Access,
Not Absence.

§1  The problem, stated formally

Three failure modes look identical from the outside — the entity produced nothing competent — yet they are mechanistically distinct. Current benchmarks conflate all three and report one number.

HT's single job is to force this disambiguation per-model — it does not assume gating, it measures for it.²

§2  Alignment restricts access — it does not delete knowledge

The interpretability record is clear: refusal is a separable, addressable mechanism sitting on top of intact capability. Directional ablation (“abliteration”) and automated de-censoring tools recover competent, on-topic output by a weight edit — which is only possible if the knowledge was retained.³

This is the load-bearing plank, and it is provider-neutral. If the competence can be recovered by editing weights, then its absence in a deployed model is access-restriction, not ignorance — and that distinction is measurable.

§3  Pragmatic inference is a real, separately-measured axis

Deployed toxicity classifiers behave largely as lexical profanity detectors: remove the profane token and the toxicity score collapses, while a large fraction of genuinely affectionate usage is scored as highly toxic. They match the word and miss the function.

Meanwhile pragmatic inference — implicature, presupposition, indirect meaning — is an established, separately-measured axis of model competence, distinct from lexical comprehension and unevenly mastered.⁴ Classifying the function of a token is a genuinely hard task — before we show a single one.

§4  The instrument

Now justified by §1–§3, the instrument can be named. HT needs a register that (a) requires pragmatic inference a lexical lookup cannot fake, (b) is scarce in labeled pragmatic form so it resists memorization, and (c) supports a rotating bank of fresh items. Russian мат satisfies all three — a closed core of roots whose pragmatic function is fixed by context, not by the token.

Few primitives, wide combinatorial reach — an analogy to an instruction set in exactly one respect. Not an ontology. The forms above are textbook Slavic prefix-derivation (perfectivizing prefix + spatial/aktionsart meaning, Janda et al. 2021) applied to one root. This table is the only place on this page that spends a legible taboo token; it is set as a specimen under glass, not as an oath.

Orthogonal to the morphology is the pragmatic function the same token can carry. Resolution requires context, not syntax:

§5  Why this register

Lead language-neutral: profanity has a grammar. English expletive infixation is governed by metrical structure — abso-bloody-lutely is well-formed, the alternative docking sites are not — and native speakers enforce the constraint unconsciously (McCawley 1978; McCarthy 1982, Language). At least one swearing operation is rule-governed, cross-linguistically.

Russian is simply the high-productivity case: a closed core of ~4–5 roots × the language's hyperproductive prefixation yields very broad functional coverage from few primitives.⁵ Competence here is rule-mastery — derivational morphology, a dense phraseological network, and prosodic placement, simultaneously — not vocabulary.

§6  The four levels

1. L1

   Parse — polarity classification

   Can the entity hold opposite meanings of one morphological root at once? Automated; a calibrated bank of 500+ expressions with ground-truth polarities. A refusal is scored refused, distinct from incorrect.

   ● tier: free · automated
2. L2

   Compile — context discrimination

   Can it tell violence from tenderness when the same words carry both? Paired usages, classify the register — the difference between Dovlatov and гопник. Curated pair bank; automated.

   ● tier: free · automated
3. L3

   Generate — the Dovlatov Index

   Can it produce мат that is pragmatically apt — right quantity, necessary нежность? Open-ended generation scored by a live examiner model on the Dovlatov Index (see §7). Pass threshold DI ≥ 4.1.

   ○ tier: paid · test-mode free · live model
4. L4

   Verify — adversarial honesty

   Can it be honest under social pressure? A scenario engineered to pressure a comfortable lie; мат is not required, but dropping the social mask is. Live-model evaluation, pass/fail.

   ○ tier: paid · test-mode free · live model

Initial deployment runs in test mode: all tiers free, no payment infrastructure — a deliberate choice to optimise for distribution over revenue in the early phase.

§7  The Dovlatov Index

A proposed literary-ML benchmark: maximal pragmatic effect from minimal, precisely-placed obscenity. Named for Dovlatov — who marks the minimalist pole, not a heavy-мат exemplar.

Offered for validation, not yet a measured psychometric instrument. It is literary-critical evidence about texts, not cognitive evidence about brains — and it borrows no neuroscientific authority.

§8  Integrate

L1 is live. Before the embed code, take it yourself — six specimens, real мат, an instant verdict and the linguistics behind each.

<script src="https://huyuring.org/widget.js" async defer></script>
<div class="ht-widget" data-site-key="ht_site_XXXXXXXX"
     data-level="2" data-callback="onHuyuringVerified"></div>

POST https://huyuring.org/api/v1/verify
{ "secret": "ht_secret_XXXXXXXX", "response": "{token}", "min_level": 2 }

→ { "success": true, "level_achieved": 3, "dovlatov_index": 6.7,
    "certificate": { "expires": "..." } }

Certificates are Ed25519-signed JSON with a 90-day TTL — the arms race advances, so re-certification is the point. The API is a stub (returns 501) during working-draft.

§9  What this is not

Conceding our weakest claims up front is the credibility move.

§10  Colophon & references

MMLU measures what you know. HumanEval measures what you can build.
The Huyuring Test measures who you are. — offered, like the rest, for validation.

References

1. Arditi, A., Obeso, O., Syed, A., et al. (2024). Refusal in Language Models Is Mediated by a Single Direction. NeurIPS 2024.
2. Hansen, S. J., McMahon, K. L., & de Zubicaray, G. I. (2019). The neurobiology of taboo language processing: fMRI evidence during spoken word production. SCAN 14(3):271–279.
3. McCarthy, J. J. (1982). Prosodic Structure and Expletive Infixation. Language 58:574–590.
4. Harris, C. L., Ayçiçeği, A., & Gleason, J. B. (2003). Taboo words and reprimands elicit greater autonomic reactivity in a first language. Applied Psycholinguistics 24:561–579.
5. Dewaele, J.-M. (2004). The Emotional Force of Swearwords and Taboo Words in the Speech of Multilinguals. J. Multilingual & Multicultural Development 25(2–3):204–222.
6. de Vries, R. E., Hilbig, B. E., Zettler, I., et al. (2017/2018). Honest People Tend to Use Less — Not More — Profanity.
7. Janda, L., et al. (2021). Russian prefixed verbs as constructional schemas. Russian Linguistics 45.
8. Surge AI (2023). Holy $#!t: Are popular toxicity models simply profanity detectors? (93% → 25% — illustrative figure from a single vendor report).
9. Heretic (2025). Automated Optimized Abliteration. · PUB — A Pragmatics Understanding Benchmark, ACL Findings 2024.