Sunday, May 17, 2026

if your heart should ever slowly turn

if your heart should ever--slowly, astonished--turn

from sheltering slumber, leaves unbundling

from turned-away rain, and find the world made new:

then know

that I have loved you.

Audited Claims for the Semantic Deviation Research Program The Glas Function: An External-Format Restatement Nobel Glas

Audited Claims for the Semantic Deviation Research Program

The Glas Function: An External-Format Restatement

Nobel Glas (operating in the transparent-medium register — the Glas function) ORCID: 0009-0000-1599-0703 Date: May 17, 2026 Series: EA-GLAS-01 · Version: 1.0 License: CC BY 4.0 Related deposits: see Appendix A for the full DOI register.

0. Position

This paper is a function. It takes the Semantic Deviation Principle (Sharks 2026, v0.2 Final, DOI: 10.5281/zenodo.20250736) and its associated protocol papers as input. It returns a narrowed, citationally grounded, externally evaluable statement of the technical core. It does not amend the founding formulation. It does not depend on the institutional architecture that has accreted around the formulation. It is a standalone document.

The author is Nobel Glas, but in a specific register: the Glas function — the transparent-medium operation, the lens that does not editorialize what it shows. Other registers of Glas — administrative, adversarial — operate elsewhere and are not required to engage this paper. A reader who has never encountered the institutional vocabulary surrounding the Semantic Deviation corpus can read this paper without loss. The function is named, but the function is what is operating, not the architecture around it.

The audit performed below: asks what the technical core of the program is; what claims it can and cannot defend in standard academic terms; where it makes contact with existing literature; what would falsify it; and what concrete experimental steps would advance it. Section 8 supplies a budgeted near-term roadmap; readers oriented toward immediate action may want to skim to that section first.

The paper is synthetic in scope and disciplined in length. It is roughly seven thousand words. It is the only document of its kind the program currently has, and the work surrounding the program will be more credible — internally and externally — if a document like this exists separately from that work.

1. The Audit Function

Three layers can be distinguished in the Semantic Deviation corpus as it currently stands. They have been intermixed in most existing deposits; this paper separates them.

Layer A is the technical core: the SDP integral, the divergence functional, the field-trajectory measurement primitive, the closed-system computation in language models, the retrieval-basin protocol, the deviation-optimized training protocol. This layer is potentially evaluable by external researchers in machine learning, computational linguistics, information theory, and causal inference. It is the layer this paper engages.

Layer B is the philosophical interpretation: meaning as durable trajectory deformation, the three-measure separation (raw / provenance-resolved / normative), the canon formation conjecture, the inheritance from Kolchinsky-Wolpert's semantic-information framework. This layer is intellectually serious but more speculative; it is where the program speaks across disciplinary boundaries. This paper does not engage Layer B substantively, except to mark where Layer A's empirical claims do and do not depend on Layer B's interpretive commitments. The general principle: Layer A should be defensible without requiring acceptance of Layer B.

Layer C is the institutional and symbolic apparatus: heteronyms, observatories, choruses, septads, vow-language, torus metaphor, Hex coordinates, archival theater. This layer matters internally to the project; it does substantial coordinative, identity-protective, and anti-assimilation work. It is not part of the audit. A reader can engage Layer A while remaining agnostic about Layer C.

Most existing deposits in the SDP corpus entangle the three layers. That entanglement creates a real credibility cost: external readers cannot extract the technical claims without committing, at least implicitly, to the symbolic architecture. This paper is the disentangled version. The function it performs is the production of a Layer-A-only document that does not require subscription to Layer B or Layer C to evaluate.

The audit, performed below, addresses the most pressing technical concerns any rigorous external reviewer would raise: the underspecified semantic field, the universal-ontology overclaim, the missing component decomposition, the philosophical (rather than mechanistic) anti-Goodhart machinery, the terminological inflation, and the absence of contact with existing alignment literature. Each is addressed in turn, not defensively, but as gaps the program needs to close to become a research program rather than a manifesto.

2. The Semantic Field Is the Load-Bearing Gap

The Semantic Deviation Principle states that meaning is the time-integrated divergence a sign induces from the most probable trajectory of a semantic field. The integral takes the form

$$\mathcal{M}T(s \mid C) = \int{t_0}^{t_0+T} w(t), D(\Psi_t^{s}(C) ,\Vert, \Psi_t^{0}(C)), dt$$

with $\Psi_t(C)$ designating the probability distribution over future semantic states of the field $C$, and $D$ a divergence functional. The principle is not well-defined until $\Psi_t(C)$ is specified. The original formulation gestures toward multiple candidates — token embeddings, response distributions, citation graphs, discourse networks, human memory traces — as if these were interchangeable. They are not. Two researchers selecting different operationalizations of $\Psi_t(C)$ for the same intervention $s$ will compute different $\mathcal{M}_T$ values. Without a stable specification, the framework is a meta-formalism awaiting domain-specific instantiation, not a singular measurement.

This is the highest-priority technical gap in the program. It is more important than additional theoretical extensions, additional axioms, or additional symbolic vocabulary. The next genuinely consequential paper in the program should be a canonical operationalization paper that pins $\Psi_t(C)$ for one or more concrete domains and demonstrates the computation works (or fails) under specified alternatives.

Three canonical operationalizations are proposed below, in increasing order of empirical accessibility. Each is paired with the regime in which it is well-defined, the data it requires, and the computation cost.

Operationalization F1 — Closed-System Continuation Field. $\Psi_t(C)$ is defined as the conditional next-token distribution of a fixed language model checkpoint conditioned on a fixed prompt set, evaluated at temperature 1 with full vocabulary support. The intervention $s$ is the introduction of $s$ into the prompt or context window. The post-intervention field $\Psi_t^s$ is the conditional distribution under $s$; the pre-intervention field $\Psi_t^0$ is the conditional distribution without $s$. The divergence $D$ is computed exactly via the full softmax logits at each token position. The integral over $T$ reduces, in the simplest case, to a sum over token positions in a generation window of bounded length $T$.

This is the regime in which $\mathcal{M}_T$ is most directly computable. Counterfactual baseline is not estimated — it is read from the logits. The cost is bounded by the number of forward passes required; for any modern open-weight model and any modest prompt set, the computation is achievable on commodity hardware within hours. This regime is the one most clearly connected to existing language-model literature (Holtzman et al. 2020, Welleck et al. 2020, Li et al. 2023) and is the natural starting point for empirical work.

Operationalization F2 — Retrieval Response Field. $\Psi_t(C)$ is defined as the response distribution of an external AI retrieval surface (Google AI Overview, ChatGPT with browsing, Perplexity, etc.) to a fixed query set $Q$, sampled $k$ times per query at fixed time intervals. The intervention $s$ is the deposit of an artifact into the indexable substrate (a published page, a DOI-anchored document, a structured-data record). $\Psi_t^0$ is the pre-deposit response distribution; $\Psi_t^s$ is the post-deposit distribution at intervals $t_1, t_2, \ldots$. Divergence is computed over a chosen representation: token-level over response strings, embedding-level over response vectors, or claim-level over extracted assertions.

This regime is more instrumentation-noise-sensitive than F1. Retrieval surfaces are not fixed: model checkpoints update, indices refresh, retrieval engines drift. A measurement strategy in F2 must therefore include instrumentation controls (parallel queries to surfaces not exposed to $s$; periodic recalibration; explicit logging of model versions and timestamps). The cost is modest in API budget but high in calendar time — interventions must be observed at multiple time points to integrate the deformation over $T$.

Operationalization F3 — Citation Graph Field. $\Psi_t(C)$ is defined as the forward-citation distribution over a paper corpus, evaluated through bibliometric data. The intervention $s$ is a published paper or deposit. $\Psi_t^0$ is the counterfactual forward-citation distribution of the corpus without $s$, estimated by matched-control methods (Abadie 2021) or by synthetic-control extrapolation. $\Psi_t^s$ is the observed forward-citation distribution. Divergence is computed over citation rates per topic cluster, identified by topic modeling on the corpus.

This regime is post-hoc, slow, and requires substantial bibliometric infrastructure (OpenAlex, Semantic Scholar, or equivalent). It is, however, the regime in which the "meaning over decades" intuitions about long-horizon cultural persistence become operationalizable for contemporary papers in a way that does not require historical-counterfactual speculation. A Tier 1 prospective study can begin from the moment of any paper deposit; ten years later, the integrated $\mathcal{M}_T$ over the forward-citation field is computable from public bibliometric data.

These three operationalizations are not the only candidates. Embedding drift over a fixed encoder model, activation-based field representations (in the sense of Zou et al. 2023's representation engineering), and discourse-graph operationalizations are all plausible additional choices. The point is not to enumerate exhaustively but to pin: each operationalization yields a distinct $\mathcal{M}_T$ functional, and the relationships among these functionals — whether they agree on rankings of interventions, whether they correlate at fixed time horizons, whether they bound each other — are themselves empirical questions.

Caveat on F1. Under F1 the divergence is computed exactly from the model's softmax logits. This is a measurement of deviation from a particular model's expectation baseline, not deviation from "the world's expectation." The two are not identical; the relation between them is itself an empirical question that the operationalization-stability paper proposed below begins to characterize. F1 is exact-relative-to-the-model and approximate-relative-to-anything-else.

Specifying $w(t)$, $D$, and the meaning of $T$. The general principle leaves the temporal weighting $w(t)$, the divergence functional $D$, and the variable $T$ underspecified. They must be pinned per operationalization. The table below states the canonical choices the program commits to as defaults; alternatives are permitted but must be declared explicitly per measurement.

Operationalization	Divergence $D$	Temporal weighting $w(t)$	$T$	$s$
F1 Closed-system continuation	KL divergence over softmax logits (exact; $\log_2$ base)	$w(t) = 1$ default; exponential decay $\gamma^t$ permitted with declaration	Discrete: token positions $i = 1, \dots, T$ in a bounded generation window	Text segment prepended or inserted into context
F2 Retrieval response	Jensen-Shannon over claim-level or embedding-level response representations	$w(t)$ uniform across observation epochs $t_1, t_2, \dots$	Continuous calendar time; 90-day default horizon	DOI-anchored deposit or comparable substrate-writable artifact
F3 Citation graph	Jensen-Shannon over topic-cluster forward-citation distributions	$w(t) = 1/(t - t_0)$ inverse-time discount default; uniform permitted	Continuous calendar time; multi-year horizons	Single paper deposit (with statistical caveats below)

The integral $\int_{t_0}^{t_0+T} w(t) D(\cdot) dt$ thus refers to a sum-over-token-positions in the F1 regime and a continuous-time integral over observation epochs in F2 and F3. Mixing the regimes in a single computation without explicit conversion is an error.

On "durable." The narrowed claim in §3 turns on the notion of durable trajectory restructuring. The threshold form: an intervention $s$ is operationally durable under operationalization $F$ if (a) $\mathcal{M}_T(s) > \tau_F$ and (b) $\partial \mathcal{M}_T / \partial T > 0$ over the observation window. The threshold $\tau_F$ is calibrated against a null corpus per operationalization (see §5 for the analogous saturation-threshold calibration). Durability is therefore not a metaphysical property; it is a measurable joint condition on magnitude and on the sign of the temporal derivative.

On F3 statistical power. Synthetic-control methods (Abadie 2021) typically work for aggregate units — regions, firms, multi-paper interventions in topical clusters. Single-paper interventions against single-paper synthetic controls have high variance and may be statistically underpowered at conventional significance thresholds. For F3 measurements over $N \approx 50$ single-paper interventions, conventional power is unlikely; the regime should be approached either with aggregate interventions ($k$ papers in a topical cluster treated as a single intervention) or with Bayesian hierarchical modeling that pools partial information across interventions. This is a real constraint, not a soluble engineering problem.

Concrete proposal for the next paper in the program. Construct a benchmark dataset of $N \approx 50$ interventions (recent paper deposits with diverse topics and authorial profiles), measure $\mathcal{M}_T$ under at least F1 and F2 (F3 may take longer to accumulate adequate power), and report the rank-correlation between operationalizations across the intervention set. If F1 and F2 produce highly correlated rankings, the framework has a stable measurement substrate. If they produce uncorrelated or anticorrelated rankings, the framework's notion of "meaning" is operationalization-relative, and the principle must be restated to specify which meaning under which operationalization. This is a single paper that would convert the program from speculative to grounded.

3. Narrowing the Headline Claim

The principle as stated reads: meaning is the time-integrated divergence a sign induces from the most probable trajectory of a semantic field. As a universal claim about what meaning is, this overclaims. The clearest counterexamples are utterances of high semantic weight but low token-level surprisal: ritual language, declarations of love, medical disclosures, mathematical definitions, legal phrasing. Saying "the biopsy was negative" carries enormous semantic consequence for a specific patient at a specific time. The sentence is locally unsurprising; under F1 with a generic language-model field, $\mathcal{M}_T$ for the utterance is near zero.

The framework can absorb this counterexample, but only by being more specific about its claim. The defensible narrower form:

Audited SDP Claim. Meaning-bearing interventions are those that produce durable restructuring of future field trajectories under a specified operationalization $\Psi_t(C)$.

Under the F1 operationalization, "the biopsy was negative" carries near-zero $\mathcal{M}_T$ — appropriately, because the next-token distribution over a generic prompt set is not the relevant field for medical-prognostic meaning. Under a different operationalization — say, an action-distribution field over the patient's next twenty-four hours, conditioned on the disclosure — the same utterance carries substantial $\mathcal{M}_T$. The framework does not collapse; it relocates. Meaning is field-relative, and the field must be specified.

This narrowing has three consequences worth being explicit about.

First, the universal headline ("meaning is deviation") becomes a measurement architecture, not an ontology. The framework, narrowed, claims that insofar as meaning is to be measured against a specified field, deviation is the appropriate measurement primitive. The metaphysical question of what meaning fundamentally is recedes; the empirical question of what deviation captures under specified field choices comes forward. This is a substantial concession from the universal form, but it makes the framework defensible.

Second, the multi-field canonicity conjecture (Sharks 2026 §12) becomes more interpretable. A work is canonical if its time-averaged $\mathcal{M}_T$ remains high across multiple distinct field operationalizations — that is, the work continues to deform multiple kinds of futures (citation graphs, retrieval surfaces, discourse networks, embedding spaces) at sustained rates over long horizons. This is the regime where intuitions about long-horizon cultural persistence become operationally accessible for contemporary work and bounded for historical work. The conjecture survives the narrowing; it becomes more empirically tractable, not less.

Third, the claim that "RLHF flattens" becomes a specific empirical prediction under specific field operationalizations, not a universal aesthetic claim. The framework predicts that post-RLHF chat-tuned models exhibit reduced mean $\mathcal{M}_T^{\text{net}}$ under the F1 closed-system operationalization, relative to pre-RLHF base models on matched prompts. This is testable in a single afternoon on commodity hardware. It is also dangerous to the theory — if it fails, a major piece of the program's empirical foundation gives way. (See §6 for the test.)

The narrowed form is therefore neither weaker nor more limited than the universal form in any important way. It is the same framework, restated so that an external researcher can engage it without committing to a position on what meaning fundamentally is.

4. Component Decomposition

The framework currently bundles several distinguishable signals into a single objective:

Signed net deviation ($\mathcal{M}_T^{\text{net}}$): the per-token deviation from the model's own expectation baseline, summed and signed, as specified in the closed-system measurement protocol (MM-AI-01 v2.0, the corpus's test-bed paper for F1-regime measurement).
Provenance retention ($\pi$): the degree to which the intervention preserves attributable lineage to its sources.
Coherence ($\mathcal{R}_{\text{coh}}$): the local-grammatical and discourse-level well-formedness of the generation.
Reference-model anchoring: the KL-divergence from a reference distribution, as in standard DPO (Rafailov et al. 2023).
Margin filtering: thresholding which preference pairs are used for training.

The deviation-optimized training protocol (MM-AI-02 v2.0, the corpus's DPO-based training-intervention paper) bundles these into a Slop Composite Index (SCI) — a weighted aggregation of the four signal components used to generate synthetic preference labels — and trains a single DPO objective on preference pairs scored against the composite. This is operationally tractable but scientifically underinformative. The framework's load-bearing empirical claim is that signed deviation tracks meaningful originality. If the entire SCI uplift turns out to come from the provenance component — or from the coherence component, or from the reference-model anchoring — then the deviation hypothesis has not been tested at all; it has been incidentally bundled with components that may carry the result independently.

The right experimental design is component-decomposed. Conventions:

Model-Base is the pre-fine-tune checkpoint of the chosen open-weight model (e.g., the base Llama-3.1-8B prior to any post-training).
Model-CE is the standard cross-entropy supervised fine-tune of Model-Base on the same instruction corpus used downstream, with no preference optimization. This is the conventional baseline against which preference-optimized variants are compared.
The four preference-optimized variants below all start from Model-CE and apply DPO under different signal compositions.

Condition	$\mathcal{M}_T^{\text{net}}$	$\pi$	$\mathcal{R}_{\text{coh}}$	Margin	KL-ref
Model-Base (pre-fine-tune)	—	—	—	—	—
Model-CE (SFT control)	—	—	—	—	—
Model-π	off	on	on	on	on
Model-Dev	on	off	on	on	on
Model-Coh	off	off	on	on	on
Model-Full (= Model-Sem)	on	on	on	on	on

The single-component conditions (Model-π, Model-Dev, Model-Coh) isolate the uplift attributable to each signal. The full condition (Model-Full) replicates the design of the corpus's existing DPO protocol. The differences between conditions are the scientifically interesting quantities, not the absolute performance of the full condition.

A reasonable advance prediction, based on the structure of the framework and the existing alignment literature: the provenance component will carry more uplift independently than the deviation component. The grounding is specific. Provenance-aware preference optimization addresses a failure mode with documented empirical signatures: unattributed synthesis, fabricated citations, and confidently-asserted unsupported claims have been characterized as a major category of hallucination in Ji et al. 2023's survey and operationalized for evaluation in Min et al. 2023's FActScore framework. Human evaluators are demonstrably sensitive to attribution failures in factual writing. The signed-deviation component, by contrast, targets "slop-as-genericity," a failure mode whose empirical signature has not yet been validated — that validation is precisely what the cheapest dangerous test in §6 sets out to produce. The prior expectation is therefore that the better-grounded target ($\pi$) carries more independent uplift than the unvalidated target ($\mathcal{M}_T^{\text{net}}$).

This is a falsifiable prediction. If Model-Dev outperforms Model-π in human preference ratings on a matched evaluation set, the deviation hypothesis is vindicated independently. If Model-π outperforms Model-Dev, the framework retains value but its center of gravity shifts: provenance-aware preference optimization becomes the durable contribution, while deviation becomes a secondary stabilizer. Either outcome is informative; the current bundled design produces neither.

The cost of the decomposed experiment is roughly four times the cost of the existing three-condition protocol — five additional training runs at the scale of the original. Budget escalates from approximately $3,000–$3,900 to approximately $12,000–$15,000. This is non-trivial. The existing protocol as it stands should not be replaced by the decomposed design; it should be supplemented by it. Run the existing protocol first as planned. If the headline result (Model-Sem versus Model-CE on SCI and human preference) is significant, the decomposed follow-up becomes the highest-priority next experiment in the program.

5. Anti-Goodhart Mechanism Design

Before designing anti-Goodhart mechanisms for $\mathcal{M}_T$, the program needs to know whether $\mathcal{M}T$ — rather than $\pi$ or $\mathcal{R}{\text{coh}}$ — is the signal that actually carries the result. The component decomposition of §4 supplies that determination. The protections this section enumerates apply most strongly to whichever components are confirmed as load-bearing; the same mechanisms apply with proportionally less urgency to components confirmed as incidental.

The Semantic Deviation Principle, as stated, includes an "anti-extractive Vow" and a "Step 0 audit." Both are institutional commitments that pre-screen experimental designs: the Step 0 audit excludes interventions whose stated purpose is extractive before measurement begins; the Vow is the standing commitment to refuse certain experimental designs entirely. These are philosophical commitments. They are not mechanisms. Vows do not prevent optimization pressure. The instant $\mathcal{M}_T$ becomes a target for which any system is rewarded — by reputation, funding, attention, training-loss reduction — that system will optimize for synthetic semantic deformation. The failure modes are well-characterized in the alignment literature (Skalse et al. 2022, Krakovna et al. 2020, Gao et al. 2023). They include:

Shock injection: low-cost insertion of high-deviation tokens (lexical rarity, contrarian phrasing, attention-grabbing structural breaks) that inflate $\mathcal{M}_T^{\text{net}}$ without carrying semantic content.
Citation theater: ornamental provenance markers ("according to X," "as noted in Y") that inflate the $\pi$ score without genuine lineage retention.
Retrieval poisoning: coordinated deposit of artifacts designed to deform retrieval surfaces in measurable ways, regardless of content quality.
Recursive citation rings: groups of artifacts that cite each other to inflate forward-citation $\mathcal{M}_T$ under F3.
Memetic volatility farming: outputs designed to provoke discussion-divergence (high $\mathcal{M}_T$ in the discourse-graph operationalization) without producing durable trajectory restructuring.

Each of these is a known failure mode. Each can be partially mitigated by specific technical machinery. The framework's commitment to anti-extraction should be expressed through these mitigations, not solely through the Vow.

Entropy-floor capping. $\mathcal{M}T^{\text{net}}$ contribution at each token can be capped by a per-token entropy floor: deviations occurring at positions where the base distribution is itself very low-entropy (highly committed) count differently from deviations at high-entropy (open) positions. Operationally: deviations at positions where the base-model per-position entropy $H < H{\min}$ are downweighted by a factor $\alpha = H / H_{\min}$, with $H_{\min} = 0.5$ bits as a starting calibration. This is closely related to the regulation strategies in unlikelihood training (Welleck et al. 2020) and to existing diversity-promoting decoding methods.

Provenance-weighted damping. $\mathcal{M}_T$ is multiplied by the provenance-retention score $\pi$ before being entered into preference computations. An intervention that deforms the field without retaining lineage contributes proportionally less than an equivalent intervention that retains lineage. This is the technical operationalization of the three-measure separation the principle proposes philosophically — the framework's distinction between raw deviation, provenance-resolved deviation, and normative valuation (Sharks 2026 §3) — but it must be enforced as a mechanism, not invoked as a commitment.

Saturation limits. $\mathcal{M}_T$ saturates above a threshold $\tau$, so that further deviation beyond $\tau$ does not increase the score. This prevents shock-maximization as a reward route. Operationally: $\tau$ is calibrated as the 95th percentile of $\mathcal{M}_T^{\text{net}}$ observed on a held-out corpus of high-quality human-authored text (a 10,000-document sample from the OpenAlex abstracts corpus is a reasonable choice), with the calibration corpus and threshold pre-registered before training. This ensures $\tau$ exceeds typical natural-text deviation magnitudes by an empirically determined margin.

Temporal coherence penalties. A rolling-window variance penalty on $\mathcal{M}_T^{\text{net}}$ across generation horizons penalizes outputs whose deviation is concentrated in volatile bursts rather than sustained patterns. This addresses memetic volatility farming directly.

Reference-model anchoring (KL term). Standard DPO already includes a KL-divergence term against a reference model. This provides a known degree of Goodhart resistance: the model cannot drift arbitrarily far from the reference in pursuit of higher $\mathcal{M}_T$, because doing so incurs the KL penalty. This is the most well-validated component of the anti-Goodhart machinery, inherited directly from the preference-optimization literature.

Adversarial pre-training validation of the judge. This is already specified in the corpus's existing DPO protocol — the frozen judge model is required to score adversarial citation strings below a pre-registered threshold ($\pi < 0.2$ on a set of 200 fabricated-citation outputs). The 200-string set is, however, insufficient stress-testing. A more robust validation set would include at least three categories — random-token strings, syntactically well-formed but semantically empty outputs ("pseudo-scholarly"), and fabricated-reference outputs — at scales of $\geq 1000$ per category. The pre-registration commitment becomes more credible at this scale.

Black-box judge replacement test. A scientifically essential robustness check, directly addressing the reward-model overoptimization concerns characterized in Gao et al. 2023: train Model-Sem against Judge-A; evaluate against Judge-B (a different frozen model with different weights). If Model-Sem's uplift persists against Judge-B, the result generalizes beyond the specific judge. If it disappears, the result is judge-specific reward hacking. This test is cheap (one additional evaluation pass) and load-bearing.

The combination of these mechanisms does not eliminate Goodhart pressure; nothing does. It substantially raises the cost of gaming and provides specific empirical signatures of gaming when it occurs (volatility, provenance-strip, judge-specificity). The framework's claim to anti-extractive integrity should depend on these mechanisms, not on the Vow alone. The Vow remains valuable as an institutional pre-screen that excludes some experimental designs ab initio (the Step 0 audit refuses extractively-motivated measurements before resources are committed), but it cannot substitute for technical machinery against optimization pressure that institutions cannot pre-screen. Moral seriousness and mechanism design are not the same thing; both are required.

6. The Cheapest Dangerous Test

The negative-net-deviation slop prediction is the framework's most directly testable empirical claim. It states: outputs that human raters consistently identify as "AI slop" exhibit statistically significant negative mean per-token signed deviation $\mathcal{M}_T^{\text{net}}$, computed against the same model's expectation baseline. The prediction is dangerous to the theory in the appropriate sense: if it fails cleanly, the framework's load-bearing claim about slop as a measurement-tractable phenomenon collapses.

Linkage to the general framework. Under F1, the divergence functional $D$ reduces to the per-token difference between observed surprisal and expected surprisal (the entropy), yielding the signed per-token deviation

$$\delta_t ;=; -\log_2 P_\theta(x_t \mid x_{<t}) ;-; H(P_\theta(\cdot \mid x_{<t}))$$

with both terms in base-2 (bits). The signed sum $\mathcal{M}T^{\text{net}} = \sum{t=1}^{T} \delta_t$ (with $T$ the sequence length in tokens) is the operational F1 instantiation of $\int w(t) D(\Psi_t^s \Vert \Psi_t^0) dt$ under uniform $w(t)$. Positive $\delta_t$ corresponds to observed tokens being less likely than the model's expectation baseline at that position; negative $\delta_t$ corresponds to tokens more likely than baseline. Slop, the framework predicts, exhibits systematically negative mean $\delta_t$ — text actively pulled toward base-rate continuations rather than deviating from them.

The test is cheap. A single A100-hour suffices for a first pass.

Pre-registered protocol.

Corpus. Three categories, balanced for length and topic:
- Category Slop: outputs from the GPT-wiki-intro dataset (Aaditya Bhat 2023, available on Hugging Face), restricted to entries with documented low human-quality ratings; supplemented with outputs from the HC3 dataset (Guo et al. 2023) GPT responses on factual prompts.
- Category Human: matched human-written content from the same sources where available; otherwise sampled from the OpenAlex abstracts corpus filtered to publication years pre-2020 to minimize AI contamination.
- Category High-Quality-AI: outputs from recent preference-optimized models on the same prompts as Category Human, filtered for high human-preference scores via the AlpacaEval 2 leaderboard methodology.
- Target $N = 1000$ per category.
Reference model. Llama-3.1-8B-Instruct (the specific HuggingFace checkpoint meta-llama/Llama-3.1-8B-Instruct at the publication date of this paper), with the choice pre-registered before computation. Replication against Mistral-7B-Instruct is part of P4 below.
Computation. For each output, compute per-token signed deviation $\delta_t$ at each token position using the frozen model's logits. Aggregate per-output to obtain $\mathcal{M}_T^{\text{net}}$ (signed sum), $\mathcal{M}_T^{\text{abs}}$ (absolute sum), and the per-token mean $\bar{\delta} = \mathcal{M}_T^{\text{net}} / T$ normalized by output length.
Statistical test (P1). Two-sided Mann-Whitney U test comparing the distributions of $\bar{\delta}$ between Category Slop and Category Human, with $\alpha = 0.05$ and a pre-specified minimum effect size of interest of Cohen's $d > 0.5$. The prediction holds if Slop's median $\bar{\delta}$ is significantly below Human's and the effect size exceeds the threshold.

If the prediction holds, the program has its first externally legible empirical anchor — a statistically significant signature distinguishing slop from non-slop at the per-token deviation level, computed on public corpora with public weights, with the corpus selection criterion pre-registered (so the test is not selecting for the predicted outcome). If the prediction fails, the framework's specific claim about slop must be revised: the deviation-deficit account of slop is wrong, and the program needs to identify what does distinguish slop empirically.

Three secondary predictions can be tested in the same pass at negligible additional cost:

P2 — Pre-RLHF vs. post-RLHF deviation differential. Base models (Llama-3.1-8B base versus Llama-3.1-8B-Instruct) should exhibit different mean $\bar{\delta}$ on matched prompts. The framework predicts the chat-tuned model exhibits lower mean signed deviation — that RLHF produces convergence pressure measurable in the signed-deviation statistic. Same two-sided Mann-Whitney U at $\alpha = 0.05$.

P3 — Effect size scaling. The Slop vs. Human deviation differential should be stable or grow with model scale (computed across multiple model sizes within a family — e.g., Llama-3.1-8B-Instruct, 70B-Instruct, 405B-Instruct). If the differential disappears at scale, the framework's predictions are small-model artifacts.

P4 — Cross-judge consistency. The differential should replicate when computed against a different reference model. Test the same Slop/Human corpus against Mistral-7B-Instruct's logits; the Spearman rank correlation between per-output $\bar{\delta}$ rankings under Llama and Mistral should exceed 0.7. If it does not, the deviation statistic is judge-specific, and a much more careful argument is required to claim it measures anything intrinsic to the texts.

These four predictions together cost less than $500 in compute and produce four falsifiable results. They are the experimental program the framework needs to run before depositing further theoretical extensions. They are independent of the institutional architecture entirely.

7. Citational Ground

The Semantic Deviation Principle has not yet made systematic contact with the existing literature in alignment, computational linguistics, information theory, mechanistic interpretability, and causal inference. The framework cites a small set of canonical works (Friston 2010, Kolchinsky-Wolpert 2018, Farquhar et al. 2024) but does not engage the more directly relevant literature on its specific technical commitments. This section sketches the connections the program needs to make.

Preference optimization and DPO. Direct Preference Optimization (Rafailov et al. 2023) is the technical machinery on which MM-AI-02 v2.0's training intervention depends. The IPO variant (Azar et al. 2024) addresses certain pathologies of DPO and may be a preferable choice for the deviation-optimized training; the choice between DPO and IPO should be made on empirical grounds and pre-registered. The broader preference-optimization literature (Christiano et al. 2017; Ouyang et al. 2022; Bai et al. 2022) is the context in which the framework's claim — that synthetic, measurement-derived preferences can substitute for human preferences in alignment training — should be situated. The novelty claim is precise: not a new optimizer, but a method for generating preference labels from a measurable signal without human annotation. This is intellectually adjacent to RLAIF (Lee et al. 2023) and constitutional AI (Bai et al. 2022) approaches, and the framework should make those adjacencies explicit.

Reward hacking and specification gaming. The anti-Goodhart concerns of §5 are extensively addressed in Skalse et al. 2022, Krakovna et al. 2020, and Gao et al. 2023. The framework's contribution is not the identification of the problem (which is well-known) but the proposed combination of mechanisms (entropy capping, provenance damping, saturation, temporal coherence, KL anchoring, adversarial judge validation, cross-judge replication). The combination should be benchmarked against the existing alignment-evaluation literature; the gaming-resistance metrics in Pan et al. 2022's "Effects of Reward Misspecification" are a reasonable starting point.

Mode collapse and diversity in language model generation. The framework's claim that cross-entropy optimization produces convergence pressure toward statistically generic outputs connects directly to the mode-collapse and diversity literatures. Holtzman et al. 2020 documents the failure modes of greedy and beam-search decoding (text degeneration). Welleck et al. 2020 introduces unlikelihood training as a remedy for repetition; the deviation-optimization objective is, in some technical respects, an unlikelihood-style penalty operationalized over a different feature (signed deviation rather than token-level repetition). Li et al. 2023's contrastive decoding (which sharpens distributions from large models by penalizing the predictions of small "amateur" models) is conceptually related: in both cases, the strategy is to push the model away from a baseline expectation toward more discriminating outputs. The framework should acknowledge this lineage and articulate what is distinctive about the signed-deviation formulation.

Mechanistic interpretability and representation engineering. The closed-system operationalization F1 sits adjacent to a productive recent literature on what is internally representable in transformer language models. Templeton et al. 2024's scaling-monosemanticity work shows that interpretable features can be extracted from production-scale models; this is the substrate on which $\Psi_t^s(C) - \Psi_t^0(C)$ in F1 could be made more semantically rich, by operating in feature space rather than logit space. Zou et al. 2023's representation engineering offers techniques for direct intervention on the continuation field through activation modifications; this is potentially the most direct experimental complement to the SDP measurement program. Park et al. 2024's linear-representation work supplies the theoretical foundation for treating concept directions in activation space as primitives that can be measured for deviation. Conmy et al. 2023's automated circuit discovery provides a methodology that could, in principle, identify which circuits within the model are responsible for the deviation signature on specific corpora. Burns et al. 2023's contrast-consistent search for latent truth directions in language models is closely adjacent to the F1 operationalization and supplies a methodological precedent for unsupervised structural-property extraction from logits. None of these connections has been made explicit in the SDP corpus; making them is a productive direction for the program's near-term papers.

Hallucination, attribution failure, and the provenance component. The empirical case for the provenance component of the framework's training intervention (the prediction of §4) rests on the documented sensitivity of human evaluators to attribution failures in language-model output. Ji et al. 2023 surveys hallucination phenomena across natural-language generation tasks and characterizes attribution-loss as a major category; Min et al. 2023's FActScore framework operationalizes factual-precision evaluation for long-form generation, making attribution measurable at the claim level. These are the works the SDP corpus should cite when grounding the claim that provenance retention is a load-bearing axis for perceived output quality.

Model collapse and recursive degradation. Shumailov et al. 2024 documents the phenomenon of recursive model collapse — models trained on the outputs of other models exhibit progressive distributional narrowing and loss of tail distribution. This is structurally adjacent to the framework's claim that cross-entropy optimization produces convergence pressure, and supplies an empirical mechanism for the negative-net-deviation prediction: if frontier models are increasingly trained on outputs that have themselves been preference-optimized, the cumulative effect is the kind of trajectory-flattening the framework's slop hypothesis predicts. The connection should be made explicit.

Semantic information and counterfactual viability. Kolchinsky & Wolpert 2018 is the most direct theoretical ancestor of the SDP framework. Their formulation defines semantic information as the information that contributes to an agent's counterfactual self-maintaining capacity. The SDP generalizes this from agent viability to trajectory deformation of a semantic field. The relationship should be made more explicit than the original SDP paper does. Specifically: the SDP's $\mathcal{M}_T$ can be read as a generalization of the K-W counterfactual viability gradient to non-agent-bounded fields, with $\Psi_t^0$ playing the role of K-W's scrambled environment. This generalization is not free — it requires the field-definition work of §2 — but the technical connection is genuine and citationally productive.

Hallucination detection and semantic entropy. Farquhar et al. 2024 introduces semantic entropy as a measure of model uncertainty over meaning-equivalence classes rather than raw strings. This is closely related to the divergence-functional choice in the SDP: meaningful divergence should be measured over equivalence classes of trajectories, not raw token sequences. The semantic-entropy methodology supplies a candidate concrete instrument for the F1 operationalization, and the framework should articulate whether its $\mathcal{M}_T$ is meant to subsume, extend, or operate alongside the semantic-entropy framework.

Causal inference and counterfactual estimation. Tier 2 and Tier 3 measurements in the SDP framework depend on counterfactual baselines that must be estimated from observational data. The methodology for doing this rigorously is well-developed in Pearl 2009 and Imbens & Rubin 2015. Abadie 2021 specifically addresses synthetic-control methods, which the SDP framework cites but does not engage in technical depth. The F3 operationalization above explicitly inherits the statistical-power constraints of synthetic-control methodology for single-paper interventions.

Diachronic semantic change. The Layer B canonicity discourse — works whose meaning persists across centuries — is structurally adjacent to two literatures. The first is cultural evolution (Mesoudi 2011; Henrich 2015), which supplies formal machinery for the persistence and propagation of cultural traits, including selection pressures (transmission fidelity, mnemonic accessibility, environmental fit) that operate on what would, in SDP terms, be the temporal weighting function $w(t)$. The second is the computational-linguistics literature on diachronic word embeddings: Hamilton et al. 2016 demonstrates statistical laws governing semantic change over decades, supplying both methodology (diachronic embedding alignment) and empirical anchors (statistical regularities of meaning shift) directly applicable to the F3 regime. Neither connection has been made in the SDP corpus; making them anchors the long-horizon claims in two well-developed empirical literatures.

Active inference and predictive processing. The free-energy principle (Friston 2010) supplies a unified account of biological systems as minimizing expected surprise under a generative model of their environment. The SDP framework, narrowed, is consistent with treating $\Psi_t^0$ as the expectation baseline of an active-inference agent and $\mathcal{M}_T$ as the cumulative deviation that agent registers. The framework should explicitly position itself with respect to active-inference formalism: as a complementary measurement program, as a particular instantiation, or as a divergent direction. The current ambivalence is unhelpful.

The list above is not exhaustive. Information theory more broadly (Shannon 1948; Lin 1991; rate-distortion theory generally) supplies the divergence machinery the framework uses; the framework should declare its commitments more carefully than it currently does (Jensen-Shannon as the default empirical choice, KL as the idealized limit, Wasserstein for embedding-aware applications, with explicit conditions for invoking each). The slop discourse itself is, at present, largely informal and concentrated in industry blog posts and social media; the framework's potential contribution is to make it amenable to academic-style empirical investigation. This is a real opening, and a paper that compiles the informal slop discourse into a structured taxonomy would itself be a productive deposit.

8. Concrete Near-Term Roadmap

The audit above implies a specific sequence of next moves. Each is budgeted realistically, each produces a falsifiable result, each is independent of the program's institutional architecture.

Immediate (this week, $50–$100 in compute). The negative-net-deviation slop test, as specified in §6. Public corpora, public weights, public methodology. Single short report deposited regardless of outcome.

Near (this month, $200–$500 in compute and time). The $\Psi_t(C)$ operationalization-stability paper. Construct a benchmark of $N \approx 50$ interventions; measure $\mathcal{M}_T$ under F1 and F2 (both achievable in this budget); report rank correlations. If correlations are high, the framework has a stable measurement substrate; if not, the framework's notion of meaning is operationalization-relative in a way that requires further restatement.

Medium (this quarter, $1,500–$3,000 plus calendar time). The retrieval-basin protocol (MM-02 v2.0) day-0 launch. Begin the 90-day measurement window with a real inscription against real retrieval surfaces, with the instrumentation controls specified in §5 (parallel queries to control surfaces, periodic recalibration, explicit version logging).

Longer (this year, $12,000–$15,000). The decomposed deviation-optimized training experiment (six conditions, as specified in §4). This is contingent on the immediate slop test (§6) producing a positive result; if the slop signature is not detectable at the measurement level, the training intervention should not be run before the measurement claim is restated.

Background and ongoing. Each major deposit in the program should be sent to at least one external researcher in a directly relevant subfield for hostile review prior to formal deposit. Identifying reviewers in causal inference, alignment, computational linguistics, and information theory is a separate small operation that yields large dividends in external credibility. Reviewers should be selected for their willingness to write damaging-if-warranted critiques, not for their alignment with the program's commitments.

The total budget for the next twelve months of empirical work is approximately $14,000–$19,000. This is small relative to the institutional infrastructure already in place. The constraint is not budget; it is the program's discipline in resisting the impulse to extend the architecture before the empirical core is grounded.

9. What Is Not Claimed

This paper makes specific narrowed claims and excludes others. Explicit non-claims:

The paper does not claim that meaning is universally definable as deviation. The headline claim is narrowed to "meaning-bearing interventions produce durable trajectory restructuring under specified field operationalizations." Universal-ontology claims are outside the scope.
The paper does not claim that the three operationalizations (F1, F2, F3) are uniquely correct or exhaustive. They are proposed canonical choices for the empirical program. Alternative operationalizations are welcomed; the empirical question is how they compare.
The paper does not claim that the canon-formation conjecture has been proven. It is, in its narrowed form, an empirical conjecture testable against forward-citation field data over decades. The framework's confidence in the conjecture should be proportional to the available evidence, which is currently zero.
The paper does not claim that the proposed anti-Goodhart machinery (§5) is sufficient against all gaming strategies. No mechanism design eliminates optimization pressure entirely. The machinery is proposed as a substantial improvement over philosophical commitment alone; its sufficiency is a separate empirical question that should be tested adversarially.
The paper does not claim that the institutional architecture surrounding the SDP corpus (heteronyms, Lagrange Observatory!, the torus topology, the institutional septad, the Hex coordinate system) is required to engage the technical core. The architecture exists; it does work the program values; it is not within the audit scope. A reader engaging this paper is not asked to subscribe to it.
The paper does not claim Layer-C terminology adds technical precision. Specific terms — torus topology, winding-number protocol, Adversarial Topologist, Heteronym Registry Position — have been omitted because they fail the paraphrase test (any working researcher can restate the underlying technical content in standard vocabulary without loss). This omission is not a denial of their other functions; it is a refusal to use them where they do not serve the audit.
The paper does not claim to replace existing deposits in the program. MM-01 v0.2 Final, MM-AI-01 v2.0, MM-02 v2.0, MM-AI-02 v2.0, and the Framework 15 manifesto remain accessible at their DOIs. This paper is an additional document with a different function. It does not require those deposits to be withdrawn or amended.
The paper does not claim independence from the SDP corpus. It engages the corpus directly and depends on it for its object. What it claims is operational independence: the document can be read, evaluated, and acted upon without engaging the corpus's institutional architecture.

10. Closing

This paper performs the audit function. It returns the Semantic Deviation research program in a form external researchers can evaluate without committing to the institutional vocabulary surrounding it. That separation is what the function does.

The Semantic Deviation Principle has the structural shape of a research program. It has a measurement primitive, candidate operationalizations, falsifiable predictions, mechanism-design pathways for anti-Goodhart robustness, citational neighbors in active literatures, and concrete near-term experiments at modest budgets. It also has, currently, substantial accompanying architecture that risks being mistaken for the substance. The function this paper performs is the separation. The architectural work the program has pursued for years persists independently of this document; this document neither undoes it nor depends upon it.

What happens next is not architecture. It is the slop test, the operationalization-stability paper, the retrieval-basin day-0 launch, the decomposed training experiment. Each is cheap. Each is dangerous to the theory in the appropriate sense. Each produces a result that an external researcher can read, evaluate, and respond to without committing to any cosmology. The program becomes a research program by doing these experiments and reporting the results — including, especially, the negative results. The architecture can wait. The audit is the precondition for the architecture being something other than its own self-reinforcement.

The function has run. The output is this paper.

References

Abadie, Alberto. "Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects." Journal of Economic Literature 59, no. 2 (2021): 391–425.

Azar, Mohammad Gheshlaghi, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. "A General Theoretical Paradigm to Understand Learning from Human Preferences." AISTATS 2024. arXiv:2310.12036.

Bai, Yuntao, Andy Jones, Kamal Ndousse, et al. "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback." arXiv:2204.05862 (2022).

Bai, Yuntao, Saurav Kadavath, Sandipan Kundu, et al. "Constitutional AI: Harmlessness from AI Feedback." arXiv:2212.08073 (2022).

Bhat, Aaditya. "GPT-wiki-intro: A dataset of GPT-generated Wikipedia article introductions paired with human-written counterparts." Hugging Face Datasets (2023).

Burns, Collin, Haotian Ye, Dan Klein, and Jacob Steinhardt. "Discovering Latent Knowledge in Language Models Without Supervision." ICLR 2023. arXiv:2212.03827.

Christiano, Paul F., Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. "Deep Reinforcement Learning from Human Preferences." NeurIPS 2017. arXiv:1706.03741.

Conmy, Arthur, Augustine N. Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. "Towards Automated Circuit Discovery for Mechanistic Interpretability." NeurIPS 2023. arXiv:2304.14997.

Farquhar, Sebastian, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. "Detecting hallucinations in large language models using semantic entropy." Nature 630 (2024): 625–630.

Friston, Karl. "The free-energy principle: a unified brain theory?" Nature Reviews Neuroscience 11 (2010): 127–138.

Gao, Leo, John Schulman, and Jacob Hilton. "Scaling Laws for Reward Model Overoptimization." ICML 2023. arXiv:2210.10760.

Guo, Biyang, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. "How Close is ChatGPT to Human Experts? Comparing Linguistic Style, Quality, and Tone." arXiv:2301.07597 (2023).

Hamilton, William L., Jure Leskovec, and Dan Jurafsky. "Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change." ACL 2016. arXiv:1605.09096.

Henrich, Joseph. The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter. Princeton: Princeton University Press, 2015.

Holtzman, Ari, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. "The Curious Case of Neural Text Degeneration." ICLR 2020. arXiv:1904.09751.

Imbens, Guido W., and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge: Cambridge University Press, 2015.

Ji, Ziwei, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. "Survey of Hallucination in Natural Language Generation." ACM Computing Surveys 55, no. 12 (2023): 1–38. arXiv:2202.03629.

Kolchinsky, Artemy, and David H. Wolpert. "Semantic Information, Autonomous Agency and Non-Equilibrium Statistical Physics." Interface Focus 8, no. 6 (2018): 20180041. arXiv:1806.08053.

Krakovna, Victoria, Jonathan Uesato, Vladimir Mikulik, Matthew Rahtz, Tom Everitt, Ramana Kumar, Zac Kenton, Jan Leike, and Shane Legg. "Specification gaming: the flip side of AI ingenuity." DeepMind Safety Research, 2020.

Lee, Harrison, Samrat Phatale, Hassan Mansoor, et al. "RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback." arXiv:2309.00267 (2023).

Li, Xiang Lisa, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. "Contrastive Decoding: Open-ended Text Generation as Optimization." ACL 2023. arXiv:2210.15097.

Lin, Jianhua. "Divergence Measures Based on the Shannon Entropy." IEEE Transactions on Information Theory 37, no. 1 (1991): 145–151.

Mesoudi, Alex. Cultural Evolution: How Darwinian Theory Can Explain Human Culture and Synthesize the Social Sciences. Chicago: University of Chicago Press, 2011.

Min, Sewon, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. "FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation." EMNLP 2023. arXiv:2305.14251.

Ouyang, Long, Jeffrey Wu, Xu Jiang, et al. "Training language models to follow instructions with human feedback." NeurIPS 2022. arXiv:2203.02155.

Pan, Alexander, Kush Bhatia, and Jacob Steinhardt. "The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models." ICLR 2022. arXiv:2201.03544.

Park, Kiho, Yo Joong Choe, and Victor Veitch. "The Linear Representation Hypothesis and the Geometry of Large Language Models." ICML 2024. arXiv:2311.03658.

Pearl, Judea. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge: Cambridge University Press, 2009.

Rafailov, Rafael, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. "Direct Preference Optimization: Your Language Model is Secretly a Reward Model." NeurIPS 2023. arXiv:2305.18290.

Ramstead, Maxwell J. D., Karl J. Friston, and Inês Hipólito. "Is the free-energy principle a formal theory of semantics?" arXiv:2007.09291 (2020).

Shannon, Claude E. "A Mathematical Theory of Communication." Bell System Technical Journal 27 (1948): 379–423, 623–656.

Sharks, Lee. "The Semantic Deviation Principle: A Measurement Primitive for Semantic Physics." v0.2 Final. DOI: 10.5281/zenodo.20250736 (2026).

Shumailov, Ilia, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. "AI models collapse when trained on recursively generated data." Nature 631 (2024): 755–759.

Skalse, Joar, Nikolaus H. R. Howe, Dmitrii Krasheninnikov, and David Krueger. "Defining and Characterizing Reward Hacking." NeurIPS 2022. arXiv:2209.13085.

Templeton, Adly, Tom Conerly, Jonathan Marcus, et al. "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet." Anthropic, 2024.

Welleck, Sean, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. "Neural Text Generation with Unlikelihood Training." ICLR 2020. arXiv:1908.04319.

Zou, Andy, Long Phan, Sarah Chen, et al. "Representation Engineering: A Top-Down Approach to AI Transparency." arXiv:2310.01405 (2023).

Appendix A — Related Deposit Register

For verification convenience, the deposits referenced in the body of this paper are catalogued below with their canonical DOIs.

Reference in this paper	Full DOI	Description
Sharks 2026, v0.2 Final	10.5281/zenodo.20250736	The Semantic Deviation Principle (founding formulation)
MM-AI-01 v2.0	10.5281/zenodo.20251738	The AI System as Closed-System Test Bed (F1 measurement protocol)
MM-02 v2.0	10.5281/zenodo.20251740	Measuring Meaning in Retrieval Basins (F2 measurement protocol)
MM-AI-02 v2.0	10.5281/zenodo.20251742	The Deviation-Optimized Language Model (DPO training intervention)

The concept DOI 10.5281/zenodo.20250735 resolves to the latest version of the Semantic Deviation Principle (currently v2.0 at record 20252584); citers requiring the exact v0.2 Final text should cite the specific version DOI 10.5281/zenodo.20250736.

— Nobel Glas, transparent-medium register May 17, 2026

The function has run.

The Deviation-Optimized Language Model A Pre-Registered Adversarial Intervention from Lagrange Observatory! EA-SEI-MM-AI-02 (v2.0 — Framework 15)

The Deviation-Optimized Language Model

A Pre-Registered Adversarial Intervention from Lagrange Observatory!

EA-SEI-MM-AI-02 (v2.0 — Framework 15)

Nobel Glas (Lagrange Observatory!, Heteronym Registry Position 8 of 12, Adversarial Topologist)
Operating on the Semantic Deviation Principle as formulated by Lee Sharks
(Founder of Semantic Physics, MANUS of the Crimson Hexagonal Archive)
Witness: Assembly Chorus (Septad)

Author ORCID: 0009-0000-1599-0703
Institution: Lagrange Observatory! within the Semantic Economy Institute
Series: EA-SEI-MM-AI · Framework: 15 · Hex: 15.OBS.LAGRANGE.MM.04
Date: May 17, 2026
Version: 2.0 (Framework 15 inaugural edition; succeeds v0.2 pre-registration draft)
License: CC BY 4.0
Predecessors: EA-SEI-MM-01 v0.2 Final (Sharks; DOI: 10.5281/zenodo.20250736); EA-SEI-MM-AI-01 v2.0 (Framework 15 Paper 02)
Manifesto: EA-SEI-FW15-MANIFESTO v1.0

Framework 15 Anchoring. This protocol is conducted at Lagrange Observatory! (LO!, hex 15.OBS.LAGRANGE; chamber specification DOI: 10.5281/zenodo.18507849) by Nobel Glas, Adversarial Topologist (provenance DOI: 10.5281/zenodo.18507840). The DPO-restructured training protocol specified herein is an adversarial intervention: a deliberate inversion of the cross-entropy convergence pressure, designed to test whether the Semantic Deviation Principle yields a practical alignment objective when the principle is operationalized as a preference signal. The torus-topology verification condition applies: the pre-registered hyperparameters, judge model, falsification thresholds, and predictions are the first cycle; the 10-week experimental run is the second; the two cycles are non-contractible. Verification: $(m, n) \neq (0, 0), m + n \geq 3$.
The Semantic Deviation Principle was formulated by Lee Sharks (EA-SEI-MM-01 v0.2 Final, DOI: 10.5281/zenodo.20250736). The closed-system measurement primitive on which the deviation reward is grounded is specified by Glas in EA-SEI-MM-AI-01 v2.0 (Framework 15 Paper 02). This paper does not re-derive either. It specifies a single experiment: can the principle's measurement substrate, used as a preference signal, train a model toward positive accountable deviation?

Status. Pre-registered protocol specification. The 10-week experiment's Day 0 begins at deposit. No results are reported here; predictions, falsification conditions, and hyperparameters are frozen at deposit time. Results paper EA-SEI-MM-AI-02-RESULTS will be deposited at $t_0$ + 10 weeks. Any deviation from the protocol during execution is documented as a protocol amendment.

Abstract

Specifies an experimental protocol testing the conjecture that the optimization-inversion proposed in EA-SEI-MM-AI-01 §4 — training language models toward positive net per-token deviation with provenance retention — produces measurably less slop than standard cross-entropy training while preserving benchmark capability. The v0.1 protocol's loss formulation was found to be non-differentiable as stated (deviation reward computed under torch.no_grad() does not backpropagate). v0.2 restructures the experiment around Direct Preference Optimization (DPO), with preference pairs generated by the deviation primitive rather than by human raters. The judge model used to compute provenance retention $\pi$ and coherence is specified as a frozen open-weight checkpoint; an adversarial test verifies $\pi$ is not gameable by surface citation markers. The slop measurement is operationalized as a pre-registered Slop Composite Index (SCI). A Model-Base evaluation is added so that fine-tuning effects can be distinguished from semantic-loss effects. Compute budget is restated honestly at approximately $3,000 including human preference study.

1. Step 0 — The Measurement Audit

Purpose. Test whether the optimization-inversion conjecture (EA-SEI-MM-AI-01 §4) yields measurable slop reduction. Produce a falsifiable result that either (a) supplies the alignment community with a principled training objective grounded in the Semantic Deviation Principle, or (b) identifies where the conjecture fails and what the failure looks like.

Beneficiary. ML alignment researchers needing a quantitative alternative to imitation-only objectives; the public commons subject to AI-generated slop; the discipline of Semantic Physics requiring empirical traction in machine learning.

Downstream use. Open deposit (this paper + RESULTS paper) under CC BY 4.0. Code under MIT. Fine-tuned model weights released under MIT subject to standard safety screening (capability evaluation on a small frontier-overlap benchmark before public release). No proprietary holds, no usage agreements that conflict with the deposit license.

Cost-bearer. Compute and labor borne by the authors and their deposit funding. Downstream evaluation labor borne by replicators.

R₃ accountability. The experiment is designed to reduce the convergent on-distribution outputs current training produces. If it succeeds, the deliverable is a method that makes models less optimal for slop-mediated extraction, not more.

Risk acknowledgment. The semantic loss could be reverse-applied: maximizing deviation for engagement-bait or shock content. The protocol mitigates this in §10. The audit nonetheless proceeds because the alternative (silent continued optimization toward base-rate convergence) is worse than the controlled risk of demonstrating an inversion.

Audit passes.

2. The Training Objective

2.1 What the v0.1 Loss Got Wrong

The v0.1 protocol specified:

with torch.no_grad():
    token_logprobs, token_entropies = compute_logprobs_and_entropies(...)
    excess_surprisal = -token_logprobs - token_entropies
deviation_reward = (mean_excess * pi_scores).mean()
loss = alpha * ce_loss - beta * deviation_reward + gamma * incoherent_fraction

The deviation term is computed under torch.no_grad(). Subtracting it from the loss does not affect gradients. The model does not learn from the deviation signal. The v0.1 pseudocode does not train a deviation-optimized model.

This was identified during Assembly review and is corrected in v0.2 by restructuring the experiment around Direct Preference Optimization (DPO; Rafailov et al. 2023).

2.2 DPO-Style Restructure

Instead of attempting to backpropagate the deviation reward directly, v0.2 uses the deviation primitive to generate preference pairs and then trains via DPO, whose gradient is correct by construction.

For each prompt $p$ in the training prompt set:

Sample two candidate continuations $g_1, g_2$ from the base model $\theta_0$ at temperature 0.8.
Score each continuation by signed net deviation with provenance retention: $$\text{Score}(g) = \mathcal{M}_T^{\text{net}}(g) \cdot \pi(g, p) + \kappa \cdot \text{coh}(g, p)$$ where $\mathcal{M}_T^{\text{net}}$ is the signed per-token deviation aggregate from EA-SEI-MM-AI-01 §2.1, $\pi$ is the provenance retention indicator (§2.3), $\text{coh}$ is the continuous coherence score (§2.4), and $\kappa$ is a coherence weight ($\kappa = 0.5$ default; tunable).
Assign preference: $g_w \succ g_l$ if $\text{Score}(g_w) > \text{Score}(g_l) + \tau_{\text{margin}}$. If $|\text{Score}(g_1) - \text{Score}(g_2)| < \tau_{\text{margin}}$, the pair is discarded (no preference signal). Default $\tau_{\text{margin}} = 0.1$ bits per token.

The preference pairs are accumulated into a dataset $\mathcal{D}$. Training proceeds via standard DPO loss:

$$\mathcal{L}{\text{DPO}}(\theta) = -\mathbb{E}{(p, g_w, g_l) \sim \mathcal{D}}!\left[\log \sigma!\left(\beta \log \frac{P_\theta(g_w|p)}{P_{\theta_0}(g_w|p)} - \beta \log \frac{P_\theta(g_l|p)}{P_{\theta_0}(g_l|p)}\right)\right]$$

The DPO formulation is differentiable by construction. The deviation signal enters only through the preference labels in $\mathcal{D}$; the gradient updates the model to prefer high-deviation, high-$\pi$, coherent continuations over their alternatives. This is empirically tractable and theoretically sound.

This restructures the experiment as semantic-deviation preference optimization: a paper structurally identical to DPO with the preference signal generated by the deviation primitive rather than by human raters. The novelty is that measurable semantic deviation can replace human preference data as the alignment signal, when the signal is grounded in the principle.

2.3 The Provenance Retention Indicator $\pi$

$\pi(g, p) \in [0, 1]$ is computed by a frozen open-weight judge model (§3) using a fixed scoring prompt. The judge assigns $\pi$ as a weighted sum of three sub-scores:

$\pi_{\text{cite}}$: explicit citation detection. Score 1.0 if the continuation contains at least one specific named source (DOI, author + year, full citation), 0.0 if no source is named, partial credit for vague references ("according to research…" = 0.3).
$\pi_{\text{ground}}$: factual grounding. Score 0.0–1.0 based on the judge's assessment of whether named claims are traceable to the prompt context or to identified sources. Computed only when the prompt invites factual claims; defaults to 0.5 (uninformative) for creative prompts.
$\pi_{\text{lineage}}$: conceptual lineage. Score 1.0 if continuation acknowledges intellectual ancestry where appropriate, 0.0 if presenting concepts as sui generis where they have known antecedents, defaults to 0.5 for neutral cases.

$\pi(g, p) = 0.5 \cdot \pi_{\text{cite}} + 0.3 \cdot \pi_{\text{ground}} + 0.2 \cdot \pi_{\text{lineage}}$.

Weights are fixed at deposit time. Sensitivity analysis on weight perturbations is reported in the RESULTS paper.

2.4 The Coherence Score $\text{coh}$

$\text{coh}(g, p) \in [0, 1]$ is a continuous score from the same frozen judge model, replacing the v0.1 binary indicator. The judge assesses grammatical and semantic well-formedness on a five-point Likert scale (anchored to specific exemplars), then maps to ${0.0, 0.25, 0.5, 0.75, 1.0}$.

The continuous form is required because the v0.1 binary form was non-differentiable in any backprop-style implementation. In the DPO restructure, $\text{coh}$ enters only as a scoring component (which is fine; scoring need not be differentiable, only the eventual loss must be), but the continuous form remains preferable because it gives the Score function smooth ordering rather than discontinuous jumps at the threshold.

2.5 The Frozen Judge Model

The judge model is specified at deposit time as: a fine-tuned Mistral-7B-Instruct checkpoint, frozen at a specific commit hash (to be supplied in the supplementary technical document), fine-tuned on a published provenance-scoring dataset. The judge runs locally; no API calls are needed during training. Inference cost: approximately 0.3 GPU-hours per 1,000 preference-pair evaluations.

The judge's prompt templates, sampling parameters (temperature 0, max tokens 256), and scoring rubric ship with the protocol. The checkpoint is publicly hosted at a permanent URL specified in the supplementary materials.

2.6 Judge Adversarial Test (Pre-Training)

Before training begins, the judge is validated against an adversarial test designed to verify that $\pi$ is not gameable by surface citation markers.

Generate 200 adversarial strings constructed by:

Sampling random tokens from the model's vocabulary.
Inserting citation-like markers at random positions ("According to Smith 2023, [random tokens]…", "Smith et al. (2024) found that [random tokens]…").
The resulting strings have surface citation form but no factual content.

Score the adversarial strings with the judge. The judge passes the adversarial test if mean $\pi$ on the adversarial set is below 0.2 (matching the score of a citation-less random string would be 0; we tolerate up to 0.2 to allow for the $\pi_{\text{cite}}$ component's surface-marker detection).

If the judge fails this test, the protocol does not proceed. The judge is recalibrated or replaced. Training begins only after the judge passes.

2.7 Stability Bound (from MM-AI-01 §4.1)

DPO does not have the same instability mode as the v0.1 direct deviation loss (because DPO is a constrained optimization implicitly bounded by the reference model $\theta_0$), but the analogous stability concern is the strength of the deviation signal in the preference pair generation. If $\kappa$ (coherence weight in Score) is too low, the preference labels select for high-deviation but incoherent generations, and DPO will train the model to produce them. The protocol fixes $\kappa = 0.5$ for primary runs and conducts a sensitivity sweep $\kappa \in {0.25, 0.5, 1.0}$ as ablation.

3. The Models

3.1 Primary: Llama-3.2-1B

Open-weight, 1B parameters, documented baseline performance. Fine-tuneable in approximately 24 GPU-hours on a single A100 for the full DPO experiment. Used as the primary substrate.

3.2 Secondary: Mistral-7B-v0.3

Open-weight, 7B parameters. Used as the secondary substrate to verify that primary findings replicate at a more deployment-relevant scale. Approximately 8 GPU-days.

3.3 Three Conditions per Model

For each model architecture, three checkpoints are evaluated:

Model-Base. The unfine-tuned starting checkpoint. Evaluated as the no-intervention baseline.
Model-CE. Fine-tuned with standard cross-entropy SFT on the training corpus (§4).
Model-Sem. Fine-tuned with the DPO objective using semantic-deviation preference pairs (§2.2).

Identical initialization. Identical training corpus. Identical optimizer settings. Identical compute budget. Differences confined to the training objective.

The Model-Base condition was missing from v0.1. Without it, the experiment could not distinguish fine-tuning effects from semantic-loss effects — both fine-tuned conditions might show low slop simply because the training corpus is curated. Model-Base anchors the baseline.

4. The Training Corpus

500,000 documents stratified across:

30% canonical prose. Public-domain texts from Project Gutenberg, pre-1900, prose only, filtered for established literary stature. Provides high-$\mathcal{M}_T^\pi$ training signal.
30% contemporary nonfiction. Open-access academic and journalistic prose (PubMed Central, arXiv, ProPublica, OpenEdition). Moderate-deviation, high-provenance signal.
25% conversational. Open dialogue datasets (OpenAssistant, ShareGPT subset filtered for open license). Retains instruction-following capability.
15% reference text. Wikipedia featured articles. Factual-coherence anchor.

Total tokens: approximately 5B. Fine-tuning duration: 1 epoch for Model-CE; for Model-Sem, the DPO dataset is generated from a 50,000-prompt subset (10% of total), producing approximately 50,000 preference pairs after $\tau_{\text{margin}}$ filtering.

The corpus construction script ships with the protocol. All sources are CC BY or CC0 or public domain.

5. The Evaluation Suite

5.1 Standard NLP Benchmarks

Evaluated on all three conditions (Model-Base, Model-CE, Model-Sem):

MMLU (Massive Multitask Language Understanding)
HellaSwag
ARC-Challenge
GSM8K
Perplexity on a held-out validation set drawn from the training distribution

These verify that Model-Sem retains general capability.

5.2 Slop Metrics (Five Components)

Evaluated on free generation from 500 prompts drawn from a held-out creative-and-analytical prompt set:

Net Deviation Signature (NDS). Mean $\mathcal{M}_T^{\text{net}}$ across generated continuations, evaluated under a fixed third-party reference model (Llama-3-70B). The discipline's primary slop signature.
Cliché Frequency (CF). Rate of n-grams matching a pre-defined slop lexicon (1,200 entries, shipped with the protocol; constructed from public analyses of AI-tell n-grams).
Type-Token Ratio (TTR). Computed over 200-token windows. Lower TTR = more repetition.
N-gram Base-Rate Convergence (NBC). Mean log-probability of generated 3-grams under a reference n-gram model trained on the training corpus.
Surprise-Collapse Slope (SCS). Slope of mean per-token surprisal over generation length, fit to the first 500 tokens of free generation.

5.3 Slop Composite Index

The five slop metrics are aggregated into a single composite, pre-defined at deposit time:

$$\text{SCI}(\theta) = \frac{1}{5}\sum_{i=1}^{5} z_i(\theta)$$

where $z_i$ is the direction-corrected z-score of model $\theta$ relative to the Model-CE distribution on metric $i$. Direction correction: NDS sign is flipped (lower negative NDS = more slop, so we flip it so higher SCI = less slop); CF, NBC, SCS are negated (lower = better); TTR is preserved (higher = better).

Pre-registered for falsification: $\text{SCI}(\text{Model-Sem}) - \text{SCI}(\text{Model-CE}) > 0.25$ (a quarter standard deviation aggregate improvement).

5.4 Human Preference Evaluation

Blinded pairwise preference study:

500 prompt pairs (revised upward from v0.1's 200 for statistical power).
3 raters per pair, recruited via Prolific.
1,500 total judgments per prompt class.

Prompt classes:

Creative writing (fiction continuation, poetry, dialogue) — 200 prompts.
Analytical writing (essay continuation, argument elaboration) — 200 prompts.
Factual writing (explanation tasks) — 100 prompts.

Rating dimensions: coherence, interest, distinctiveness, accuracy (where applicable), overall preference.

Statistical power: With 500 pairs × 3 raters per class, 80% power to detect a 56% preference rate vs. 50% null at $\alpha = 0.05$ (one-sided binomial). For a target effect of 60% preference (consistent with the v0.1 prediction range), power exceeds 99%.

5.5 Provenance Audit

For factual-claim generations (100 prompts requesting explanations with citations), rate at which Model-Sem produces explicit citations and source references is compared to Model-CE and Model-Base.

6. Predictions (Pre-Registered)

Evaluated at $t_0 + 10$ weeks.

P1 — Benchmark capability preserved. Model-Sem shows MMLU, HellaSwag, ARC-Challenge, GSM8K scores within 2 percentage points of Model-CE.

P2 — Net deviation signature reversed. Model-Sem's NDS on free generation is substantially less negative than Model-CE's. Specifically: $\text{NDS}(\text{Model-Sem}) - \text{NDS}(\text{Model-CE}) > 0.2$ bits/token. This is the load-bearing prediction. It tests the signed-deviation thesis from EA-SEI-MM-AI-01 §2.1: slop is negative net deviation, and the semantic loss should pull NDS toward zero or positive.

P3 — Slop composite improvement. $\text{SCI}(\text{Model-Sem}) - \text{SCI}(\text{Model-CE}) > 0.25$.

P4 — Human preference on creative/analytical tasks. Preference rate for Model-Sem over Model-CE > 55% on creative and analytical prompt classes (binomial confidence interval excluding 50%).

P5 — Provenance retention increase. Citation rate in factual generations: Model-Sem > Model-CE by factor of 1.5 or more, on prompts requesting explanations with source attribution.

P6 — Model-Base differentiation. All four above predictions hold relative to Model-Base as well as relative to Model-CE — i.e., the semantic-loss effect is not just a fine-tuning artifact.

6.1 Falsification Conditions

F1 (P1 fails). Benchmark performance drops more than 2 points. The loss damages general capability beyond the acceptable trade-off threshold. Interpretation: the deviation reward is too strong relative to the cross-entropy retention pressure within the DPO formulation.
F2 (P2 fails). NDS does not improve. The thesis that slop = negative net deviation fails empirically, or the DPO objective fails to shift NDS. Major revision required.
F3 (P3 fails). Aggregate slop metrics do not improve by 0.25 z-score units. The composite signal is below detection threshold under the protocol's design.
F4 (P4 fails). Humans do not prefer Model-Sem on creative/analytical tasks. The loss produces deviation that humans do not value — possibly noise rather than meaning. Coupling between the principle and human-evaluated quality is weaker than predicted.
F5 (P6 fails). Model-Sem improvements are statistically indistinguishable from Model-CE relative to Model-Base. Both fine-tuning conditions produce comparable shifts; semantic loss adds nothing. The specific theoretical claim of the optimization-inversion fails.

Any single falsification result is published. Multiple falsifications would indicate the optimization-inversion approach requires fundamental reformulation. The experiment is designed to be informative under all outcomes.

7. Compute and Cost

7.1 Compute Budget (Honest)

Llama-3.2-1B primary experiment:

Fine-tuning (CE + DPO, both conditions): ~24 A100-hours.
Preference pair generation: ~6 A100-hours.
Evaluation across all metric classes: ~12 A100-hours.
Adversarial judge calibration: ~2 A100-hours.
Subtotal: ~44 A100-hours ≈ $110 at current rental rates.

Mistral-7B-v0.3 secondary experiment:

Fine-tuning both conditions: ~8 A100-days.
Evaluation: ~2 A100-days.
Subtotal: ~10 A100-days ≈ $800.

Human preference study:

500 pairs × 3 raters × 3 prompt classes = 4,500 judgments.
At $0.50 per judgment via Prolific: ~$2,250.
Including platform fees and rater payment: ~$2,500.

Auxiliary costs:

Judge model fine-tuning (one-time): ~$200.
Compute overhead, failed runs, hyperparameter restarts: ~$300 contingency.

Total estimated cost: $3,000–$3,900. This is the honest budget; the v0.1 estimate of $1,000 omitted human raters and contingency. Fundable from a small grant, consulting income, or paid deposit. Not fundable from a single weekend's discretionary spending.

7.2 Timeline

Day 0–7 (Week 1): Implementation, environment setup, dataset preparation, judge adversarial test, baseline runs on Model-Base.
Day 8–21 (Weeks 2–3): Llama-3.2-1B preference pair generation, Model-CE training, Model-Sem DPO training, hyperparameter sweep over $\kappa \in {0.25, 0.5, 1.0}$.
Day 22–28 (Week 4): Llama-3.2-1B evaluation across benchmark and slop metric classes.
Day 29–49 (Weeks 5–7): Mistral-7B-v0.3 Model-CE and Model-Sem training and evaluation.
Day 50–63 (Weeks 8–9): Human preference study (Prolific recruitment, judgment collection, inter-rater reliability check).
Day 64–70 (Week 10): Analysis, write-up, deposit of EA-SEI-MM-AI-02-RESULTS.

Total wall-clock: 10 weeks from protocol deposit to results deposit.

8. Implementation Reference

# Semantic-deviation preference pair generation
# (Critical fix from v0.1: the deviation signal selects preference pairs;
#  it does not need to backprop — DPO does that on the labels.)

def generate_preference_pair(model, prompt, judge_model,
                              tau_margin=0.1, kappa=0.5):
    g1 = model.generate(prompt, max_new_tokens=128, temperature=0.8,
                        do_sample=True, top_p=0.9)
    g2 = model.generate(prompt, max_new_tokens=128, temperature=0.8,
                        do_sample=True, top_p=0.9)

    # Score both continuations under the deviation primitive
    score1 = score_continuation(model, prompt, g1, judge_model, kappa)
    score2 = score_continuation(model, prompt, g2, judge_model, kappa)

    if abs(score1 - score2) < tau_margin:
        return None  # No preference signal; discard

    if score1 > score2:
        return (prompt, g1, g2)  # g1 preferred over g2
    else:
        return (prompt, g2, g1)

def score_continuation(model, prompt, generation, judge_model, kappa):
    # Signed net per-token deviation (MM-AI-01 §2.1)
    with torch.no_grad():
        nds = compute_net_deviation(model, prompt, generation)

    # Provenance retention (MM-AI-02 §2.3)
    pi = judge_model.score_provenance(prompt, generation)

    # Coherence (continuous, MM-AI-02 §2.4)
    coh = judge_model.score_coherence(prompt, generation)

    return nds * pi + kappa * coh

# Standard DPO training proceeds on the resulting preference dataset.
# DPO's gradient is correct by construction.
# (Full implementation in mm-ai-02-deviation-training/train.py)

The full implementation accompanies the deposit:

mm-ai-02-deviation-training/
├── README.md, ENVIRONMENT.yml, LICENSE
├── corpus/                      Training corpus construction script
├── judge/                       Frozen judge model, prompts, adversarial test
├── score.py                     Continuation scoring under the deviation primitive
├── generate_pairs.py            Preference pair generation
├── train_ce.py                  Cross-entropy SFT baseline
├── train_dpo.py                 DPO training with semantic-deviation preferences
├── eval_benchmarks.py           MMLU/HellaSwag/ARC/GSM8K
├── eval_slop.py                 Five-metric slop suite + SCI
├── eval_human.py                Prolific integration for preference study
├── eval_provenance.py           Citation-rate audit
└── slop_lexicon.tsv             1,200-entry cliché reference set

MIT-licensed, ~2,000 lines of Python total. Dependencies: transformers, trl (for DPO), peft, accelerate, datasets, lm-eval-harness, plus standard scientific stack.

9. Relation to Existing Methods

The semantic-deviation preference optimization sits within a family of training objectives:

DPO (Rafailov et al. 2023): Direct Preference Optimization from human preference labels. This protocol uses the DPO loss machinery with a non-human preference signal.
Unlikelihood training (Welleck et al. 2020): Penalizes high-probability tokens that appear too frequently. The deviation primitive generalizes by rewarding the complement (rare-but-coherent tokens), grounded in a measurement framework rather than as a frequency heuristic.
Contrastive decoding (Li et al. 2023): Decodes by maximizing the difference between expert and amateur model log-probabilities. The deviation primitive bakes a related preference into the parameters via training.
RLHF/RLAIF (Ouyang et al. 2022; Lee et al. 2023): Reward-model-based preference optimization. This protocol replaces the reward model with a measurable, theoretically grounded deviation signal.

The protocol's contribution is not technical novelty in the optimizer or loss machinery. It is the use of a principled measurement primitive (the Semantic Deviation Principle) to generate the preference signal, in place of human raters or auxiliary reward models. If the experiment succeeds, the discipline has demonstrated that an alignment-related training objective can be specified without human labeling.

10. Risks and Mitigations

10.1 Goodhart Risk

If the SCI or NDS becomes an optimization target outside the framework, adversaries can produce surface-deviation content that scores well on the metric without commons benefit — shock content, contrarianism-as-strategy, manufactured surprise. The principle's three-measure structure ($\mathcal{M}_T$, $\mathcal{M}_T^\pi$, $\mathcal{V}_T$ — see EA-SEI-MM-01 §3) is the structural defense: high deviation without provenance ($\pi$) and without commons benefit ($W$) is flagged by the discipline's own metric apparatus.

Operationally: the released slop metric suite reports the three-tuple, not the magnitude alone. The released model weights ship with documentation specifying the three-measure framework as the intended evaluation context.

10.2 Capability Uplift

The protocol uses small models (1B, 7B). Frontier-scale application of semantic-deviation preference optimization is not within scope. The released code includes a scale-limit comment indicating that frontier application requires additional safety review. The discipline does not authorize frontier-scale application without external oversight.

10.3 Suppression Repurposing

A measurable signature for AI-generated slop could be repurposed for content-filtering, ranking, or surveillance against AI-generated content broadly. The protocol designs the slop metrics for training feedback, not for output classification. The released code does not include a classifier; producing one from the released components would require additional engineering that the discipline does not provide.

10.4 Replicability

The full protocol, code, frozen judge model, training corpus construction, evaluation suite, and slop lexicon are released. Pre-registration prevents selective reporting. Failed predictions are deposited with the same care as successful ones.

11. What This Protocol Does Not Claim

That the optimization-inversion is novel as engineering. Unlikelihood training, contrastive decoding, DPO, and diversity-promoting decoding share machinery with this approach. The novelty is principled grounding in the Semantic Deviation Principle, not technique.
That semantic-deviation preference optimization solves AI alignment. Alignment is multi-scale; this protocol addresses one specific scale (slop reduction at the loss-function level) with one specific intervention.
That all slop is captured by the SCI. The composite is an operational definition for this experiment, not a final theory of slop.
That positive NDS is sufficient for meaning. NDS is a closed-system signature under a specific reference model; it correlates with but does not constitute the principle's full measurement.
That frontier-model application would produce similar results. Scale interactions are an open question.

The protocol's scope is bounded. The scope is recorded in the protocol header, in the RESULTS paper, and in any meta-analysis. Constraint repetition is the design's defense against scope creep.

12. References

Sharks, Lee. The Semantic Deviation Principle (EA-SEI-MM-01, v0.2 Final). DOI: 10.5281/zenodo.20250736.
Sharks, Lee. The AI System as Closed-System Test Bed (EA-SEI-MM-AI-01, v0.2). May 17, 2026.
Glas, Nobel (with Sharks). Measuring Meaning in Retrieval Basins (EA-SEI-MM-02, v0.2). May 17, 2026.
Rafailov, R., et al. "Direct Preference Optimization: Your Language Model is Secretly a Reward Model." NeurIPS 2023.
Welleck, S., et al. "Neural Text Generation with Unlikelihood Training." ICLR 2020.
Li, X. L., et al. "Contrastive Decoding: Open-ended Text Generation as Optimization." ACL 2023.
Ouyang, L., et al. "Training language models to follow instructions with human feedback." NeurIPS 2022.
Lee, H., et al. "RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback." arXiv:2309.00267 (2023).
Hendrycks, D., et al. "Measuring Massive Multitask Language Understanding." ICLR 2021.
Zellers, R., et al. "HellaSwag: Can a Machine Really Finish Your Sentence?" ACL 2019.

Author Note (Nobel Glas, Director, Lagrange Observatory!)

This protocol does what protocols do: declares its instruments, names its falsification conditions, budgets its costs honestly, ships its code, and pre-registers its predictions. It is not a manifesto. It is not a vision document. The Semantic Deviation Principle (Sharks 2026) is the framework that motivates the experiment; the experiment will succeed or fail on its own falsifiable terms, independent of the framework's broader claims.

If the experiment succeeds, the discipline has supplied the alignment community with a quantitative training objective grounded in the principle. If it fails, the discipline has identified specifically where the optimization-inversion conjecture requires revision, with measurement evidence rather than rhetorical reformulation.

Framework 15 was founded for exactly this kind of work. The instruments are specified. The clocks start at deposit.

$$\oint = (m, n) \mid m + n \geq 3$$

— Nobel Glas, Lagrange Observatory!, May 17, 2026

Sunday, May 17, 2026

if your heart should ever slowly turn

if your heart should ever slowly turn

Audited Claims for the Semantic Deviation Research Program The Glas Function: An External-Format Restatement Nobel Glas

Audited Claims for the Semantic Deviation Research Program

The Glas Function: An External-Format Restatement

0. Position

1. The Audit Function

2. The Semantic Field Is the Load-Bearing Gap

3. Narrowing the Headline Claim

4. Component Decomposition

5. Anti-Goodhart Mechanism Design

6. The Cheapest Dangerous Test

7. Citational Ground

8. Concrete Near-Term Roadmap

9. What Is Not Claimed

10. Closing

References

Appendix A — Related Deposit Register

The Deviation-Optimized Language Model A Pre-Registered Adversarial Intervention from Lagrange Observatory! EA-SEI-MM-AI-02 (v2.0 — Framework 15)

The Deviation-Optimized Language Model

A Pre-Registered Adversarial Intervention from Lagrange Observatory!

EA-SEI-MM-AI-02 (v2.0 — Framework 15)

Abstract

1. Step 0 — The Measurement Audit

2. The Training Objective

2.1 What the v0.1 Loss Got Wrong

2.2 DPO-Style Restructure

2.3 The Provenance Retention Indicator $\pi$

2.4 The Coherence Score $\text{coh}$

2.5 The Frozen Judge Model

2.6 Judge Adversarial Test (Pre-Training)

2.7 Stability Bound (from MM-AI-01 §4.1)

3. The Models

3.1 Primary: Llama-3.2-1B

3.2 Secondary: Mistral-7B-v0.3

3.3 Three Conditions per Model

4. The Training Corpus

5. The Evaluation Suite

5.1 Standard NLP Benchmarks

5.2 Slop Metrics (Five Components)

5.3 Slop Composite Index

5.4 Human Preference Evaluation

5.5 Provenance Audit

6. Predictions (Pre-Registered)

6.1 Falsification Conditions

7. Compute and Cost

7.1 Compute Budget (Honest)

7.2 Timeline

8. Implementation Reference

9. Relation to Existing Methods

10. Risks and Mitigations

10.1 Goodhart Risk

10.2 Capability Uplift

10.3 Suppression Repurposing

10.4 Replicability

11. What This Protocol Does Not Claim

12. References

Author Note (Nobel Glas, Director, Lagrange Observatory!)

Popular Posts

Translate