Skip to content
Affective
01

Research-stage · Fine-tuning an EI model

Find where AI loses the human context.

Affective is fine-tuning an EI model that chooses the right conversational decision, with person state as an external, person-editable memory component. Transcript evaluations of how agents mishandle people are the first state add-on.

See current evidence
POC underwayNo public API or live evaluation service yet.
Evaluation specimenConcept, not a result
01

Trace the moment

Find the exact turn where context was lost or a response missed a change in need.

02

Inspect the evidence

Record the observed signal, comparator, and uncertainty instead of assigning a hidden emotion score.

03

Test the decision

Ask whether structured state adds value beyond a strong transcript or summary baseline.

Ink study of Eros's residual-stream architecture: a shaft, a helical affect filter, and three instrument taps. Pre-POC. No live model state.
Residual-stream study. Pre-POC. No live model state.
02

The first commercial question

Before an agent remembers people, test where it fails them.

On the v2 sequence, the API's first offer is the fine-tuned EI model. The first state add-on is evaluation of existing agent transcripts, the part we can discuss in detail today. The report format is designed to show a specific failing turn, the method used to identify it, an appropriate comparator, and what remains uncertain. A production Observe or Runtime service is a later gate.

01 / INPUT

An authorized conversation

Data rights and privacy terms come before any transcript review.

02 / REVIEW

Inspectable mishandles

Missed escalation, lost state after a topic shift, and inappropriate agreement are initial evaluation targets.

03 / OUTPUT

Evidence, not a magic score

The planned deliverable is an auditable trace with limits, not an unverified emotional diagnosis.

Ratchet study: emotional context advances turn by turn and is held by the pawl, carried across turns rather than re-derived.PAWLT1T2T3T4T5T6SIDE ELEVATIONPERSISTENT ACROSS TURNSPRE-POC
Ratchet study. Context should be carried from turn to turn, not re-derived. Illustration.
How an evaluation would work
03

Where this is headed

Emotionally intelligent AI, in five gated stages.

The long-term product is AI that keeps track of the person it is talking to, reached through an API and the Affective platform. Nothing is live yet. Each stage opens only if the evidence supports it.

  1. 01

    EI model

    Research now

    A fine-tuned open-weight chat model that infers what a person feels, how strongly, and why, chooses a response strategy, and replies truthfully. Served through an OpenAI-compatible API with strategy and appraisal readouts.

  2. 02

    Memory

    After the model validates

    Persistent, person-editable memory per end user, used by the model. The person can see and delete it.

  3. 03

    State add-ons

    Each behind its own gate

    Evals (mishandle suites), Observe (production scoring), Runtime (state for any model), and Scenario Data, sold on the same API.

  4. 04

    Voice

    Separate track, own gates

    Affective Voice: our own realistic voice models. V1 half-cascade first; speech-to-speech only if it measurably wins.

  5. 05

    Foundation model

    Series A plan

    The from-scratch EI-native foundation model behind the same endpoints.

See the full product direction
04

Evidence, with its limits

A result earns a claim only at its own gate. We report synthetic, development, and real-person evidence separately, including negative results.

Matched-pair balance: a loaded scenario and its neutral control, same domain, characters and length, hang level. Only one line differs.LLOADEDLCONTROLMATCHED PAIRSAME DOMAIN · SAME CHARACTERS · SAME LENGTH
Matched-pair study. Every comparison is read against a matched control.
M0 ArenaPassed (conditional)A sealed, person-disjoint longitudinal benchmark with real people and larger-model controls. Track-P gate evidence is incomplete.
M1 Kill testStopped the state thesisA shuffled-order placebo kept 85.6% of the ledger's advantage (limit 50%). State stays a memory component; amendment A24 re-scoped the program to the fine-tuned EI model.
M2 EI arena v2ActiveSealed capability-spec suites with human labels and the standard controls. The E52 preregistration is frozen; no human labels exist yet.
EI modelNot yet evaluatedNo fine-tuned EI chat model has been evaluated against the capability spec. Base model and training data come next.

Snapshot: October 25, 2026. Read the evidence map

05

Research track: Voice

Speech is evidence. Core decides.

Affective Voice is a separate research track building the V0 cascade toward a V1 half-cascade, with an offline path for transcript and acoustic observations, plus an initial TTS adapter. Speech signals carry quality, uncertainty, provenance, and consent limits. They are not a direct reading of someone’s private emotion.

Read the voice research

Current boundary

Offline observation is implemented.

Offline only

Validated production speech-emotion inference, Core integration, live streaming, and release remain future gates.

07

Join the team

Chief Marketing Officer (CMO)

We are looking for a marketing leader who can make a rigorous, research-stage story legible to AI teams without outrunning the evidence.

Review the role · founder@affective-llc.site