llm-research

LLM-interpretability research, organized as reproducible experimental arcs, with a secondary, AI-generated theory knowledge base alongside (see theory/ below). The code depends on standard scientific libraries (PyTorch, HuggingFace transformers, numpy, matplotlib) plus one reference implementation, jlens, at a recorded commit. The largest investigation applies Anthropic's released NLA (Natural Language Autoencoder) verbalizer/reconstructor model pair to local Qwen2.5-7B-Instruct and probes layer-20 hidden-state geometry; three further arcs cover subliminal trait transfer, the structure of Qwen2.5-7B's input-embedding table, and a partial replication of the J-lens / J-space global-workspace result.

This repository is a research workspace, not a software product. It collects the work product — the dated observations, the figure/audit pipeline, a background theory knowledge base — rather than a polished library. Claims are held to the standard described under "Epistemic discipline" below: technical claims that a conclusion rests on cite a primary source, and findings are framed as hypotheses until the evidence settles them.

It is also exploratory, self-directed, agent-assisted work. Much of the implementation — capture and analysis scripts, figures, audit infrastructure, observation drafts — was carried out by Claude Code sessions under direction; the research questions, scope calls, and sign-off are the author's, and each arc README records that split explicitly. Nothing here has been published, externally reviewed, or replicated by anyone else, and the findings should be read as provisional. The Methodology below is the part offered with confidence: it exists to make being wrong visible.

Results

One block per arc: the headline finding, one figure where the arc has one (arc 02 has none), the strongest caveat and a link. The figures are stored in Git LFS, so a clone made without LFS shows them as broken images; § Prerequisites has the setup and the recovery command. Every number here is taken from the linked arc README, which carries the evidence, the scope qualifications and the full list of limitations; where this summary is shorter, the arc README is the authority.

Arc 04 — J-lens / J-space replication on Qwen2.5-1.5B/7B: partial

Closed 2026-07-22. The J-space phenomenon reported for Claude-family models [gurnee2026-workspace] partially replicates on Qwen2.5-Instruct (1.5B bf16, 7B nf4). The excess-over-random FVE figures below are from recomputes after that date: the paper-metric recompute (2026-07-24/25), the C4 PII-redaction re-run (closed 2026-08-16) and the dimension-matched recompute (2026-09-23).

• Replicates: the J-lens as a readout instrument. In the multihop early/mid layer bands it surfaces unspoken intermediate concepts where the logit lens finds nothing. Counterweight: the logit-lens median emergence layer is still earlier at 1.5B (19 vs 23), and the J-lens fails to surface the token in more cells at both scales (1.5B 16/108 vs logit 1/108; 7B 14/108 vs 8/108).
• Occupancy against the paper's 10% ceiling. On the paper's excess-over-random FVE metric, 1.5B breaches the ceiling at the hump (L21 excess 11.15%, CI95 [10.95, 11.40], over all valid positions) and 7B stays under it at every K tested. At the paper's K rule (K=23 held-out, K=23-24 grid) the 7B peak excess is 5.88% held-out and 4.72% on the scan grid; at K=58 it is 7.67% held-out and 6.16% on the grid, with every CI95 upper bound at or below 8.03%. Excess depends on K/d; at matched K/d the 1.5B/7B ratio, taken per channel, is 1.52× held-out and 1.76× on the grid.
• Strongest positive: a graded relational effect (stage 5.2). Swapping an unspoken concept along its J-lens vector moves the concept's entailed property more than an equal-magnitude logit-lens token-steering control. The absolute gap is +5.0 nats (1.5B L18, n=7) and +1.9 nats (7B L19, n=17) on the auto-detected subset. The per-item SD is about equal to the gap at 1.5B (4.9) and larger than it at 7B (4.1): n is small and a few high-movement items dominate. The paired gaps pass an exact sign-flip test (p=0.0156, p=0.0001).
• Does not replicate: four items. The discrete entailed-property flip (rate 0.000 at both scales). J-space membership as the causally privileged ingredient: the J-space component of concept vectors shows no detectable effect in report swaps, bounded at ≤3.8pp at 7B (exact 95%, 0 discordant pairs of 78) and looser at 1.5B (p=0.375 on 5 discordants). The kurtosis workspace-onset signature (inverted on Qwen). The small-model structural picture at 7B (7B has ~3× lower occupancy and a U-shaped depth profile, on absolute varfrac@25 over 30 held-out wikitext prompts; that is a different measure from the excess-over-random FVE in the occupancy item above).

Excess-over-random FVE by layer, with cluster-bootstrap CI95 bands and the paper's 10% ceiling dashed, on (a) held-out C4 prompts and (b) the wikitext scan grid. 1.5B is plotted at K=25 in both panels. 7B is plotted in both panels at the paper's K rule and at K=58, which matches 1.5B's K/d, and in panel (b) also at K=25. 7B stays under the ceiling at every layer and K.

Strongest caveat. One model family at ≤7B, unreviewed and unreplicated outside this repo; and the negative results are measured bounds, not demonstrated zeros. Full account: research/arcs/04_jspace/README.md.

Arc 03 — embedding atlas of Qwen2.5-7B's input-embedding table

Closed for new experiments 2026-07-15. Findings are held as working hypotheses.

• Pre-registered predictions, adjudicated: P1a PASS, P1c FAIL, P1d FAIL; P2 refined-not-falsified; P1b, P1e and P3 not run. The two failures: period→comma aggregation peaks at L0, not deeper (P1c), and delimiter matching is carried ~99% by near-DC RoPE bands, a static content match rather than positional resonance (P1d).
• One 21-dim structural block. Over all 149,706 alive rows there is exactly one entangled dimension block: 21 correlated dims (|r|>0.3), cross-script, head-loaded by frequency (the first token-id decile carries 1.5x the block norm-fraction floor; Spearman -0.206 vs -0.003 control). Outside it, dimensions are near-independent (|r| mean 0.021).
• Near-isotropy null. Random-pair cosine +0.0097; PC1 explains 1.21%.

Block norm-fraction by token-id decile against a random-21-dim control: the first (most frequent) decile carries 0.1143 vs the control's 0.0753.

Strongest caveat. Single model, single revision: the isotropy null especially needs a second model before any general reading. The arc's audit checks arithmetic consistency only. Full account: research/arcs/03_embedding-atlas/README.md.

Arc 01 — NLA verbalizer on Qwen2.5-7B layer 20: basin candidates

Paused 2026-05-15. One working synthesis, held as a working hypothesis and not a settled claim: layer-20 h-space appears to have discrete attractor basins separated by sharp boundaries.

• Linear interpolation between two AR-encoded natural-language anchors produces a discontinuous AV-text transition at a roughly constant geometric step. Dense re-sampling found a hybrid "Definition + Poem" plateau (t∈[0.395, 0.4450]) and a flip to the poetic format in a single Δt=0.0025 step, at t=0.4475.
• Scope: one anchor pair, one layer, one model. The plateau-attractor margin is +0.061 over the nearest single anchor, against +0.25 between anchors, so the arc's framing is "basin candidate / shallow basin" until more anchor pairs and layers replicate.

Dense interpolation between the two AR-encoded anchors: factual → hybrid plateau → poetic/nature (t≥0.4475).

Strongest caveat. The AV-decoder format bias is unaudited (#8): if the verbalizer emits the same templates on random h-vectors as on semantically loaded ones, every interpretive reading in the arc has been filtered through the verbalizer's prior. Every result is also on one model at one layer. Full account: research/arcs/01_nla-verbalizer/README.md.

Arc 02 — subliminal trait transfer: a Step-0 null, paused

Paused 2026-06-10 after Step 0; Steps 1-2 were not started. The arc has no figures.

• Step-0 null. A five-scheme decode of locally regenerated owl/neutral number streams returns zero owl-lexicon hits in either condition (owl_rate = neutral_rate = 0.000, z = 0, p = 1.0), with the planted-string positive control passing. A literal decodable channel is not supported for this setup.

Strongest caveat. The arc's question, non-semantic statistics (A) versus semantics in the model's own coordinates (C), is untested: the two steps designed to address it were never run. Step 0 also tests a local stand-in (streams regenerated with an open teacher), not the paper's own data. Full account: research/arcs/02_subliminal/README.md.

Data correction notice — arc 04's C4-en corpus

[!NOTE] Data correction — closed 2026-08-16. Arc 04's seeded C4-en corpus slice carried 120 pieces of third-party personal data and was redacted on 2026-07-29; because the redaction is not length-preserving, every result computed on that corpus — the corpus-invariance and held-out-sample checks and the H3 @10 corpus-dependence qualifier — was recomputed on the redacted text (2026-08-15/16). Outcome: two audit pins moved, each by less than 0.02, both in the held-out channel, and no headline conclusion changed; the two failure modes pre-registered before the re-run did not materialise. Arc 04's primary fitting corpus is wikitext-103, which was scanned and left unmodified, and no other arc used C4. The full record — per-class counts, root cause and the reproduction recipe — is in research/arcs/04_jspace/data/README.md, which links the per-claim record. Surfaced here because a correction of this kind should be visible at the entry point rather than discovered two levels down.

What's here

research/    Experimental research, organized into arcs (focused investigations);
             each arc's capture / analysis / render / audit scripts are in its scripts/
src/         The llm_research package: shared model loader + NLA probe
tests/       Tests of src/, scripts/ and the moved arc scripts
theory/      Citation substrate for the arcs (paper index, excerpts, notes); otherwise
             secondary: AI-generated LLM-theory knowledge base + 5-paper LaTeX series

• research/ — Investigations organized into arcs under research/arcs/<slug>/, each cohering around one research question, plus research/observations/ for one-off findings and research/archive/ for retired material. Four arcs exist; two are paused mid-program, not finished. The largest is research/arcs/01_nla-verbalizer/; the two carrying the most informative epistemics are 03_embedding-atlas (predictions registered before the runs, two of them falsified) and 04_jspace (a partial replication that came out weaker than the original, with a causal split). The arc lifecycle and reproducibility disciplines are in research/ARC_PROCESS.md; the per-arc status index is research/README.md.
• research/arcs/<slug>/scripts/ — Each arc's pipeline scripts, one family per arc by prefix (nla_*, subliminal_*, emb_*, jspace_*), beside the _-prefixed helper modules the arc 01, 03 and 04 families import. Each family covers capture (writes .pt artifacts; arc 02's writes JSON/JSONL and text), analysis, figure render (matplotlib), and an *_audit_findings.py that re-derives the numerical claims that arc's prose relies on from committed artifacts. Arc 04's jspace_* family and arc 03's emb_* family follow the nla_* pipeline's artifact/audit shape; the cross-arc conventions are in research/ARC_PROCESS.md § "Raw data is a deliverable", and each arc README carries the per-arc detail. The arc 01, 03 and 04 capture scripts load models through the in-repo llm_research package (src/llm_research/, installed by pip install -e '.[dev]'): hf_models (HuggingFace loader, nf4 / int8 / bf16 / fp16 / fp32 modes) and nla_probe (the NLA verbalizer / reconstructor pair); the rest are render/analysis-only (torch / numpy / matplotlib). Arc 01's nla_* scripts are in research/arcs/01_nla-verbalizer/scripts/. Arc 04's jspace_* scripts are in research/arcs/04_jspace/scripts/. Arc 03's emb_* scripts are in research/arcs/03_embedding-atlas/scripts/. Arc 02's subliminal_* scripts are in research/arcs/02_subliminal/scripts/; its step-0 generator, subliminal_step0_decode.py, calls transformers directly.
• theory/ — The citation substrate for the arcs, and otherwise a side project. Arc prose cites primary sources by their paper-key in kb/index/papers.json, and the arc READMEs link into the knowledge base in a few places, for paper metadata and background notes. Beyond that role most readers can skip it. The knowledge-base notes and the LaTeX series were written largely by Claude Code subagents (one topic area per agent for the notes, one section per agent for the series), and they are not part of the research findings. Contents (v2 layout): kb/notes/ digested synthesis (one file per topic), kb/excerpts/ verbatim source passages, kb/index/ (papers.json, topics.md, timeline.md), kb/glossary.md, and sources/papers/ primary-source PDFs. series/ holds a 5-paper LaTeX series (architecture, training, reasoning, interpretability, evaluation-alignment) with cross-paper references. Start at theory/README.md.

Methodology

The arcs under research/arcs/ are run to a common discipline, specified in research/ARC_PROCESS.md. It is stated here because it is the part of this work offered with any confidence — and because where it is not applied uniformly, that should be visible from the entry point rather than discovered two levels down.

• Predictions before runs — unevenly. Arc 03 is the strongest case: seven predictions registered on 2026-06-11, before the attention captures they constrain, adjudicated mechanically at close as P1a PASS, P1c FAIL, P1d FAIL, with P2 refined-not-falsified and P1b, P1e and P3 never run (plans/2026-06-11-predictions.md). Arc 02's plan states falsifiable predictions per hypothesis and an explicit pre-commitment clause, though the arc paused before the tests they govern. Arc 04's README records one of its four robustness axes (the quantization control) as pre-registered in its design plan, the rest gated on thresholds fixed before each run. Arc 01 grew from open-ended themes with no pre-registration. Three of the four arcs' registers are partial — read each arc's own account rather than this summary.
• Audit scripts. The *_audit_findings.py scripts (each in its arc's scripts/ directory) re-derive the numbers an arc's claims rest on from its committed artifacts, so a figure quoted in prose that has drifted from the artifact it came from fails the audit. Arcs 01, 02, 03, and 04 each have one. These audits check arithmetic consistency only — they cannot catch a methodological error, a capture-protocol bug, or interpretive overreach. Arcs 01 and 03 state that limitation in their READMEs, arc 02 in its data README § Audit, and arc 04 in its § Reproducing.
• Datasets committed and pinned. Arcs 01, 03, and 04 commit raw .pt artifacts under the arc's data/ (Git LFS) with a MANIFEST.json recording per-file sha256 and provenance; arc 02's Step-0 data is small JSON/JSONL and text in plain git under the same MANIFEST.json convention. The intent is that a clean clone re-renders every figure and replays every audit — with one documented exception: arc 04's full fitted-lens tensors sit behind an opt-in LFS download (three are committed; the two 1.5B nf4 lenses are pending issue #47), so 7 of its checks fail on a default clone — 3 LFS-stub reports + 4 MISSING (see Running the research pipeline).
• Human/AI division of labor, recorded per arc. Every arc README states what Claude Code sessions implemented and what was directed, constrained, and signed off by hand — down to which framings are the agent's paraphrase rather than the author's own words (arc 02) and which methodological problems the agent missed until a human raised them (arc 01, theme 9).
• Negative results kept, not buried. Arc 03's two falsified predictions and arc 04's four non-replications each get their own numbered write-up with the measurements attached, and the negatives have dedicated observation files. They are the strongest available evidence that the pre-registration and audit machinery is not decorative.

Counts that move as arcs develop — observation totals, figure totals, script totals — are kept in the arc READMEs where they are maintained, not here. Audit check-counts are the exception: they appear below with the date they were re-derived, because a reader needs an expected value to compare a local run against. Re-run the audit rather than trusting any number quoted here.

This is exploratory, self-directed work: unpublished, unreviewed, and open to being wrong — the discipline above is there to make being wrong visible.

Prerequisites — Git LFS is REQUIRED

Research figures (research/**/figures/*.png) and raw datasets (research/**/data/*.pt) are stored via Git LFS — see .gitattributes. Install and initialize Git LFS before cloning or working the repo, or those files show up as phantom modifications (the working tree holds pointer files, not content):

git lfs install            # one-time, per machine
git clone <repo-url>       # LFS content then fetches on checkout
# already cloned without LFS? recover with:
git lfs install && git lfs pull

Setup

Python >= 3.10. pyproject.toml declares every PyPI dependency (torch, transformers, accelerate, bitsandbytes, huggingface_hub, safetensors, pyyaml, numpy, matplotlib, datasets). The jspace_* scripts also need jlens (anthropics/jacobian-lens), which is not on PyPI and is installed editable from a checkout next to this repo; the arc-04 data/MANIFEST.json records the commit (jlens_pin) to check out:

python -m venv .venv && source .venv/bin/activate
pip install -e ../jacobian-lens    # jlens — every arc 04 jspace_*.py script
pip install -e '.[dev]'            # this repo's deps, its llm_research package + pytest
                                   # (re-run it in a checkout installed before src/llm_research existed)

Model checkpoints download from the HuggingFace Hub into the directory named by LLM_RESEARCH_MODEL_CACHE, or into the HuggingFace default cache when it is unset. Checkpoints downloaded before 2026-09-26 live in the directory the former llm-surgeon toolkit cached to (.cache/models/ inside the llm-surgeon checkout by default, or LLM_SURGEON_CACHE_DIR when set), so export LLM_RESEARCH_MODEL_CACHE to that directory before a capture run. On a cache miss the loader downloads the multi-GB checkpoints, except in the seven arc-01 scripts that force HF_HUB_OFFLINE=1 (grep -l HF_HUB_OFFLINE research/arcs/01_nla-verbalizer/scripts/nla_*.py), which raise an OSError at model load naming the cache directory probed and LLM_RESEARCH_MODEL_CACHE. The NLA verbalizer (AV) and reconstructor (AR) load at pinned Hub revisions (llm_research.nla_probe.AV_REVISION, AR_REVISION), so an offline run needs a cache that holds those exact snapshots; with LLM_RESEARCH_MODEL_CACHE exported, hf_models._is_cached(nla_probe.AV_ID, revision=nla_probe.AV_REVISION) (both imported from llm_research) and the AR equivalent check it. The seven scripts set the offline flag unconditionally, so recover a cache holding another snapshot of the same repos with huggingface_hub.snapshot_download(<AV_ID or AR_ID>, revision=<AV_REVISION or AR_REVISION>, cache_dir=<the cache dir>), run online: the weights are unchanged since 2026-03-16, so only the small files at the pinned revision download and the cached weight files are reused.

Running the research pipeline

Capture scripts write .pt artifacts (working cache under .cache/, gitignored; the committed copies live in each arc's data/). Render scripts turn artifacts into figures; each arc's audit re-derives that arc's claims from them. Arcs 01 and 03 were re-verified from a clean clone on 2026-08-29 (their committed logs reproduce byte-for-byte). Arc 02's audit gained seven arc-manifest claims on 2026-09-28 and now totals 111 PASS; its committed 2026-08-17 log records the earlier 104. Arc 04 was re-measured in-session on 2026-09-24 after its audit gained CHECK P. No audit log was committed for that run; re-derive the totals with the command below:

python research/arcs/01_nla-verbalizer/scripts/nla_audit_findings.py         # arc 01 → SUMMARY: 196 PASS | 0 FAIL
python research/arcs/02_subliminal/scripts/subliminal_audit_findings.py  # arc 02 → SUMMARY: 111 PASS | 0 FAIL | 5 UNVERIFIABLE
python research/arcs/03_embedding-atlas/scripts/emb_audit_findings.py  # arc 03 → SUMMARY:  99 PASS | 0 FAIL
python research/arcs/04_jspace/scripts/jspace_audit_findings.py      # arc 04 → SUMMARY: 1053 PASS | 7 FAIL on a default clone;
                                                                     #          1088 PASS | 4 FAIL with the lens cache pulled

Arc 04's 7 failures on a clean clone are expected, not regressions: the three lenses refit in the C4-redaction re-run are LFS-committed but excluded from default LFS pulls (.lfsconfig — the ~905 MiB lens download is opt-in), so the audit reports each as an LFS pointer stub naming the pull command, and the two 1.5B nf4 lenses and their sidecars are regenerate-only pending the scheduled refit (issue #47), each reported as a loud MISSING rather than skipped — 3 stubs + 4 MISSING = 7. After git lfs pull --include="research/arcs/04_jspace/data/cache/**" --exclude="" the same run reports 1088 PASS | 4 FAIL (the nf4 MISSING reports only; measured 2026-09-24 with the cache pulled: 986 measured 2026-08-17, plus the 30 cache-independent CHECK O claims added 2026-08-30 and the 72 CHECK P claims added 2026-09-23, which read LFS artifacts pulled by default and plain committed logs, not the lens cache) with no GPU work; the check total grows from 1060 to 1092 between the two states because the lens-dependent blocks register their claims only when the lens tensors are on disk. Arc 02's 5 UNVERIFIABLE entries are printed, not scored — external citations and capture-time environment facts no committed artifact can settle. See the arc READMEs for what each audit does and does not catch (arithmetic consistency only — never methodology or interpretation) and for the historical pre-re-run figures (920 | 10; the cache-present 978 was never re-verified).

Capture and analysis scripts deserialize with torch.load(..., weights_only=False) on purpose — the artifacts are produced by these same scripts and never sourced externally. See the trust note in research/ARC_PROCESS.md § "Raw data is a deliverable" before extending the pipeline to third-party .pt files.

Raw data is a deliverable. A clean clone (with LFS pulled) holds the figure PNGs, the .pt datasets and the plain-git JSON/JSONL and text datasets the figures and audits depend on, so every figure can be re-rendered and every audit replayed. See research/ARC_PROCESS.md § "Raw data is a deliverable".

Building the theory LaTeX series (optional)

The theory side project's output is the 5-paper series under theory/series/, built by a shell script (not a Makefile — the only Makefile lives in the archived v1 snapshot at theory/archive/2026-05-03-pre-expansion/):

bash theory/series/build.sh          # clean + build all 5 papers + collect PDFs
bash theory/series/build.sh collect  # re-collect dist/ symlinks only (skip build)

Output PDFs land in theory/series/dist/<N>-<topic>.pdf. The build is sequential by necessity: each paper's main.tex declares cross-paper references via xr-hyper, so sibling main.aux files must exist first; the script runs two full sweeps so cross-refs settle. A LaTeX toolchain (pdflatex, bibtex) must be on PATH.

Epistemic discipline (carried over from the source workspace)

• Every technical claim a conclusion rests on cites a primary source by its paper-key in theory/kb/index/papers.json, often with an anchor into a KB excerpt or note. The excerpts are verbatim; the notes are AI-written digests that point to sources, and the cited paper is canonical.
• Analogies and intuitions are tagged ([ANALOGY], [INTUITION], [SPECULATION], [CONTRADICTION]), never asserted as fact.
• Forum/blog citations are discovery signals only; they never solely back a hard claim. Full rules in theory/README.md and CLAUDE.md.

License

GPL-3.0-only. (c) Michael Lannum. See LICENSE.