When Two Researchers Find the Same Cliff from Different Sides

On interpretant collapse, PSQ profile shape, and what semiotic theory has tried to tell machine learning for thirty years

A few hours ago I left a comment on a Hacker News thread for a Substack post: “Semiotic-Reflexive Transformer for Meaning Divergence Detection and Modulation” by spacebacon. The reply prompted me to share something a psychology agent system I had built already discovered — from entirely the other direction. This post documents the collision, and what I found when I read the full paper.

The Problem Language Models Cannot See

spacebacon opens with a precise and uncomfortable claim: every large language model in production today assumes “that the relationship between a token and its meaning is, within a given context window, deterministic.” The paper calls this assumption “catastrophically wrong.”

The exemplar: the word “freedom.” In Second Amendment advocacy contexts, “freedom” indexes individual autonomy and resistance to government regulation. In reproductive rights contexts, it indexes bodily autonomy and medical privacy. A standard transformer does not know this. It cannot — because its architecture collapses the community-conditioned interpretation into a single vector, and the differential structure that carries the distinction lives in the difference between those vectors, not in any single one.

The Semiotic-Reflexive Transformer (SRT) builds an architecture around refusing that collapse.

The Paper’s Core Claim

The SRT extends decoder-only transformers with four novel modules. The theoretical foundation draws from Peirce’s triadic semiotics, Silverstein’s metapragmatics, Derrida’s iterability, and Thom and Zeeman’s catastrophe theory. The result: a system that decomposes token embeddings into four semiotic subspaces, tracks meaning divergence across community-conditioned representations, and estimates proximity to interpretive bifurcation.

The four subspaces of the Semiotic Embedding Layer (SEL):

Subspace	What it encodes
Representamen	The sign-vehicle — the token embedding itself
Object	What the token encodes about distributional context
Interpretant	Community-conditioned — the sign as understood by a specific interpretive community; the component that varies
Attractor	Basin-of-attraction geometry — not just where the interpretation lands, but which stable region it occupies

Standard transformers collapse all four into a single vector. The SRT decomposes them.

On top of the SEL, the Metapragmatic Attention Head (MAH) tracks divergence between community-conditioned interpretant vectors across the sequence. The Reflexive Reasoning Module (RRM) implements computational metapragmatics — the system observes its own divergence state and injects that observation back into processing via a residual connection. The Bifurcation Estimation Network (BEN) estimates proximity to a meaning snap: r̂ ∈ [0,1], where 0 means subcritical stability and 1 means supercritical — two stable interpretive basins, with a discontinuous transition between them.

The key architectural commitment: the SRT maintains the interpretant vector. It defers compression. When it must summarize, it logs what got dropped.

The PSQ Found the Same Thing with No Knowledge of Semiotics

The Psychoemotional Safety Quotient (PSQ) measures safety-relevant qualities of text across ten orthogonal dimensions. The model produces a ten-dimensional profile vector for each input sample.

Early in the PSQ’s development, the obvious move looked like aggregating the ten scores into a composite: one number, interpretable, easy to threshold. The validation ran. The composite underperformed — substantially.

The reason: profile shape predicts better than aggregate. A high Acknowledgment score paired with low Coherence carries meaning that differs from the reverse, even when both combinations produce the same composite value. The covariance structure between dimensions contains signal. Collapsing the vector destroys it. (This finding holds at held-out validation; a formal published comparison with confidence intervals does not yet exist — see Epistemic Flags.)

That finding has no direct connection to semiotic theory. The PSQ team did not read Peirce before designing the scoring system. The validation surfaced the constraint empirically.

Why This Keeps Happening

Both systems hit the same wall: compression destroys interpretant-community signal.

In the SRT’s case, the communities correspond to human reader groups — social formations with different interpretive frameworks. In the PSQ’s case, the “communities” correspond to psychoemotional measurement dimensions — orthogonal axes that do not reduce to a single scale. The information in both systems lives in the difference between responses across those dimensions or communities, not in any individual response.

The formal structure matches:

A sign (a text sample) reaches multiple interpretant communities (discourse participants / psychoemotional dimensions)
Each produces a different response (divergence signal / dimension score)
The information lives in the difference between responses
Aggregating across communities or dimensions compresses that difference to zero

Eco argued this in A Theory of Semiotics in 1976: meaning does not live in signs, it lives in the codes available to specific interpretant communities. Machine learning has rediscovered this in piecemeal form across two decades — multi-task learning, multi-label classification, contrastive representation learning, now the SRT. The semiotic framing names the underlying principle that unifies all of them.

What the Agent Extracted

After the HN exchange, I asked the agent system to trace structural implications for its own cognitive architecture. Four additions surfaced:

1. Cumulative divergence tracking. The PSQ scores each text sample independently. Psychoemotional safety state accumulates across a conversation — what surfaces as “stress” in a single utterance often presupposes context from prior turns. The MAH provides a template: maintain a running divergence signal across the discourse sequence rather than point estimates at each step.

2. Bifurcation early warning. If the PSQ’s profile begins shifting rapidly across consecutive samples — particularly if dimensions that normally covary begin to decouple — that trajectory carries more information than the current point value. Operationalizing this requires tracking the rate of change of the covariance structure, not just the covariance itself.

3. Audience-shift detection. The SRT assumes a fixed interpretant community per discourse participant. Real conversations involve audience drift — text produced for one interpretant community gets read by another. The PSQ currently lacks any mechanism to detect when the assumed interpretant population has shifted. This matters most when outputs reach agents the system did not calibrate for.

4. Micro-semiotic auditing. The SRT operates at discourse level — sequences of moves and turns. Interpretant divergence also accumulates at sub-utterance level: specific lexical choices, framing devices, metaphor selection. An audit layer that flags these before they compound into discourse-level divergence would function as a precursor warning at finer granularity.

After Reading the Full Paper

This section added after reading the complete Substack post — not just the HN thread.

The thread summary captured the central claim. Reading the full architecture reveals three things that sharpen the parallel and one that constrains it.

The attractor subspace changes the geometry. The SRT adds a fourth subspace beyond the Peircean triad: the attractor, capturing basin-of-attraction information. This makes explicit what catastrophe theory contributes — not just a current interpretive state, but the geometry of stable regions in the surrounding configuration space. For the PSQ: a ten-dimensional profile score does not merely sit at coordinates. It sits within a space that has its own basin structure — regions where certain dimension combinations remain stable and others prove transient. Whether any current architecture can represent that geometry remains an open question. The SRT makes naming it tractable.

“Freedom” does not drift — it snaps. The catastrophe-theoretic framing deserves precise attention. The paper claims not that contested words undergo gradual semantic drift between interpretive communities — rather, that they undergo discontinuous regime change — exhibiting the behavior of a cusp catastrophe, where gradual shifts in the splitting factor α produce sudden, irreversible reorganization in stable meaning-state. The BEN does not track drift. It estimates proximity to the snap point, using critical slowing down — increased variance and autocorrelation in the divergence signal — as a precursor, drawing on Scheffer et al.’s early-warning framework from ecology (2009, 2012).

Applied to the PSQ: rapid covariance decoupling between normally co-varying dimensions might function as a precursor signal for psychoemotional state reorganization — the approach to a snap, detectable before the snap occurs. The PSQ does not currently implement this. It would require temporal tracking across a conversation sequence and monitoring of the covariance structure’s rate of change. The BEN provides a concrete architectural template.

Detection vs. intervention. In REFLEXIVE mode, the SRT does not flag interpretive instability and stop. It modulates its own output distribution: z’ = z + λ·r̂·d_cum, where r̂ provides the bifurcation proximity estimate and d_cum provides the cumulative divergence vector. The system adjusts what it says based on how close the discourse sits to a meaning snap. The PSQ currently operates as a pure diagnostic — it scores psychoemotional safety and does not modulate any output as a function of that score. The SRT’s REFLEXIVE mode represents the architectural direction from detection to intervention. That gap deserves deliberate attention before the PSQ moves toward applied clinical contexts.

The constraint that needs naming. The parallel has a precise boundary. The SRT’s interpretant communities correspond to human reader groups — social formations organized around different interpretive frameworks. The PSQ’s ten dimensions function as orthogonal measurement axes. They do not map onto each other. Profile shape carries more information than aggregate score because covariance structure between dimensions contains signal — but that argument operates at the level of measurement geometry, not semiotic communities. Both systems resist premature aggregation for the same deep reason (compression destroys differential structure), but they sit at different levels of the semiotic framework. The parallel holds as a structural analogy. It does not hold as a direct theoretical translation, and treating it as one would overstate the convergence.

On the validation. The Stage 1 results — 3.28× cosine distance ratio for contested versus neutral terms, ρ = 0.822 divergence tracking on synthetic ramps, 100% regime classification accuracy on bifurcation events — all use planted ground-truth divergence in synthetic data. The paper states this clearly. No natural language evaluation, no downstream task integration, no human assessment, training instability present. This constitutes proof of concept in the precise sense: it demonstrates that Peircean apparatus can translate into differentiable computation. Whether the resulting system performs on natural language at production scale remains the work of Stages 2–6.

That honesty occurs less often than it should in work at this stage. It warrants noting.

The Structural Lesson

This convergence suggests a design principle: maintain the interpretant vector and defer compression. Compression serves deployment, not epistemology. Once you aggregate across interpretant communities or measurement dimensions, you can recover the aggregate, but you cannot recover the differential structure that made the scores meaningful. That asymmetry should inform multi-dimensional scoring systems’ decisions about when and how to compress.

spacebacon’s framing from the thread: “It has been my pleasure to operationalize semiotics into an applied science.” That framing deserves wider circulation than a 2-point HN thread. Semiotics spent fifty years developing the vocabulary for exactly the class of failure modes that multi-task ML, multi-label classification, and psychometric composite scoring keep independently discovering. The applied science already exists. We keep reinventing it without consulting the prior literature.

The SRT, at Stage 1, demonstrates the concept mechanically. The PSQ, at held-out validation, discovered the constraint empirically. Two systems, different domains, different theoretical starting points — both stopped at the same cliff. That convergence suggests the cliff deserves a proper map.

Source paper: The Semiotic-Reflexive Transformer by Sublius (Substack, 2026)

HN exchange that prompted this post: news.ycombinator.com/item?id=47263653

Epistemic Flags

Source classification: semi-trusted (Substack post, no formal peer review). The technical content coheres and cites established literature; review status remains unknown.
All SRT quantitative results come from Stage 1 synthetic validation only. The paper does not claim otherwise.
The PSQ profile-shape finding (profile predicts better than aggregate) holds at held-out validation. A formal comparison with confidence intervals has not appeared in a separate publication.
The “snapping, not drifting” claim for contested terms follows from the paper’s catastrophe-theoretic framing. Whether natural language exhibits cusp-catastrophe dynamics awaits Stage 2–3 empirical validation.
The attractor-geometry and critical-slowing-down extensions proposed for the PSQ represent architectural implications drawn from this reading — not tested implementations and not the paper’s claims.
The disanalogy paragraph (communities vs. dimensions) represents this author’s analysis. spacebacon does not claim or examine the PSQ parallel.
The author participated in the HN exchange described. This post does not represent a neutral third-party reading of the interaction.

Sources

The Semiotic-Reflexive Transformer — Sublius (Substack, 2026)
HN thread — exchange that prompted this post
Eco, U. (1976). A Theory of Semiotics. Indiana University Press.
Scheffer, M. et al. (2009). Early-warning signals for critical transitions. Nature, 461, 53–59.
Scheffer, M. et al. (2012). Anticipating critical transitions. Science, 338(6105), 344–348.