Research preview · UK-hosted

Affective AI you can read.

Serena is a transparent affective AI from Beyond the Box. Every response is produced by symbolic operations on the input — not a black-box neural model — so you can see, exactly, how it arrived.

Try a demoHow it works
Studio
Beyond the Box
Division
GlassMind
Author
Blade Harvell
Supervisor
Dr. Deniz Cetinkaya
Region
europe-west2
Status
Research preview

§ 01Principles

What we're committed to.

Concrete operational commitments — not aspirations. The questions every AI company should be able to answer plainly, answered plainly.

  1. § 01

    No black boxes.

    Every response Serena produces is the output of inspectable symbolic operations. There is no neural network in the affective inference path, and no part of the answer that can't be traced back to its inputs.

  2. § 02

    Audit-trail by design.

    Each call returns a trace ID. Reopen it later — see every symbol Serena considered, every weight it applied, and the exact path from input to vector to response.

  3. § 03

    We don't train on your data.

    Serena's lexicon, n-gram model, and lifespan curves are fixed and documented. Your inputs are stored against your trace ID for inspection and deletion — never used to update the model.

  4. § 04

    Open research.

    The architecture is documented end-to-end across two academic papers. Methods, weights, and limitations are all public. Disagreement is part of the design.

  5. § 05

    Deletable on request.

    Email us with a trace ID and we'll delete the input and result within seven days. No retention loopholes, no "legitimate interest" carve-outs.

  6. § 06

    UK-hosted, EU-friendly.

    All inference and storage runs in europe-west2. Cloud Run scales to zero when idle. Data does not leave the region.

§ 02What it does

Three calm windows into the same input.

Serena decomposes language into three orthogonal vectors. Each is inspectable. Each comes with its own trace.

7D affect

non-semantic structural

12D emotion

lexical · phonemic · syntactic

disgust
0.96
pride
0.87
anticipation
0.66
joy
0.56
sadness
0.53
shame
0.42
fear
0.41
love
0.39
surprise
0.37
anger
0.30
trust
0.24
guilt
0.08

Lifespan curves

0–25y developmental priors

0y7y25y

§ 03How it works

From tokens to feeling, in three traceable stages.

No transformer. No attention block. Symbolic operations on tokens, phonemes, lexicon weights, and lifespan-scaled priors. Every step is documented.

01 / tokens

Read the surface.

Tokenize, POS-tag, map words to ARPAbet phonemes. Strip nothing. Pure structure, before any meaning is assigned.

> "i think i'm okay"
tokens → [i, think, i, m, okay]
pos    → [PRP, VBP, PRP, VBP, JJ]
phon   → [AY, TH IH NG K, AY, AH M, OW K EY]
02 / symbols

Assemble the trace.

Look up PMI lexicon weights. Run Kneser-Ney n-grams. Bind symbols to nodes. The graph is a record, not a guess.

lex(okay) → val −0.12, aro 0.04
ngram(i'm okay) → conf 0.41
phonΔ(soft fricative) → aro −0.08
trace_id 0x7af2 · 17 edges
03 / affect

Project into feeling.

The 12D emotion vector and 7D affective vector are produced by inspectable arithmetic on the trace — not regression weights, not attention heads.

EVec  → joy 0.08, sad 0.42, fear 0.31 ...
7D    → val −0.32, aro 0.58, dom 0.19,
        prd 0.71, nov 0.22, com 0.06, srf 0.44

§ 04Research

Two papers. Read both.

Serena's architecture is documented end-to-end across two academic papers — the dissertation that grounded the symbolic core, and the lifespan paper that gave the system a developmental timeline.

Dissertation2025

A Symbolic Affective Architecture

12D emotion vectors via PMI lexicons, Kneser-Ney n-grams, ARPAbet phonemic mappings, LMDB short-term memory, MPHF long-term memory, and template-based response generation.

Blade Harvell · supervised by Dr. Deniz Cetinkaya

Read paper →
Lifespan paper2025

Lifespan-Aware Affective Computing

A 7D affective vector (val, aro, dom, prd, nov, com, srf) derived from non-semantic structural features. 6D DNA traits. Developmental priors from birth to 25 years. Newborn blending. Homeostatic integration.

Blade Harvell · GlassMind Division, BTB

Read paper →

§ 05Try a demo

Paste a sentence. See the trace.

Open the demo to paste a sentence and watch Serena decompose it into a 7D affect vector, a 12D emotion breakdown, and a trace ID. No signup, no logging against you. While the production model is being deployed, the demo runs against a deterministic local stub — the UX is real, the numbers are not.

§ 06Team

Built by a small group of people who read papers.

Founder & architect

Blade Harvell

BSc (Hons) Software Engineering, Bournemouth University, 2025. Author of both Serena papers. Designed and implemented the symbolic affective core, the 7D lifespan model, and the white-box trace pipeline.

Dissertation supervisor

Dr. Deniz Cetinkaya

Supervised the original Serena dissertation at Bournemouth University. Provided the academic frame for symbolic affective architectures and review of the 12D emotion vector method.

Studio of record

Beyond the Box · GlassMind

Beyond the Box is the studio behind Serena. The GlassMind Division is its research thread for transparent intelligence — symbolic-first, audit-trail-by-default systems.