Skip to content
Project Synapse · v2.0 · Phase 2 running

Building the Future of Human × AI Collaboration

Project Synapse is an ongoing research initiative focused on understanding how humans and AI communicate, collaborate, and learn from each other through real-world conversations.

22,166+
Conversations analysed
multi-turn transcripts
1,064+
Context variations
prompt/context pairs
84%
Response quality
human-rated acceptance
222+
Feedback sessions
structured reviews
89+
Prompt experiments
controlled ablations
v2.0
Current version
Synapse release
About Project Synapse

Human × AI Conversation Intelligence Platform

Synapse instruments real conversations end to end — capture, context ablation, blind scoring and human review — so every claim about model behaviour traces back to a scored transcript.

01

Grounded in transcripts

No synthetic benchmarks. Every metric derives from consented, real multi-turn conversations.

02

Blind evaluation

Reviewers never see which system produced an answer, and rubric weights are fixed before scoring.

03

Replicable by design

Harness, rubrics and anonymised slices are published so external teams can reproduce results.

Research areas

Four questions we keep measuring

01

Conversation Intelligence

How meaning survives across multi-turn dialogue, and where it degrades.

multi-turn coherence
02

Context Engineering

Measuring which context actually improves answers versus which only inflates tokens.

retrieval budgeting
03

Human Feedback Loops

Structured review rubrics that turn subjective preference into comparable signal.

rubrics annotation
04

Evaluation Harnesses

Blind scoring pipelines that keep model comparisons honest and reproducible.

blind eval replication
How it works

Capture, ablate, score, publish

STEP 1

Capture

Consented transcripts are normalised and stripped of identifying detail.

STEP 2

Ablate

Context slices are removed and restored to isolate what actually helps.

STEP 3

Score

Blind reviewers apply a fixed rubric; agreement is tracked per reviewer.

STEP 4

Publish

Findings, harness and data slices ship together as research notes.

Programme phases

The Synapse roadmap

Phase 1 · Complete

Pilot corpus

Collect and clean a multi-domain conversation corpus with consented transcripts.

Phase 2 · Running

Context ablations

Systematically remove and restore context to isolate what drives answer quality.

Phase 3 · Queued

Feedback modelling

Model reviewer agreement and calibrate rubric weights against outcomes.

Phase 4 · Planned

Open replication

Publish harness, rubrics and anonymised slices for external replication.

Dashboard

Everything Synapse measures, in one surface

Scored conversations per month

Feb Mar Apr May Jun Jul Aug
Appreciation

Built with reviewers, annotators and open collaborators

Project Synapse runs on the patience of the people who score transcripts line by line. If you would like to contribute or replicate a result, we would like to hear from you.

Start a conversation