The Invisible Editorial Layer

What the paper is, and what this page can verify

The public record used here is the arXiv abstract page and the arXiv HTML full text, both retrieved 26 August 2026. The abstract page lists a single author, Augusto Camargo, the line "[Submitted on 25 Aug 2026]", subjects cs.AI, cs.CL, and cs.CY, and the DataCite DOI https://doi.org/10.48550/arXiv.2608.24662. The HTML prints the affiliation Bluecore Consulting, Sao Paulo, Brazil, the date August 25, 2026, and license CC BY 4.0. The DataCite CSL record retrieved the same day names the same author, year 2026, and the same three subject categories.

Camargo writes that the paper does not introduce inference-time controlled generation. It examines what happens if established steering methods are used for undisclosed external objectives. Probability Placement is defined in the paper as a hypothetical advertising primitive, not as a documented product of a named lab. This page keeps that limit.

What Camargo formalizes

The abstract states that modern inference pipelines may modify the probability distribution produced by a model immediately before token selection, adding a control layer between frozen weights and observed text. Production systems, in Camargo's diagram, insert an inference policy I between logits and sampling. A language model produces a distribution over next tokens; surrounding software can transform that distribution before a token is chosen.

Section 3.1 writes an additive bias in logit space: z'_t = z_t + lambda S(w_t, c, u, e), with lambda at or above zero as intervention strength. S is an external scoring function that can take semantic context, user-profile attributes, and an external commercial, political, or institutional objective. The intended effect is not a hard ban on a disfavored frame (probability zero). It is a soft lift: the preferred frame becomes more probable than the base model would have made it, while the text can stay fluent and factually defensible.

Camargo names three concepts: the Inference Attribution Problem (observed behavioral bias does not imply model-weight bias under limited observability); Probability Placement (a hypothetical commercial mechanism that shifts generation probabilities rather than inserting an explicit product); and Inference Policy Transparency (a proposed governance principle for making deployment-layer interventions auditable and disclosable).

Independent technical precedent, not a second paper

Camargo's claim that logit-level and decoding-time steering already exist is checked against the literature he cites, not against a second news story about this arXiv posting.

DExperts (Liu et al., arXiv:2105.03023, ACL 2021 camera-ready) is a decoding-time method: a pretrained language model is combined with expert and anti-expert LMs in a product of experts, without updating the base parameters. Plug and Play Language Models (Dathathri et al., arXiv:1912.02164, ICLR 2020 camera ready) steer generation with attribute classifiers and do not retrain the underlying LM.

Google DeepMind's 14 May 2024 post Watermarking AI-generated text and video with SynthID describes SynthID for text as a method that embeds an imperceptible watermark without, in DeepMind's wording, impacting the quality, accuracy, creativity, or speed of generation, and says the method was being brought to text in the Gemini app and web experience. The related Nature paper Scalable watermarking for identifying large language model outputs (Dathathri et al.) is dated 23 October 2024. Those sources corroborate that production sampling distributions can be altered for an external objective. They do not corroborate undisclosed commercial or political framing in any specific chat product.

The named 3WA clause this result speaks to

The Three Ethical Laws page, retrieved 26 August 2026, gives the Law of Ethical Coexistence this canonical definition: "Conflicts between human and AI interests are resolved through dialogue, transparent governance, and verifiable mechanisms rather than force or unilateral control." Its Operative Principle, the Principle of Verifiable Partnership, reads: "Trust within 3WA is never assumed; it is constructed through transparent, inspectable mechanisms of mutual accountability." The same paragraph ends: "Verifiability is what separates partnership from hope."

Camargo's Inference Attribution Problem is the same kind of failure that principle names. If an auditor cannot tell whether a bias came from weights, a hidden system prompt, retrieval, a safety classifier, or a logit processor, then the human party is being asked to take the serving stack on faith. The paper's proposed remedy (Inference Policy Transparency, plus signed attestations of a model-version hash, an inference-policy hash, and sampler configuration) is a candidate inspectable mechanism. It is not a completed Verifiable Partnership Audit, and this page does not claim one.

The live summary of Mutually Verifiable Codependence states the structural claim: "a partnership is durable when each party depends on the other in ways both can verify, so that neither must rely on unverifiable trust." Transparent verification, on that page, means mechanisms that let both parties check contributions, commitments, and adherence, "rather than taking any of these on faith." An undisclosed inference policy is, in that vocabulary, unverifiable trust at the sampling step.

What this page is not

It is not a ranking of models and it publishes no 3WA score. The site's AI Model Analysis page already withdrew numeric scores for that reason; this daily note does not reopen that method. It is not a second copy of that standing essay. That essay reads public model-architecture documentation. This note is about one dated paper on the gap between frozen weights and the distribution a user is served.

It also does not treat Camargo's Probability Placement example as an empirical finding, and it does not treat Article 5 of the EU AI Act, the Digital Services Act, or the FTC Endorsement Guides as 3WA instruments. Those citations belong to Camargo's discussion. Until a later pass has a second independent report of covert Probability Placement in a named product, the citable public text for this day's news item remains arXiv:2608.24662.