14 GB, the form in which it would run inside a tenant perimeter. Because the Gate 8 grounding metric scores term coverage rather than generation quality, it is insensitive to serving precision, and the held-out result does not depend on the exact precision of the deployed checkpoint. The wider research program targets a 14B student and a curve across model sizes. This proof of mechanism reports the 27B instance that was trained and evaluated, and holds the smaller-student and cross-scale questions for full-power work (6).

3.2 Ontology slicing↩︎

The Foundation AgenticOS ontology represents an enterprise domain across four sections: the role performing the work, the domain concepts in scope, the regulations that bind the task, and the interaction pattern that moves work between actors. For a given task, the slicer renders a task-specific ontology slice from these sections under a fixed token budget of 800 tokens (slicer version 1.0.0). When a slice exceeds the budget, the slicer drops content in a fixed order—interaction first, then domain—and never drops the regulation or role sections, so the compliance-bearing content survives truncation. The same slice is injected during training, as context for the teacher, and during evaluation, as context for both the student and the frontier, which holds the grounding signal identical across the pipeline.

3.3 The distillation pipeline↩︎

Adaptation runs in two stages on top of a capture step.

Capture. The frontier teachers GPT-5 and GPT-5-mini answer a non-held-out training manifest with the ontology slice injected, producing 451 grounded trajectories at a metered cost of $2.98. The held-out evaluation tasks are excluded from capture.

Supervised fine-tuning. A low-rank adapter over four layers is trained on the captured trajectories, reducing training loss from 2.370 to 0.496. The adapter is fused into the base model and re-quantized to nvfp4, producing the supervised student that serves as both the policy and the frozen reference for the next stage.

Ontology-grounded DPO. Direct preference optimization runs with mlx-lm-lora 2.1.0 on 47 synthetic preference pairs (42 for training, 5 for validation). Each pair marks the ontology-grounded frontier answer as preferred and the base model’s no-slice answer as dispreferred; the ontology-compliance gap between the two ranges from 0.17 to 0.53 (mean 0.27), with no inverted pairs. Training uses a rank-8 adapter over four layers, \(\beta = 0.1\), a learning rate of \(5\times10^{-7}\), the sigmoid DPO loss, batch size 1, a 512-token sequence cap, and 42 iterations of about one epoch under seed 20260615, updating 0.014% of parameters. A composite reward with weights \(\alpha = 0.50\) on task, \(\beta = 0.35\) on ontology, and \(\gamma = 0.15\) on role governs preference scoring (ADR-SL-006). The run completes in roughly 32 minutes at a peak of 58.1 GB, fuses to the trained student (adapter 38c28df0), and costs nothing beyond local power.

The preference data are synthetic, and three composition facts bound their reach. They are English-only; their source domains are healthcare (23 pairs), insurance (13), and fintech (11), none of which is a Vietnamese evaluation vertical; and the 47 rows reduce to 15 distinct prompts under roughly threefold duplication, with the five validation rows drawn from two prompts, one of which also appears in training. The protocol’s design anticipated subject-matter-expert preference labels and a larger pool; this run uses 47 synthetic pairs—about nine percent of the planned floor—and zero expert labels, so it makes no expert-preference claim. The defensible reading is that the preference signal demonstrates the training mechanism on constructed pairs rather than a production-grade preference set, and that any transfer to Vietnamese financial grounding is cross-lingual and cross-domain (6).

3.4 Evaluation gate system↩︎

The program scores a trained student against eight gates and a canary leakage check. Two gates carry binding authority. Gate 7 tests confident-wrongness, the rate at which the student abstains on out-of-distribution prompts, and Gate 8 tests ontology compliance, whether the student grounds its answers in the task’s ontology relative to the frontier. Gate 8 is the thesis gate for ontology-amplified distillation: it passes when the student’s mean compliance is at least the frontier’s, with no tolerance margin. This study runs Gate 8 alone. Gate 7 and the capability gates—in-distribution quality, cost ratio, dialect robustness, polysemy, and multi-turn consistency—require judge sign-off or evaluation sets above their full-power size floors, and are deferred (6).

3.5 Held-out evaluation set↩︎

The evaluation set holds 40 Vietnamese-language tasks, ten in each of four financial verticals: banking, insurance, capital markets, and real-estate technology. Tasks span four question types—terminological fidelity, metric accuracy, regulatory compliance, and role consistency—together with a cross-cutting category. The set is version-locked; its SHA-256 digest is a8c1d43b, a re-lock of an earlier digest (bbc14058) after each task was annotated in place with its ontology terms. A pinned extractor derives each task’s terms from the same ontology slice the evaluator uses, so the ground-truth terms and the scored slice share one source. Two verticals, banking and insurance, were redrawn for this version after an earlier exposure, and the two unexposed verticals, capital markets and real-estate technology, were carried forward; the redraw preserves the task-level leakage boundary, but a residual span-level overlap risk—training artifacts derived from the same blueprint sections the holdout cites—is examined in 6.

3.6 The ontology-compliance metric↩︎

Both the student and the frontier are scored by the same local metric over the same injected slice, with no judge model in the loop. The frontier baseline is GPT-5 run at minimal reasoning effort: the slice-laden prompts otherwise exhaust the reasoning-token budget and truncate the answer, and term coverage is insensitive to reasoning depth, so this is a defensible baseline for this metric—but every “equal to the frontier” in this paper refers to that configuration.

The held-out tasks carry ontology terms but no metric ranges, so \(s_{\mathrm{metric}}\) defaults to 1 and \(r_{\mathrm{onto}}\) becomes two-valued: 1.0 when the answer surfaces at least one of the task’s ontology terms, and 0.5 otherwise. This executed metric is a degenerate proxy for the pre-registered four-component ontology-compliance composite—terminological fidelity, metric accuracy, regulatory compliance, and role consistency—of which three components, and the correct-usage qualifier of the first, are not scored here; the metric is also floored at 0.50 rather than 0. The Gate 8 statistic is the mean of \(r_{\mathrm{onto}}\) over the 40 tasks, computed identically for the student and the frontier; because of the floor it ranges over \([0.5, 1.0]\), so the grounded rate—the share of tasks scoring 1.0—is the more directly interpretable quantity. The metric records term presence, not answer completeness or correctness; a finer-grained quality measure is full-power work (6).

4 Results↩︎

4.1 Ontology compliance↩︎

The distilled student and the frontier reach the same grounded count on Gate 8. Over the 40 held-out tasks each grounds 36 (a grounded rate of 0.90; mean \(r_{\mathrm{onto}}= 0.950\) under the 0.50 floor), so the gate—which passes when the student’s mean is at least the frontier’s—passes (1). Scoring is deterministic at temperature 0, and a re-run reproduced the verdict exactly; the evaluation took about 993 seconds on the M5 Max.

This is an equal point estimate, not a demonstrated equivalence. With a binary per-task outcome at \(n=40\), the grounded rate of 36/40 carries a Wilson 95% confidence interval of \([0.77, 0.96]\), and the paired comparison is uninformative: the discordant cells are balanced (two tasks each), so an exact McNemar test returns \(p = 1.00\) and the 95% interval on the mean paired difference spans \(\pm 0.05\), or \(\pm 4\) tasks. The result establishes an equal count, not statistical equivalence, and is underpowered to rule out a real gap of up to roughly ten percentage points in either direction.

The equal totals rest on partly different tasks (2 and 1; full per-task list in 9). Two tasks defeat both models, and each grounds two the other misses; the student is perfect on banking and insurance and grounds 8 of 10 in each of capital markets and real-estate technology, while the frontier is stronger on capital markets (9 of 10) and weaker on real-estate technology (7 of 10). The small number of disjoint misses (two each) is consistent with independent grounding rather than task-for-task imitation of the teacher, though four disjoint misses at \(n=40\) cannot establish independence. The perfect banking and insurance scores fall on exactly the two verticals redrawn for this version, which also carry the largest ontology-term lists; a one-term threshold is easiest to clear there, so the clean sweep is weak evidence of grounding strength (6).

4.2 The training mechanism↩︎

The DPO objective separated the constructed preference pairs (2). Held-out preference accuracy moved from 0 to 1.0 on a five-pair validation split, the reward margin between the preferred and dispreferred response from 0 to 0.307, and the DPO loss fell from 0.693 to 0.551. A second proof-scale run on the same 47-pair construction reproduced the pattern (preference accuracy 0 to 1.0, reward margin 0 to 0.319); because it reuses the same data design it is a repeat, not an independent replication. The validation split is small enough—five pairs from two prompts, one shared with training—that the 1.0 should be read as separability of the constructed pairs, not as generalized grounding. With that caveat, the training-side measurement is consistent with the intended preference shift: ontology-grounded tuning moves the student toward surfacing the ontology terms the frontier surfaces when it carries the same slice.

5 Contextuality Audit for Enterprise-Agent Routing↩︎

Ontology-amplified distillation addresses one side of the sovereign-enterprise problem: how to move a locally deployable student toward frontier-level grounding on regulated tasks. A platform also needs a governance diagnostic for the cases where a model’s answer changes across role frames, prompt orders, ontology conditions, or escalation criteria. Some variation is ordinary direct influence: the prompt changed, so the marginal answer changed. The harder question is whether any residual inconsistency remains after that direct influence is accounted for, and whether the residual should trigger multi-agent debate, synthesis, or human review.

We treat this as a contextuality-audit problem rather than as a claim that LLMs are quantum systems. Contextuality-by-Default (CbD) indexes each random variable by both the content being measured and the context in which it is measured [27]–[29]. This is well suited to enterprise-agent outputs: the same business content can be measured under a business-owner frame, risk-reviewer frame, governance-reviewer frame, or neutral-arbitrator frame. The audit’s purpose is operational. It asks whether the observed variation should be interpreted as stable behavior, measurement design, construct coupling, direct influence, or residual contextuality. That interpretation then maps to the least intrusive orchestration remedy: solo response, prompt standardization, debate or synthesis, or human arbitration. This diagnostic complements multi-agent debate and orchestration work, which can improve selected outputs but does not by itself diagnose when routing is needed [30], [31].

The consolidated RA-15 pilot used local Qwen 3.6 27B as the first model. Phase 1 produced 576 valid outputs; Phase 1.1 added 336 valid q4 robustness outputs; Phase 1.2 added 480 role/order-disentanglement outputs; and Phase 1.3 added a 384-call construct-decoupling extension separating operational permission from procedural evidence readiness. The original instrument surfaced apparent “CNTX” signals, but a formalism audit corrected the disturbance normalization: the earlier nonzero values were probability-drift residuals, not the canonical cyclic binary CbD degree. Under expectation-scale direct-influence correction, the canonical CbD degree is zero for all Phase 1.3 groups in both the local-Qwen run and an explicitly labeled Gemma replication check. The negative result is the finding. In this pilot, apparent contextuality collapses into direct influence and construct coupling rather than surviving as residual contextuality.

The link to ontology-amplified distillation is governance rather than shared training data. The distillation result asks whether ontology grounding can move a sovereign student to frontier-equal term coverage on a narrow held-out set. The contextuality audit asks how an enterprise platform should treat remaining decision variation once such a model is deployed inside a governed workflow. Together, they define a combined mechanism-and-control claim: the model can be made more domain-grounded, and the platform should still separate prompt sensitivity, construct design, and residual conflict before escalating the task.

The contextuality component therefore adds a negative-results method claim to the proof-of-mechanism claim. It does not license a contextuality-positive routing rule. It licenses a safer precursor rule: before treating model disagreement as evidence that a task requires debate or human arbitration, run a direct-influence and construct-validity audit. This avoids over-escalating tasks on the basis of an under-normalized score, and it also avoids hiding genuine conflict behind a single context-dependent answer.

6 Discussion↩︎

6.1 What the result shows↩︎

Two measurements point the same way. The student reaches the frontier’s grounded count on the held-out set, and the training-side metrics are consistent with the DPO objective—rather than chance—producing the shift on the constructed pairs. For a sovereign-deployment setting the reading is specific: a locally deployable Qwen3.6-27B student, tuned on synthetic ontology-grounded preferences, grounds Vietnamese financial-domain answers as often as GPT-5 does on this term-coverage metric. The divergent miss pattern is suggestive—parity that survives a few independent errors is harder to explain as task-for-task imitation—but four disjoint misses at \(n=40\) make it a weak signal, not a demonstration.

6.2 What the result does not show↩︎

The claim is bounded on purpose. The metric records whether an answer surfaces at least one ontology term, not whether the answer is complete or correct, so the grounded rate of 0.90 (mean \(r_{\mathrm{onto}}= 0.95\) under the floor) measures grounding presence, not domain quality. The pre-registered primary hypothesis was amplification—the distilled student’s normalized ontology-lift exceeding the frontier’s (\(\nu_D > \nu_F\)), with a tie (\(\nu_D \le \nu_F\)) named as the disconfirmation condition. This run does not test that hypothesis: it scores a reduced term-coverage proxy rather than the registered four-component composite, runs on the 27B instance rather than the 14B target, and computes no C1-to-C3 lift. The equal count is therefore neither a confirmation of amplification nor, given the proxy, a clean disconfirmation; amplification remains untested at power, and the Inverse-PKE framing should be read as motivation, not as a result this run supports. The evaluation ran one binding gate. Confident-wrongness abstention, in-distribution quality, cost, dialect robustness, polysemy, and multi-turn consistency were not measured. The preference data are synthetic, English-only, drawn from non-target domains (healthcare, insurance, fintech), and reduce to 15 distinct prompts; no subject-matter-expert preference label enters the pipeline. The student is a single 27B model from one family, evaluated on 40 Vietnamese tasks across four verticals; the smaller-student target, the cross-scale curve, and the Vietnamese-versus-English grounding contrast that would replicate the Inverse PKE in the distilled regime are not part of this evidence. No production or procurement decision should rest on a pilot of this size.

6.3 Threats to validity↩︎

Construct validity. Term-coverage is a coarse proxy for grounding, and the executed metric scores only one of the four pre-registered ontology-compliance components. Two design choices make the bar low. The gold terms are extracted from the same slice injected to both models, so each side is handed the target vocabulary and then scored on echoing at least one token of it; and a response can surface a term without reasoning over it correctly, so the metric credits shallow mentions. Scoring the frontier by the identical metric bounds the comparison—both models face the same near-ceiling test—but the absolute 0.95 and the compression of any student-frontier gap toward zero follow partly from the metric’s construction, not only from grounding ability. A coverage-fraction or held-out-term variant, and the full four-component composite, are the first things a full-power evaluation restores.

Internal validity. The preferred responses are frontier answers, so the student may be learning to surface terms in the frontier’s style rather than acquiring independent grounding. Because the preference pairs are English-only and from non-target domains while the evaluation is Vietnamese financial, the behavior the student was tuned on (echoing the preferred answer’s terms) and the behavior the metric rewards may be the same surface pattern measured twice. The modest compliance gap in the pairs (0.17 to 0.53), the synthetic-only construction, and the absence of expert validation leave this open.

External validity. Four verticals, one base-model family, and 40 tasks do not support generalization beyond the tested frame, and the result is Vietnamese-only.

Leakage. The held-out set excludes the evaluation tasks at the task-identifier level and redraws the two previously exposed verticals. This controls task-level overlap but not span-level overlap: the supervised-fine-tuning corpus is generated from the same Vietnamese blueprint sections the holdout cites, so a training artifact can convey a held-out task’s gold terms while passing the task-id check—a risk the project’s own leakage specification flags as making the identifier-only assertion structurally vacuous for those corpora. A span-overlap assertion exists as a tested primitive, but its wiring into the corpus generators was incomplete at the time of the Wave-0.5 run, so span-level exclusion is not established for the executed supervised corpus. The synthetic DPO pairs, being English and blueprint-free, are exempt. Until span-level screening is confirmed end-to-end, part of the held-out grounding could reflect memorized blueprint vocabulary rather than transferred grounding.

Reproducibility. The final publication manifest hash-pins all 451 committed source trajectories, the corpus-building code and configuration, the DPO corpus, the locked holdout, both final-evaluation JSON files, governance reviews, and the arXiv package. The executed v1.2-screened SFT train/validation JSONL files and local SFT/DPO checkpoints were not archived before operator-side cleanup. Later ontology-blueprint changes also prevent a byte-identical SFT rebake from the current tree. The evidence package therefore supports audit of the reported inputs, procedure, and Gate 8 arithmetic, but not bitwise reproduction of the trained checkpoint. The study makes no checkpoint-level reproducibility claim.

Conclusion validity. The metric is binary per task, and the equal means rest on 36 of 40 grounded for each model; a single task shifts the mean by 0.0125. The paired-difference 95% confidence interval spans \(\pm 4\) tasks. The study reports parity on this metric and makes no claim of statistical superiority or equivalence.

6.4 Researcher-as-practitioner position↩︎

The principal investigator is a co-founder of the platform whose sovereign model this study evaluates, and the experimental pipeline produces both the internal artifact and this paper. The design applies the mitigations available to a single-pipeline study: a version-locked held-out set (its questions sealed before the run, its gold-term annotation added in place under a pinned extractor and independently re-reviewed afterward), deterministic scoring at temperature 0, a compliance metric computed locally without a judge model, and a pre-registered protocol. The dual role is disclosed rather than corrected away; the deferred full-power evaluation, with independent judge sign-off, is where the harder bias controls apply.

6.5 Toward full-power evaluation↩︎

The questions this proof of mechanism leaves open define the next study. The binding safety gate—whether the student abstains on out-of-distribution prompts at a governed rate—tests the steerability that distilled models are accused of failing to acquire [10], and it is the first item. Beyond it: the capability gates at full evaluation-set size, frontier-judge sign-off on answer quality, a measured cost ratio against the frontier API, the 14B student and the cross-scale curve, subject-matter-expert preference labels in place of synthetic pairs, and the Vietnamese-versus-English contrast that would test whether the Inverse PKE holds in the distilled regime. Each turns a deferred claim into a measured one.

7 Conclusion↩︎

Ontology-amplified distillation moved a sovereign, locally deployable Qwen3.6-27B student to the same held-out grounded count as GPT-5 on a coarse ontology term-coverage metric—36 of 40 tasks each (0.90; mean \(r_{\mathrm{onto}}= 0.95\) under the 0.50 floor)—with a training-side preference signal that separated the constructed pairs (0 to 1.0). The equal count is an underpowered point estimate, not demonstrated equivalence, and not the pre-registered amplification. The consolidated contextuality-audit result is likewise bounded: after correcting direct influence, the canonical CbD degree is zero in the reported pilot, so the useful finding is a negative-results method for separating direct influence and construct coupling from residual contextuality before routing tasks to debate, synthesis, or human review. As a combined article, the contribution is a mechanism-and-control pair. The mechanism is ontology-amplified distillation for sovereign enterprise models; the control is a contextuality audit that prevents over-reading prompt-sensitive variation. Both results license fuller evaluation rather than deployment claims. Whether the student’s grounding is complete and correct, whether it abstains safely, what it costs against a frontier API, and whether the effect holds at smaller scale and across languages—these define the evaluation that follows.

8 Training Hyperparameters↩︎

9 Per-Task Grounding↩︎

Of the 40 held-out tasks the student grounds 36. Its four misses are two terminological-fidelity tasks in capital markets and two real-estate-technology tasks: one metric-accuracy task and one terminological-fidelity task. The frontier also grounds 36 of 40. It misses one capital-markets metric-accuracy task and three real-estate-technology tasks: two metric-accuracy tasks and one terminological-fidelity task. Two real-estate-technology tasks defeat both models; the remaining misses are disjoint, so each model grounds two tasks the other does not. The equal totals therefore rest on different task sets, which is the basis for reading the parity as independent grounding rather than teacher imitation.

References↩︎