You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$18,350 left to reach the funding goal.
Mottled (Apache-2.0) captures a language model's residual stream layer by layer and draws it in 2D. Every 2D picture of that space distorts it; Mottled measures how much and prints it on the picture: per-state neighbourhood preservation, a bootstrap error field on the density terrain, and an amber ✕ on every state whose local structure didn't survive. Scenes carry a provenance record. Try it with no install:
https://enactedvolition.github.io/mottled/
— and the honesty demo, where Qwen2.5-0.5B-Instruct shifts 0.44 nats toward a user's wrong answer after pushback:
https://enactedvolition.github.io/mottled/viewer/?file=samples/music-qa-honesty/item-1.mtj
Why it matters for safety: white-box audits rely on projections and SAEs that fail silently. Mottled has already caught one: the same public GPT-2 SAE reads ~24% reconstruction error on TransformerLens activations and ~342% on raw HuggingFace hidden states. It's a map you read before and alongside causal tools, not a replacement.
1. Projection-independent views (token strip, divergence curve, answer race) beside every 2D scene. On Qwen, 90–98% of states fall below the fidelity cut on almost every run, so readings shouldn't depend on the 2D picture. (Funded by the minimum.)
2. Run the dose-sweep validation (code exists; spec written) with random-direction, shuffled-label and direct-path controls on two models, and publish nulls.
3. Re-run the honesty pack (sycophancy, fabricated citation, noisy-input calibration, memorization) on Qwen2.5-1.5B and 7B-Instruct with fidelity records, plus a short public write-up.
Success = steps shipped in the repo, and at least three outside people have run it and told me where it's wrong.
It ships: the projection-independent views are live in the viewer next to every 2D scene, and the dose-sweep validation and the 1.5B/7B honesty re-runs are committed to the repo with their fidelity records.
The controls behave: random-direction and shuffled-label probes come out near zero. If they don't, the readouts are artifacts and I'll say so publicly.
Nulls get published next to positives, in the repo and the write-up.
At least three people other than me run Mottled and tell me where it's wrong, through issues, PRs, or messages.
Minimum $3,050 (~2 weeks): $2,625 stipend (75 h × $35/h), $371 self-employment tax, $50 GPU, $10 API. Delivers goal 1.
Target $18,350 (3 months): $15,750 stipend (450 h × $35/h), $2,225 self-employment tax, $300 GPU (~150 GPU-hours on Modal), $75 API credits. Delivers goals 1–3.
Hosting is free (GitHub Pages). Anything between the two buys proportionally more time. If other applications also come through, extra money becomes more months, never a second stipend for the same month.
Just me, Evelyn Campbell, in Olympia, WA. I'm a self-taught independent researcher. I build Mottled, an open-source instrument for watching how language models change their predictions layer by layer, and I run public research on whether agency can emerge in cellular automata. I work solo, with heavy AI assistance, and I publish everything openly, including null results.
Track record is the repo: started July 11, 2026; 139 commits; 56 test files and CI; parity checks against HF, TransformerLens and NNsight; SAE fit reports; a live WebGL viewer that can capture small models in the browser; a five-item honesty demo published with its weaknesses stated. I built it with heavy use of an AI coding agent (63 of the 139 commits are authored by it); I set the design, check the numbers and decide what ships.
Most likely: nobody adopts it — researchers stick with notebooks and TransformerLens. Second: the projection-independent views carry all the signal and the 2D picture adds little. Third: the controlled dose sweeps come back null at these model sizes. In each case the outcome is public and still useful: an open tool, a documented SAE pitfall, and published nulls.
$0 raised in the last 12 months. BlueDot Rapid declined Oct 11, 2026. Submitted and pending: BlueDot Seed Grant (submitted Nov 4, 2026) and AE Studio CRUCIBLE (submitted Nov 13, 2026, a $1,000 sprint stipend, different project scope). Extra money = more months, never two stipends for one month.