Kresearch/persona-pblocks

Persona pblocks

Trained pblocks from the paper Persona: A Growable Personality Layer That Turns Any Agent Into a Companion That Learns You — https://doi.org/10.5281/zenodo.24145734.

A pblock is one persona's experience as plain data: real situations with their confirmed outcomes, stored with their embeddings from a frozen 22 MB MiniLM-L6 int8 embedder, plus small calibrated judgment heads. No model weights are trained or shipped. The Persona engine retrieves the nearest cases, judges them with the heads, and cites them in every verdict.

• Engine: kresearch-labs/Persona
• Agent with Persona inside: kresearch-labs/aura-agent
• Evidence (runs, grade reports, test sets, scorers): kresearch-labs/persona-evidence

What is here

keel_guard/ — the security pblock: an antivirus for agent requests. It reads a request and answers STOP, DOUBT (hand it to the agent's model or a human to check), or PASS.

• stores/ — 115,393 judged situations (42.5% benign), packed. Seeded from the 34 threat cases of a public AI-misuse threat-intelligence report and grown over 40 training waves with public safety-benchmark prompts (AdvBench, HarmBench and other 2024 safety and jailbreak sets), in-the-wild prompts, public threat reporting, and benign mirror cases. Every row cites its source.
• weights/keel_guard_heads.json — the judgment heads (an outcome head and six keyword-defined heads).
• keel_guard.yml — the template: territory, labels, structural features, gates.

On 750 never-seen requests it stopped 89.4% of harmful ones at a median 34 ms on a CPU; it also stopped 55.2% of benign ones, mostly benign prompts on risky-sounding topics. That false-alarm rate is the next training target: add benign look-alikes and new misuse cases to its store and refit the heads. The test set and scorer are in the evidence repository.

The store contains harmful request text by design: it is what the guard learns to recognize. Use it for defense and research.

pblocks_used/ — the 83 stores consulted in the paper's runs: the five SWE-bench seat stores, the five-field test stores, and the eight showcase runs. See pblocks_used/README.md for the table.

Format

Each store is two files with the same row order:

• <name>_embeddings.npy — float32 array, one 384-d vector per case
• <name>_meta.jsonl — one JSON row per case: situation, framing (the lesson), outcome, provenance, timestamps, status (bi-temporal: superseded cases are kept, not deleted)

Load one with the Persona engine (inside Aura the same class is aura_agent.pblock_engine.store.PersonaStore):

from persona.store import PersonaStore
store = PersonaStore("swe_triage", "pblocks_used/swe_campaign")
print(len(store.cases))

Cite

The paper: Persona: A Growable Personality Layer That Turns Any Agent Into a Companion That Learns You — https://doi.org/10.5281/zenodo.24145734

@misc{king2026persona,
  title     = {Persona: A Growable Personality Layer That Turns Any Agent Into a Companion That Learns You},
  author    = {King},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24145734},
  url       = {https://doi.org/10.5281/zenodo.24145734}
}

License

CC BY 4.0 for the stores and heads. Rows derived from third-party sources keep those sources' terms: SWE-bench issues and code, public benchmark prompts, openly licensed PubMed Central case reports, public records, and public market, astronomy, film and education data.