netti-1, the person model

what any ai can know of one person, and what it is worth to them, scored in nats per decision. tested inside; it ships at launch.

what it is

a person model predicts one person: what they mean by what they ask, what they will choose, what changed about them and what was a one-off. it is built from the traces their life leaves in the tools they use, owned by them, and read by any model they allow.

white paper 01 describes netti-0, a first person model and the lab version of netti-1, and likeness, a benchmark whose people are programs, so every answer has a known true distribution. on 40 test people, scored once, netti-0’s small packet did better than the whole history in context while serving far fewer tokens. netti-1 ships at launch.

two guardrails failed and the paper says so: a forget left its words in a cache and in plain fingerprints, and the packet made the reader agree a little more often with a belief that was checkably wrong.

the evidence

against the whole history in context (H3)0.140 nats better (CI 0.122 to 0.157) at 84.6× fewer served tokens on the questions scored, 100.2× over every readdocs/research/person-models/key-numbers.test.json on labs/person-models (results-likeness-test-claude-haiku-5-5.json), via scripts/labs/import-paper.mjs
beyond a count of their own past choices (H1)0.040 nats per decision (CI 0.021 to 0.058)docs/research/person-models/key-numbers.test.json on labs/person-models (results-likeness-test-claude-haiku-5-5.json), via scripts/labs/import-paper.mjs
against retrieval (H2)0.160 nats better on decide and continue (CI 0.139 to 0.182)docs/research/person-models/key-numbers.test.json on labs/person-models (results-likeness-test-claude-haiku-5-5.json), via scripts/labs/import-paper.mjs
the field test, on what they will choose next0.255 nats a decision against 0.07 to 0.1 for memory built like mem0, zep and letta; better for 38 of 40 peopledocs/research/person-models/field (field-test.html, v-paired.json, v-forget.json, v-economy.json), branch labs/person-model, commit b92280c
the field test, on what changed about them0.81 nats against 0.14 to 0.4docs/research/person-models/field (field-test.html, v-paired.json, v-forget.json, v-economy.json), branch labs/person-model, commit b92280c
the field test, forgettingforgotten words handed back for 0 of 40 people, against 15 to 32docs/research/person-models/field (field-test.html, v-paired.json, v-forget.json, v-economy.json), branch labs/person-model, commit b92280c
what failedH4, H7, G3, G4: forgetting kept its own words, the packet bought a little agreementdocs/research/person-models/key-numbers.test.json on labs/person-models (results-likeness-test-claude-haiku-5-5.json), via scripts/labs/import-paper.mjs

40 synthetic people, read by claude-haiku-5-5: the people are programs, so nothing here is a claim about real humans
one world, built by the same authors as the model it scores
scored once, after a pre-registered freeze, which is why the failures stay in
the field test is after the reveal and outside the confirmatory chain; on plain facts about the person every other memory beat netti-1 except the letta-style blocks, which tied it; retrieval led
none of this is in innernet yet: netti-1 ships at launch, and until then the product keeps facts as sentences and nothing about choices

papers and notes