Evelyn Campbell · Independent research engineer

Making hidden dynamics inspectable.

0.0000

Max absolute logit identity error

Batch-size-1 identity checks · agent-reported

Source and limitations ↗

−0.290

Same − unrelated transfer efficiency

N = 32 · 95% CI [−0.363, −0.224] · unreplicated

Source and limitations ↗

0.37–0.64

Mean neighbourhood preservation

N = 4 scenes · no task comparator

Source and limitations ↗

Reported measurements, not three independently replicated discoveries. Missing sample counts, intervals or baselines appear in the evidence ledger and machine-readable records.

I build open research instruments for investigating transformer state trajectories, causal interventions, and the limitations of visualizing high-dimensional systems. Mottled is my flagship: a capture-and-viewer pipeline that makes model states explorable while reporting what each projection loses.

Enacted Volition is an independent, self-funded research practice. Code, checks, limitations, and contrary results are published alongside the work.

Explore a Mottled sample Examine the evidence Discuss a collaboration

Research threads

Each project has a status, the claim it currently supports, and a link to the code.

Active · primary

Mottled

Mechanistic interpretability · latent dynamics

Mottled records a transformer's residual stream after every block and plays it back in the browser, so you can see where in the layer stack two similar prompts split apart and where a model's state stops changing.

Logit-lens readouts, attention patterns, SAE features (each checked for fit to the model) and an exact attention/MLP decomposition all attach to one data type, StateTrajectory.

Sample scenes cover GPT-2, DistilGPT-2 and Qwen2.5-1.5B-Instruct; the portability study and the blast samples used Qwen2.5-0.5B-Instruct. Limits: the 2D projection is lossy, and every scene reports how much it kept.

PythonPyTorch TransformerLensSAEs WebGLStreamlit

Live explorer Self-portrait scene Source Portability study

Instrument readings

max |h[l+1] − (h[l]+attn+mlp)| = 0.0000: Sanity check: the decomposition reconciles exactly on Qwen2.5 as well as GPT-2.
42,257: Shared vocabulary entries used to put Qwen and GPT-2 on one comparable terrain.
266 tests · 24 files: Including cross-language conformance between the Python BVH and its JS port.
Portability study: An AI agent (Manus) used Mottled's unmodified capture code to run a preregistered transplant study on Qwen2.5-0.5B. Its prediction reversed. Read it.

Preprint on Zenodo, 2026-04-17 · not peer reviewed

Meth × neurodivergent reward architecture

Computational neuropharmacology

ADHD stimulants and methamphetamine act on the same dopamine machinery, with opposite clinical outcomes. This simulation links six layers to ask why: DAT reverse transport and VMAT2 packing, temperature-dependent oxidative toxicity, LC–NE gain control, trial-level reinforcement learning, sleep and circadian modulation, and setpoint drift over 365 days.

In the model, temperature is a bifurcation parameter for neurotoxicity, and the same exposure gives different trajectories for neurotypical, ADHD-C and ADHD-I parameter sets. It is a model, and it is not medical guidance.

The RL layer now uses ANCCR (Jeong et al. 2022) with IRI-scaling per Burke et al. 2026. The learning rate used to be a fitted constant; it now comes from the interval between rewards, so total conditioning time no longer depends on trial frequency.

NumPyODE systems ANCCRDose–response Sensitivity analysis

DOI 10.5281/zenodo.19625787 Source

Model structure

6 layers · ms → months: Terminal kinetics, thermal toxicity, LC–NE gain, RL, sleep/circadian, chronic adaptation.
α ∝ IRI: ANCCR revision: the learning rate is derived, so conditioning time stops depending on trial frequency.
3 architectures: Neurotypical, ADHD-combined and ADHD-inattentive parameterizations compared under identical exposure.
7 figures · OAT sensitivity: Every headline claim is reproduced by a script, with a one-at-a-time sweep showing which parameters carry it.

Active · protocol design

EVT-1

Agent-native trust infrastructure

A draft protocol for letting shopping agents check product claims themselves. Each product gets a deterministic ID (a canonical URN), and claims about it are signed with Ed25519, so an agent can verify them without trusting the store that showed them.

The server stores and serves signatures. Agents verify locally, so the server is not a trust root. An MCP server exposes the workflow as agent tools. Status: spec and reference server. Outside users: none visible on GitHub (0 stars, 0 forks, no issues from others as of 2026-10-09).

Python 3.12FastAPI Ed25519MCP Canonical JSONDocker

Source & protocol spec llms.txt

Protocol surface

urn:evt:product:{mfr}-{model}-{variant}: Deterministic identity — two agents canonicalizing the same product get the same URN.
6 MCP tools: resolve, trust profile, keygen, sign, submit, verify — the full loop from an agent.
Server ≠ trust root: Assertions are verified client-side; the service is storage and transport only.

Instruments

Everything here runs. The top rows are maintained. The archive rows are older browser tools kept for reference; they run fully in the tab, with no backend.

Instrument	What it does	Stack	Status
Mottled explorer	Captures and animates transformer hidden-state trajectories over a density terrain, with SAE features and logit-lens readouts.	Python · Streamlit	Live
Mottled viewer	WebGL viewer for `.mtj` scenes with real-ray BVH picking. Bundled GPT-2 and Qwen samples, no install.	WebGL · JS	Live
meth-neurodiv-model	Six-layer dopaminergic simulation with ANCCR reinforcement learning; regenerates every published figure from source.	Python · NumPy	Maintained
EVT-1 server	Canonical product resolution plus signed, locally verifiable trust assertions, exposed over HTTP and MCP.	FastAPI · MCP	Maintained
Portability study	Preregistered residual-state transplant experiment run on Mottled by an AI agent (Manus); the compatibility hypothesis reversed under every control.	Qwen2.5 · CounterFact	Live
Catalogue analysis	Audio-feature study of a fixed set of 155 generated tracks, analysed in August 2026. Shows which features carry information and which three are track length in disguise.	librosa · SVG	Live
GHOST_CHAIN	Client-side OSINT toolchain: domain and IP reconnaissance, hash analysis, metadata extraction. Zero backend.	Vanilla JS	Archive
WHISPER_KEY	LSB image steganography with AES-256-GCM and PBKDF2 (100k iterations) via the Web Crypto API.	Web Crypto · Canvas	Archive
NET_INTERCEPT	HTTP header and response-timing inspector built on the Fetch and Performance APIs.	Performance API	Archive
bundle-analyzer	Zero-dependency CLI that flags bundle bloat and duplicate function signatures against a size budget.	Node.js · CLI	Archive
lighthouse-enforcer	Validates Lighthouse reports against Core Web Vitals budgets and fails the pipeline on regression.	Node.js · CI	Archive

Evidence and limitations

Findings, computational checks, external-agent studies, and exploratory observations are distinguished here. Follow the links to examine scope, controls, and provenance.

Claim	Evidence status	Measurement, comparator and uncertainty	Source
Pasting a hidden state from an unrelated fact moves the answer more than one from a related fact. The preregistered prediction was the opposite.	single study	Qwen2.5-0.5B-Instruct, N = 32 directional transplants. Same − unrelated mid-band transfer efficiency = −0.290; pair-clustered 95% bootstrap interval [−0.363, −0.224], paired sign-flip p = 3.05×10⁻⁵. Designed and run by Manus AI; not independently replicated. An identity check reported a maximum logit error of 0.0000 to four decimals; this validates consistency, not generalization.	Portability study
Three of the six structural audio features in my catalogue analysis were track length in disguise.	verified computation	mean_seg_s vs duration r = 0.999. One of the three comes from a segmenter cap I set myself.	Catalogue analysis
In the meth × ADHD model, temperature is a bifurcation parameter for neurotoxicity. Risk does not scale in a straight line with it.	holds in the model	A simulation with parameters from published studies, compared with group-level findings. Not fit to individual human outcome data. Not medical guidance.	DOI 10.5281/zenodo.19625787
Mottled's attention + MLP split reconstructs the residual stream to reported precision (max error shown as 0.0000).	verified computation	A sanity check on the capture code. It is not a finding about models.	Mottled tests
Projection fidelity differs across saved scenes; predictive usefulness remains untested.	exploratory	Mean neighbourhood preservation across four sample scenes, from Mottled’s own fidelity_summary: 0.37 (self-portrait, GPT-2; 71% of states under the 0.5 flag) to 0.64 (gpt2-capitals). In three of the four, more than half the states are under the flag. Not yet shown: a task where Mottled beats a linear probe.	Sample scene

Evidence labels describe the checks reported here, not calibrated probabilities. Verified computations do not, by themselves, validate wider scientific interpretations.

How I work

Four rules I apply to both the interpretability code and the pharmacology code, each with an example.

Show what the picture loses

A 2D view of a high-dimensional state is lossy, so each view reports its own fidelity. Example: the catalogue analysis flags three features that turned out to be track length.

One interchange type

Capture code and renderers only meet at StateTrajectory. Example: GPT-2 and Qwen captures load in the same viewer.

Derive, don't fit

Prefer a mechanism to a tuned constant. Example: the RL learning rate now comes from the reward interval (ANCCR).

It must run offline

Synthetic backends let the pipeline run and be tested with no model download, API key or network.

Provenance and funding

Who pays for this and where AI was involved.

Funding	Self-funded. No affiliations.
AI involvement	AI-assisted research: I lead the projects, with AI systems assisting research, code, analysis, and adversarial critique. AI critique is not independent scientific peer review. The portability study was designed and run by Manus, an AI agent, against Mottled's unmodified code. The 155 analysed catalogue tracks are made with Suno (121 of the 146 distinct titles match public tracks on suno.com/@evevev; the rest are not public). Live total on Suno: see suno.com/@evevev.
Reversals	The portability study's registered prediction came out backwards. The catalogue analysis dropped three of its six structural features as track-length proxies.

Contact

For collaborations, contract research, or private inquiries, contact me by email. For reproducibility questions, bug reports, or disagreements about published results, open a public GitHub issue. The form below prepares an issue; it does not send private messages.

Private inquiries and collaborations: robolution789@gmail.com
For scoped technical work, research collaborations, and inquiries that should not be public. This address is already published on my portfolio.
Technical issues: github.com/enactedvolition · Issues
Best for anything about a specific instrument. Public thread, so the answer helps the next person too.
Profile: github.com/enactedvolition
Source for every project listed above, and the fastest way to see what is moving.
Published work: doi.org/10.5281/zenodo.19625787
Cite the DOI rather than this page for the neuropharmacology model.
Jurisdiction: Operated by an individual, remote. Terms of use are on the terms page; what the site stores is on the privacy page.