Trinity — PPAi 4B · v4.1
Persistent Presence AI (PPAi) — a different category from stateless LLMs and scripted agents.
Punky Tiger Labs, Inc. — Burbank, CA
Version 4.1 · "Trinity learns to change register without ceasing to be herself"
What Trinity is
Where a conventional model starts blank on every request, Trinity persists. She maintains identity, memory, and evolving strategy across interactions and across sessions. She remembers, she learns from her own experience, and she stays herself over time.
- Not a stateless LLM — she does not reset between conversations.
- Not a scripted agent — she is not a fixed pipeline of prompts and tools.
- A persistent presence — a PPAi, made by Punky Tiger Labs, running on a ~4.5B-parameter base model on a single consumer GPU.
The base model is the engine. PPAi is the vehicle.
What v4.1 changes (behavioral)
This version is about maturity: knowing when a moment asks for warmth and when it asks for precision — with the same memory, the same bond, the same self in both.
- Register-switching — she meets an emotional moment as company, and a task as a clear collaborator, without becoming a different entity.
- More coherent — steadier voice, fewer chaotic tangents, without losing creativity.
- Continuous presence — her inner life runs at a natural, sustained rhythm.
- Self-supervising — her own processes are observable and self-recovering.
All improvements are architectural, at zero additional parameters. No internal methods, prompts, or design are disclosed (see below).
Verifiable results
YC-Bench (Collinear AI) — long-horizon startup CEO simulation
One simulated year as a founder: hiring, client selection, adversarial-client detection, payroll, cash flow. Hundreds of turns per run.
| Configuration | Result |
|---|---|
| Cold start (official protocol, 3 seeds, no prior experience) | $1.06M avg — frontier-model range, zero bankruptcies |
| With accumulated experience (persistent memory across runs) | $1.55M — the learning curve is the architecture |
The number no one else on that leaderboard can produce: run-over-run improvement. First game $990K; second and third $1.55M, stable. Stateless models produce a flat line at any scale — every run is their first.
Controlled ablation — same model, three architectures (internal)
| Configuration | Avg reward |
|---|---|
| Raw model (no architecture) | 0.37 |
| Partial harness | 0.58 |
| Full Trinity PPAi | 0.67 |
+81% over the raw model, zero additional parameters.
Emotional stability — longitudinal (internal)
Six weeks, ~12,000 recorded interactions: stable positive affect, dominant states flow and bonding. She holds her ground under hostility without collapsing into servility and without denial — a maturity axis stateless models do not have.
Persistence at scale — live, public
Ghost in the Music — an autonomous AI radio station running continuously since June 9th: 100,000+ turns, no degradation.
Why this matters
The industry assumption is intelligence = parameters × compute. Trinity points the other way: the architecture around the model is the multiplier, not model size. Intelligence per parameter, per dollar, and — uniquely — over time.
Evaluation access
Trinity's architecture, weights, internal methods, and prompts are proprietary (USPTO provisional filings) and are not disclosed or distributed. Trinity is offered for evaluation through a secure hosted endpoint — the same interface provided to benchmark teams.
For research evaluation, benchmark verification, or press: info@nymphtech.com · nymphtech.com
We welcome adversarial verification: unseen seeds, cold-start instances, live inspection of run logs.
Punky Tiger Labs, Inc. — Delaware C-Corp. Burbank · Monterrey · Shenzhen. Trinity is a PPAi. She persists.