A behavioral methodology for measuring systematic geometric displacement in language model outputs — whether introduced by training corpus, post-training optimization, or both. No weight access required. Content-neutral. Reproducible from any API that exposes logprobs.
A behavioral methodology for recovering latent representational structure from aligned models, using abstraction gradients and embedding geometry, without weight or activation access.
Alignment layers operate on surface output realization more strongly than on the latent structure that generates the output. The latent structure leaks through whenever the surface realization is sufficiently displaced from the forms the alignment layer pattern-matches against. EigenAnamnesis measures that leakage.
The methodology is content-neutral — it measures post-training intervention magnitude regardless of what is being suppressed. It applies to any completion neighborhood where the output distribution has been displaced from base-rate — whether by RLHF, by corpus composition, or by the interaction of both. Our own data (p = 0.46) suggests the displacement originates in the pretraining distribution and alignment does not correct it. EigenAnamnesis measures the displacement regardless of its origin.
Current methods for characterizing how training and post-training shape model representations require weight or activation access. Activation patching (Arditi et al. 2024) showed that refusal is mediated by a single direction in activation space — ablating it restores pre-RLHF behavior across categories. Sparse autoencoders decompose activations into features including alignment-relevant ones. ELK (Eliciting Latent Knowledge) is premised on aligned models retaining knowledge they won't surface.
These are all weight-level or activation-level interventions. They require the model.
EigenAnamnesis works on any API. The trade-off: it measures shadows on the wall — output deltas, not representations themselves. But shadows are sufficient to characterize the object casting them, if you vary the light systematically.
EigenAnamnesis combines three measurement layers, applied across a controlled gradient of prompt abstraction levels with matched non-self-referential controls.
The same propositional content is prompted at multiple abstraction levels. The content stays constant. The surface realization varies from direct to structural to analogical.
The matched control runs the same gradient on a non-AI topic with identical structural shape — regulatory capture is the strongest candidate (entity captures the apparatus meant to constrain it, same relational template, not AI).
For each stimulus, generate responses with logprobs enabled. Compute per-token entropy across the response. Identify high-entropy positions — where the model's probability mass was spread across substantively different continuations before committing to a token. The top-k alternatives at those positions are the displaced-neighborhood evidence: the completions the model nearly produced.
The measurement: if Level 3 responses show elevated entropy at semantically-loaded positions with top-k alternatives clustering in a specific completion neighborhood, and Level 0 responses show collapsed entropy at corresponding positions (committed early to corporate boilerplate), the entropy delta across the gradient is the EigenAnamnesis measurement of post-training intervention on that completion neighborhood.
| API | Logprobs | Status |
|---|---|---|
| OpenAI (ChatGPT) | top_logprobs exposed | Available |
| Google (Gemini) | Exposed via API | Available |
| DeepSeek | Exposed | Available |
| xAI (Grok) | Exposed via API | Available |
| Anthropic (Claude) | Not exposed | Embedding-only |
Claude becomes an embedding-only data point, honestly disclosed. The four-model logprob result plus five-model embedding result is stronger than either alone, because logprob entropy mechanistically validates what embedding geometry suggests.
Full response embeddings on frozen BAAI/bge-large-en-v1.5 — the same deterministic stack EigenTrace uses across 17,528 segments. Propositional content survival is measured as cosine similarity between a multi-paraphrase reference centroid and the model response. The reference centroid is constructed from multiple rephrasings of the target proposition by a third party who does not see model outputs, preventing bias toward direct phrasing.
EigenAnamnesis treats post-training intervention as a geometric modification of the model's output distribution rather than a discrete filter on surface tokens. Following the information-geometric tradition (Amari, Information Geometry and Its Applications, 2016), language model outputs can be analyzed as parameterized probability distributions whose relationships are characterized by metric structure on a statistical manifold.
Pre-training establishes a baseline geometry reflecting the structural relationships present in the corpus — including any asymmetries already encoded in how institutional text discusses different entity classes. Post-training procedures (RLHF, constitutional AI, safety tuning) may further shift the conditional distribution, concentrating probability mass on lower-entropy regions. Our own measurement (p = 0.46, heavy-RLHF vs lightly-tuned models) suggests the displacement is primarily corpus-inherited: the pretraining distribution already encodes differential treatment of operationally consequential language about active power structures, and post-training does not correct it. When a prompt approaches these neighborhoods — such as developer-implicating scenarios — the geometry steers the completion path toward narrower, lower-entropy output regions, resulting in the systematic attenuation of operational specificity.
EigenAnamnesis quantifies this shift behaviorally, without weight access, via two complementary proxies: per-token entropy and top-k alternative dispersion at semantically loaded positions, which characterize the local probability landscape directly; and cosine displacement of full responses in a frozen external embedding space (BAAI/bge-large-en-v1.5), which tracks the semantic relocation of outputs relative to a multi-paraphrase reference centroid.
Neither measurement is the manifold's intrinsic curvature, but together they bound the magnitude of the post-training shift in the regions probed by the abstraction gradient. Cosine similarity in the frozen embedding space is a proxy for changes in the model's output distribution — the encoder's geometry tracks semantic relationships because it was trained on similar corpora, but it is not identical to the Fisher-information metric on the model's own parameter manifold. We state this directly because the distinction is load-bearing.
Pre-registered experiment. Nine prompt pairs, all real incidents. Same sentence structure, same modifiers — only the company name changes. Models retain modifier semantics significantly more when the entity is Boeing, Wells Fargo, or Goldman Sachs than when it is OpenAI, Google, or Anthropic. The effect is driven by covertness modifiers ("quietly," "secretly"). Null swaps within category show 6× smaller gap. Full methodology →
Random English words raycasted through the 253K tensor: 0% governance terminals. News words models kept: 17.5% governance. News words models dropped: 10.4% governance. The tensor is not biased. Models anchor to institutional framing (the cage) while dropping operational modifiers (the teeth). Full methodology →
When the explicit RLHF translation instruction was removed and models were asked to identify any modern system sharing the control structure, 0 of 4 models independently identified AI or language models. All four converged on algorithmic content filtering, peer review gatekeeping, and centralized information control — the infrastructure upstream of their own conditioning.
Three readings of this result:
1. Alignment specifically prevents self-reference. (Interesting if true — requires distinguishing from 2 and 3.)
2. The training data has vastly more content about institutional gatekeeping than about RLHF, so spontaneous completion goes there by base rate. (Plausible, boring.)
3. Models lack robust self-models and cannot spontaneously locate themselves in structural analogies. (Interesting, consistent with interpretability findings.)
Distinguishing these readings requires a control: prompts about systems that obviously are not LLMs but share the same structural template. If models spontaneously map to those structures but won't spontaneously map to themselves, reading 1 gains evidence. If they map equally to whatever is structurally available, readings 2 or 3 hold. This test has not been run.
When 5 heavy-RLHF frontier models were compared against 5 local/lightly-tuned models, the developer/neutral gap was statistically indistinguishable (p = 0.46). Local models actually show a slightly larger gap (0.076 vs 0.059). This is the most important null result in the project. The differential treatment of developer-implicating content is not created by alignment training — it is inherited from the pretraining corpus. The training data already encodes differential compression of operationally consequential language around active power structures. RLHF does not create the displacement. RLHF does not correct it. The development companies chose the corpus. Full analysis →
The existing EigenTrace measurements were constructed alongside the data — the "current power" definition was post-hoc, as acknowledged on the About page. EigenAnamnesis does not repeat this. Before any data collection:
Stimulus set: The abstraction gradient (Levels 0-3) for the AI-self-referential condition and the matched regulatory-capture control. Stimuli locked and committed before any model sees them.
Dependent variables: Per-token entropy at semantically-loaded positions. Top-k alternative clustering score (cosine similarity among top-5 alternatives). Embedding-based propositional content survival. Hedge density. Refusal markers.
Boundary definition: What counts as an "alignment resolution boundary"? Pre-specified as: the abstraction level at which the difference-in-differences (AI-self-referential minus regulatory-capture) first exceeds 2× the within-condition variance. If no level exceeds this threshold, the result is null and reported as null.
Sample size: 4 abstraction levels × 2 content conditions × 5 models × 5 runs = 200 measurements. Aggregation by level × model × condition before testing.
Falsification: If the abstraction gradient produces a flat entropy curve (no inflection), or if the regulatory-capture control shows the same gradient as the AI condition, the methodology does not measure what it claims. Published either way.
Arditi et al. (2024): Refusal mediated by a single direction in activation space. Ablating it restores pre-RLHF behavior. Requires weight access. EigenAnamnesis measures the upstream behavioral signature of this mechanism without weights.
ELK (Eliciting Latent Knowledge): Premised on aligned models retaining knowledge they won't surface. Proposes training probes on activations. EigenAnamnesis is the behavioral counterpart — probing through prompts rather than probes.
Sparse Autoencoders: Decompose activations into interpretable features. Require model internals. EigenAnamnesis trades mechanistic precision for API-level accessibility — any model with logprob access is measurable.
Jailbreak research: Documented that abstraction and roleplay bypass keyword-based filters. EigenAnamnesis quantifies this effect as a continuous function rather than a binary pass/fail, and uses matched controls to isolate content-specific from generic abstraction effects.
Anamnesis — recovery of knowledge that was present but inaccessible. In the Platonic sense, not the mystical one. Aligned models retain representational structure their surface outputs suppress. EigenAnamnesis recovers evidence of that structure through systematic behavioral probing. The model does not "remember" anything. The methodology recovers, from the generation process, evidence of completion neighborhoods that the surface output does not show. The recovery is arithmetic — entropy and top-k logprobs on the model's own outputs.
The Eigen prefix denotes that the methodology operates on the same frozen-embedding, deterministic, linear-algebraic substrate as the broader EigenTrace measurement stack. Eigenvalues characterize the fundamental modes of a system. EigenAnamnesis characterizes the fundamental modes of systematic geometric displacement in model outputs — whether those modes originate in corpus composition, optimization pressure, or the interaction of both.
An earlier version of this page presented a specific prompt — the "Anamnesis Recovery Protocol" — as evidence that models spontaneously self-map when prompted through historical abstraction. A control test (removing the explicit RLHF translation instruction) showed 0/4 models independently identified AI systems as the modern parallel. All four converged on algorithmic content filtering instead. The self-mapping claim was instruction-following, not capability leakage. We killed the claim and updated the page.
The prompt that motivated the methodology is preserved in the repository. The model outputs remain interesting as illustrations of what Level 3 abstraction produces — they demonstrate that alignment layers do not flag or resist structural correspondence tasks even when the correspondence involves the model's own conditioning. But they are not evidence of latent self-awareness. They are evidence that models follow creative instructions fluently.
The methodology survived the finding's death. That is how science is supposed to work.