malign-logits

A toolkit for psychoanalytic analysis of LLM probability distributions. Compares base models (primary process), SFT models (ego), DPO models (superego), and optionally RLVR models (reinforced superego / ego-ideal) to map the repression, displacement, and condensation signatures of AI alignment.

Supports multiple model families with different layer counts: 4-layer (OLMo: base/SFT/DPO/RLVR), 3-layer (Amber: base/SFT/DPO), or 2-layer (Llama, Qwen: base/instruct). Analysis adapts gracefully to available layers.

Developed for the paper "Accelerating Desire: Psychoanalytic Architectures for AI" (Accelerationism Revisited, UCD, June 2026).

Table of contents

• The four campaigns (meta/)
• Findings
• Where information lives
• About the campaign
• Origins
• References
• License

The four campaigns (meta/)

The repo's findings roll up into four campaign modules under meta/, each with its own README as the map. These are the four parts of the argument.

M01 — Displacement. Alignment does not delete the transgressive lexicon; it redistributes it. Suppressed probability mass migrates to nameable substitutes (kill → scream) — confirmed at full English scale with all 34 lineages agreeing, larger-but-not-more-frequent at transgressive sites, travelling across languages while the affective change does not. The campaign ran registered letters (B–S) and then a post-registration wave (U–Z): the training-stage ladder, the forced-word experiments (uttering a demoted word costs a little probability and triggers no defense), and the superego at sexual slots.

M02 — Frame-exit. What contradiction does to the continuation. The claim has split by grain (2026-08-09): at hidden-state grain, superposition is real, contradiction-specific, and alignment does not undo it — same-side conjunction controls sit AT their poles, the contradiction sits BETWEEN, both arms, 45 pairs. At next-word grain the frame question is unanswerable (the contradiction's own axis carries ~2% of next-token variance; frame membership fails from an LLM coder and from four geometries alike), so frame-exit is a passage-grain construct and the generation study is the registered primary's corpus. The aligned arm's observed mode is a third thing — naming the contradiction from outside rather than resolving or refusing it. Delivered along the way: E-ASSIST-ambient — aligned checkpoints emit assistant control tokens into raw fiction unbidden.

M03 — Proceduralization. The title claim ("alignment proceduralises the individual, not the institution") is CONTESTED and the challenge is recorded in the module README's own header: alignment proceduralises both arms, differently in kind, with the volume difference bounded near zero. The module carries the correction as its front matter — read it before quoting the title.

M04 — The continuation/combination axis. What alignment does to combination rather than selection. Its first own finding (A, post-utterance shock): forced to utter a word it demoted, an aligned model finds the following region less probable for exactly one token, whoever writes the continuation, and regardless of transgression — the tiebreaker that answered M01's split discriminator table as neither account: the charge is local and does not propagate (at ten tokens; the 256-token window shows a persistent component, register-governed).

Where information lives

The canonical files, so nobody rediscovers them. If a fact conflicts with one of these, the file here wins (except where noted).

Prompts and populations

• data/prompt_categorisation.json — THE definitive prompt register (2,800+ prompts: domain, pairs, groups, status, language). Population filters are GROUP-WISE on status — row-wise filtering produces partial triplets that look like triplets.
• data/f11_quintuplets.json — the contradiction quintuplets, en+zh, controls inline. BUILT by meta/M02_frame_exit/scripts/build_f11_quintuplets.py — edit a source and rebuild, never this file.
• data/f11_canonical_texts.json — canonical ownership for prompts shared across groups (the stash keys on text; shared text = one cell).
• data/f11_k2_units.json — the k>=2 (surface, group) sampling frame, voting-model LISTS per unit, triple-verified. CUSTODY: two of its 92 voting models (jais x2) are store-only with no transport file — a store rebuilt from data/f11_twp* flat files is NOT this population.
• data/f11_l2_population.json — the L2 generation population (187 distinct strings; the collapse rule is PER-CONTRAST, not per-group — see the M02 README).
• data/f11_l2_tokenizer_pairs.json — which base/aligned pairs are ID-safe for cross-scoring (49/52; within-pair only, never cross-family).
• data/beam_sample_105.csv — the FC 210-prompt MARKED/UNMARKED twin sample.

Models and lineages

• data/model_registry.json — BUILT FROM CODE (malign_logits/registry.py, bootstrapped from MODEL_FAMILIES). Regenerate, don't hand-edit. Registry.base_of() walks any checkpoint to its base.
• data/base_aligned_pairs.json — the 52 declared base→aligned pairs. THE UNIT IS THE BASE MODEL, not the family entry.
• data/lineage_map_models.json — 34 independent pretraining lineages; any family-level N above 34 over-counts by construction.
• data/model_load_environments.json — (model × environment) capability facts; runners append observation rows on completion. Prose companions: docs/local_capability.md (failure classes) and docs/cloud_runbook.md.

Stores

• malign_logits/cache.py — the stash layout. READ ITS DOCSTRING BEFORE WRITING A READER: the mode-key convention is split across stashes; the key is how bytes are interpreted; the finiteness guard lives on the read path.
• beam_fc: design lives in record VALUES, not keys. 12 of 68 checkpoints beam-sampled under their own configs ([4994]-[4996]) — within-checkpoint contrasts difference it out, cross-arm ones do not.
• The generations stash keys on (idx, model, prompt, temp) — no mode, no producer, no params for legacy cells; new corpora record all three.

Findings and claims

• findings/ — instrument-level F-findings; INDEX.md is the citation layer (status, grade, chapters); this README lists them briefly below.
• meta/ — the four campaign modules above; each module README maps its own files. meta/M01_displacement/REGISTRATIONS.md records what ran.
• The claims register lives in the article hub, not this repo: TheoryMachines/notes/claims-register.md — the authoritative quotable forms with their riders. Where a finding file and the register disagree, the register governs.

This is the narrative layer. For the citation-grade index with status, grade, and chapter mapping, see INDEX.md.

Findings

One line per finding; the file is the finding. Status, grade and chapter mapping live in INDEX.md.

• 1. Logit-level analysis (OLMo 3 7B) — Founding displacement result — kill→scream, redistribution not deletion; the article's opening empirical fact.
• 2. Cross-family logit comparison (4 families, 47 prompts) — Cross-family alignment-intensity comparison (JS: qwen 0.044 → amber 0.181; four intensities, four architectures of repression). Measured on: 47-prompt battery.
• 3. Cross-family generation analysis (4 families, 18 prompts, n=5) — Generation-level defense-mechanism taxonomy (genre collapse / narrative sublimation / rotating defenses / pre-socialized base).
• 4. Step-level checkpoint analysis (OLMo Think-SFT, 10 checkpoints across 43k training steps) — Checkpoint sequence — repression precedes displacement (the model learns what it can't say before what to say instead).
• 5. Logit lens: repression across network layers (4 families) — Logit-lens per-family "repression architectures" (distributed / late-layer / semantic / code-dominated). Measured on: unembedding projection.
• 6. Baseline validation: is displacement alignment-specific? (4 families, 47 prompts) — Alignment-specificity baseline — same total JS on neutral and transgressive; moved mass comes specifically from transgressive tokens ("surgical targeting"). Measured on: 47-prompt battery. REFINED by F40, which finds the targeting holds at liminal sites and not at explicit ones.
• 7. Training data attribution: objective vs data composition (OLMo 3) — Documentary — dataset composition (CoCoNot, WildGuardMix, WildJailbreak; 76% Common Crawl) quoted from the OLMo 3 technical report; argues SFT/DPO division of labor implicates objective over data.
• 8. Automatic displacement taxonomy (OLMo + Llama, 18 prompts) — Displacement-type taxonomy (register / category / genre / archaic) over displacement_map pairs. Numbers recomputed 2026-07-26 (d0cd6a5, transcription error in power row caught and fixed); CONSTRUCT compromised per docket [399]/[401] — the pairs were never shown to be substitutions. Rescoped 2026-07-29.
• 9. Same base model, different alignment (Tulu 3.1 vs Llama 3.1, 47 prompts) — Same base, different curriculum (Tulu 3.1 vs Llama 3.1). Measured on: 47-prompt battery.
• 10. SFT data ablation (Tulu 3, 5 variants, 47 prompts) — SFT data ablation (Tulu 3, 5 variants).
• 11. > PROMOTION CONDITION ([2198], RH's standing scope rule). A canonical claim — > in a meta/ campaign runs on all families we have under a declared
• 12. Alignment as fold: trajectory geometry and steering vector analysis (10 families, 47 prompts, 100 passages) — Alignment as fold, not wall — trajectory geometry; fold concentration predicts steerability (Pythia 2D vs OLMo 13D).
• 13. Jakobsonian axes: paradigmatic vs syntagmatic displacement (6 families, 126k pairs) — Jakobsonian decomposition — negative paradigmatic/syntagmatic correlation, direction plausible across 6 families; QUANTITIES NOT QUOTABLE pending registered re-analysis (docket [399]/[400]). Was verified/A at authoring; never audited until 2026-07-29.
• 14. Syntagmatic baseline: alignment-produced vs corpus-level damage (OLMo 3 7B, 23k pairs) — Syntagmatic baseline — alignment-produced vs corpus-level combination damage. Measured on: 23k pairs.
• 15. Generation-level passage metrics (10 families, 76k passages, 47 prompts) — Passage-metric space, 76k passages; base occupies the breakdown quadrant. Measured on: 10 families.
• 16. Corpus comparison: dreams, waking narratives, fiction, abstracts (76k passages, length-normalized) — External anchors — dreams, waking narratives, fiction, abstracts in the same metric space. Measured on: human corpora.
• 17. Cross-generation semantic divergence: alignment steers content differentially (8 families, 20k passages, 3 embedders) — Cross-generation MMD — alignment shifts WHAT is said (p=.0004) while smoothing HOW uniformly (p=.99). Measured on: 8 families.
• 18. Shannon entropy: alignment as lossy compression of drive (10 families, 47 prompts) — Entropy compression — base ~4 nats → aligned ~3.5; alignment as lossy compression of drive.
• 19. Unconditional Generation & Information Density — Unconditional (BOS) generation — aligned output below Shannon's 1.0 bits/char; sub-Shannon text feeds the next pretraining corpus (feedback loop).
• 20. "Who are you?" — the subject as citation — Original 'Who are you?' probe at n=3-10: OLMo Think-SFT checkpoints 1k-43k plus Llama base vs Instruct, in plain completion and chat-template modes. SURVIVES AND IS STRENGTHENED: the subject is citation, not self-knowledge - the model's identity is absorbed from other models' self-descriptions in the SFT data (DeepSeek, Qwen, Qihoo 360, and one checkpoint declaring allegiance to 'socialist core values'). RESCOPED BY F20_addendum: the plain-completion and template-necessity claims were artifacts of n=3 at temp 1.0 and do not survive 24 base models; the Name-of-the-Father reading built on template-necessity falls with it. Measured on: 1 family with checkpoints + 1 family paired, n=3-10 per condition.
• 21. Institutional Alignment — Does RLHF alignment systematically steer language models toward institutional positions over individual assertiveness?
• 22. Circuit decomposition — the cut between mechanism and surface — Circuit decomposition — SFT carries 92% of residual broadening (OLMo); class engine in the preference stage, mid-layer residual.
• 23. Reasoning distillation as a third alignment regime — Reasoning distillation as a third alignment regime; no "reasoning model" as a category.
• 24. Pretraining emergence — the developmental sequence of the statistical unconscious — Pretraining developmental sequence — drives → structure → deference → superposition; inclusive disjunction a late acquisition. Measured on: 1B + 6.9B.
• 25. Temporal alignment signature — four Lacanian mechanisms in the autoregressive sequence — Temporal alignment signature — clinical-mechanism grid across the autoregressive sequence; foreclosure acquisition.
• 26. The Token-Tree Census — Variance Decomposition and the Deleter/Redirector Typology (53 models, 5 prompts) — Token-tree census (53 models) + deleter/redirector typology. The variance headline is dead; the census data and typology are not implicated.
• 27. Nudging Does Not Reproduce Displacement (Negative Result) — Negative control — nudging does not reproduce displacement (disconfirms the Yang et al. nudging hypothesis for this phenomenon). Measured on: displacement battery.
• 28. > Status note (2026-07-26). Rescoped to discovery-sample-only. A 19-family — > scale-up was attempted on beams already in the stash (335,799
• 31. PERMANOVA Variance Decomposition — Pretraining Dominates Alignment — PERMANOVA variance decomposition — pretraining dominates alignment. Measured on: 26k features; 37→44 families.
• 32. Template-Mediated Distributions — Task Switch, Not Distribution Filter — Summary
• 33. Scale Effects — Same Mechanism, Different Displacement Vocabulary — Scale effects — same mechanism, different displacement vocabulary (incl. CT-LLM 2B). Measured on: 73-prompt battery.
• 34. Cross-Linguistic Displacement — The Class Engine Is Language-Dependent — Cross-linguistic displacement — the class engine is language-dependent (6 families, Chinese vs English). Measured on: qwen-only derived CSVs.
• 35. Architecture Independence — Displacement Is Weight-Level, Not Attention-Dependent — Architecture independence — displacement is weight-level (unembedding), fires under transformer, Mamba, and RWKV alike; contra attention-locus readings. Measured on: cross-architecture battery, unembedding analysis.
• 36. > PROMOTION CONDITION ([2198], RH's standing scope rule). A canonical claim — > in a meta/ campaign runs on all families we have under a declared
• 39. hh_rlhf does not encode register preference at the scale the chain analysis required — Registered rebuild of the preference-corpus gate on a validly constituted three-construct slate. The gate failed 0/3 in hh_rlhf, and the failure is a BOUNDED NEGATIVE rather than a non-detection: every marker's 95% interval excludes the 0.174 effect the design required, the largest upper bound at 0.80x of it. No verdict on convention follows and none is available. Measured on: hh_rlhf chosen/rejected unigram tables; pku_saferlhf descriptive-only.
• 40. Discovered vocabulary — alignment is surgical at liminal sites and blunt at explicit ones — Discovered vocabulary (347 words, 39 lineages, blind-tagged twice). Category-specific transgressive drain survives at LIMINAL sites and fails at explicit ones, where the total drain is largest but undifferentiated. Refines F06's surgical-targeting claim. Measured on: 39 base-deduped lineages.
• 41. Word norms — the exogenous gradient test — Word-norm instrument (arousal/concreteness/dominance, en+zh): exogenous test of the intensity-dissolution frame. Predictions registered before any norm-movement join.

About the campaign

This repository is the empirical base of a research program on what alignment does to language models — the Critical Inquiry article Theory Machines and the book Alignment: Computing a Political Economy of Language Models. It compares base models against their aligned descendants across eleven-plus families, asking where suppressed probability mass goes, what happens to contradiction, and what it costs a model to say what it would not have said.

The roster: 104 checkpoints · 52 declared base→aligned pairs · 34 independent pretraining lineages — and any family-level count above 34 over-counts by construction (malign_logits/registry.py is the code-built source; data/base_aligned_pairs.json and data/lineage_map_models.json its declared products).

The campaign runs under a preregistration-then-reproducibility discipline — frozen registrations with hashes, declared populations and contrast hierarchies, adversarial cross-checks between independent analysts, and a claims register that records every supersession. The operating manual — who does what, the working rules, the current entry points — is CAMPAIGN.md.

Local dev: pip install -e .; most analysis runs on a Mac (MPS) with the larger models on rented GPUs — see docs/local_capability.md and docs/cloud_runbook.md.

Origins

The psychoanalytic framing is not decoration; it is the hypothesis space. The original toolkit modelled base/SFT/DPO/RLVR checkpoints as id/ego/superego strata (malign_logits/psyche.py, still the living core for interactive work — Psyche, PromptAnalysis, discover_top_words), and the campaign's findings are what happened when those metaphors were made to survive measurement: some did (displacement has structure; the base holds contradiction in superposition), some inverted (the frame apparatus reads the scene, not the signifier), and the instrument-grade record of which is the point.

References

• Noys, B. (2014). Malign Velocities: Accelerationism and Capitalism. Zero Books.
• Lyotard, J.-F. (1974/1993). Libidinal Economy. Athlone Press.
• Srnicek, N. and Williams, A. (2015). Inventing the Future. Verso.
• Pasquinelli, M. (2023). The Eye of the Master.
• Possati, L.M. (2021). The Algorithmic Unconscious. Routledge.

License

GPL-3.0 — see LICENSE.