You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Project description
exist. is a governance layer for generative AI. Generative models optimise for continuation; they have no mechanism for refusal, so identity erodes through accumulation — tone shifts, voice flattens, meaning thins — and nothing visibly breaks. exist. evaluates each output against a defined identity and rejects what drifts. I've validated the measurement on one real case; this project runs it across many people to test whether structured refusal protects identity, or quietly becomes a form of control.
Three layers: free generation; an LLM-as-judge evaluation against an explicit identity constitution (Approve / Flag / Reject) — scoring tone consistency, value alignment, structural coherence, and commercial deviation; and drift monitoring over time.
On a single anonymised corpus (six months of longitudinal text), three independent models — GPT, Claude, Gemini — evaluated the same identity against identical constraints. They converged:
94% structural agreement between independent forensic passes
Identity stability: 74–82 / 100
Commercial drift: ~35%, contained — full convergence between models
Core components (drift control, value recurrence): all models agree (stable)
Contradictory conclusions between models: 0
Classification: structurally stable, with controlled commercial elasticity
The point isn't agreement for its own sake — where the models diverge, the divergence is the diagnostic.
Take it from one identity to a small cohort (8–10 people across different disciplines). For each: compile an identity constitution from their existing work, run their output through exist., track drift, and interview them about what the refusals felt like. Deliverable: a public write-up answering whether structured refusal preserves identity or flattens it — and where the line sits.
Why this is different. Most identity/drift work operates at the model level. exist. operates on the individual human's authorship and voice — refusal as governance for the creator, not just the model.
Budget
Minimum ($6,000) — lean pilot: 5–6 participants, manual orchestration (as today), multi-model API costs, and my time. Produces the public write-up at small scale and proves the method generalises beyond n = 1.
Full goal ($24,000):
Engineering — semi-automate the evaluation pipeline (contractor): $9,000
Multi-model API / compute (3 models × ~10 corpora): $4,000
Participant honoraria (10 × $250): $2,500
Researcher time (~4 months, part-time): $7,000
Write-up, documentation, public report: $1,500
Marian Dorobanțu — independent operator and writer. For eighteen years I built artist identities in the Romanian music industry through ATOM (~1B streams, ~8% national radio share, built without paid media): constructing and protecting voice at scale was the job. I designed exist. and ran the multi-model validation above. I'm not an ML researcher — engineering will be contracted. Independent, non-institutional work.
Honest failure modes: LLM-as-judge proves too biased or unreliable to measure identity consistently across people; refusal flattens voice rather than protecting it; or the metric doesn't generalise beyond one person. Even then the output is informative — a public negative result on whether identity can be tracked and governed, plus a documented account of where the approach breaks. The real risk is that refusal turns out to be a form of control. Surfacing that is part of the point.
None for exist. Self-funded to date; this would be its first external support.
No offers yet — be the first to back this project.
You're pledging to donate if the project hits its minimum and gets approved. If not, your funds are returned.
No comments yet. Sign in to create one!