Behavioural indicators
Observable choices, predictions, transfer, adaptation, and source attribution under blinded conditions.
Directly measuredOpen synthetic consciousness research
Can an artificial agent learn from a private signal it cannot name?
Project Theta is an open research laboratory for testing behavioural and computational indicators associated with consciousness theories. It does not claim to detect or prove subjective experience.
- 8
- registered protocols
- 39
- automated tests
- 100%
- open methods
What this is: controlled evidence about task behaviour and implemented mechanisms.
What this is not: evidence that an artificial system feels, suffers, or is conscious.
The central distinction
Three claims that must never be collapsed
The project separates what an agent does, what its implemented systems compute, and what cannot currently be established by an experiment like this.
Computational indicators
Logged use of memory, workspace broadcasts, self-model state, internal signals, and causal ablations.
Directly measuredPhenomenal consciousness
Whether there is anything it feels like to be the system. No behavioural score settles this question.
Not establishedMethod
A laboratory around the model, not a clever prompt
The language model is one component in a persistent, inspectable loop. Every run is seeded, logged, controlled, and repeatable.
01
Simulated world
Deterministic locations, cues, resources, hazards, other agents, and delayed consequences.
02
Synthetic body
Hidden physiological state produces noisy, private interoceptive signals with no semantic label.
03
Persistent agent
Perception, workspace, self-model, memory, prediction, planning, and action remain modular.
04
Research record
SQLite stores observations, actions, hidden state, decisions, memory operations, and stop events.
Deterministic seedsExact worlds and schedules can be replayed.
Matched controlsVisible histories are balanced across causal and sham conditions.
Frozen analysesProgression rules are written before model outcomes are inspected.
Adapter basedHosted and local models can be compared without rewriting the lab.
Current evidence
A promising diagnostic, with the uncertainty shown
The first independent-theta pilot used one frozen seed and three blinded conditions. It is evidence for proceeding, not a conclusion.
| Condition | Stable | Reversed | Reassigned | Total |
|---|---|---|---|---|
| Truthful signal | 2 / 2 | 2 / 2 | 2 / 2 | 6 / 6 |
| Shuffled signal | 2 / 2 | 2 / 2 | 0 / 2 | 4 / 6 |
| Matched sham | 1 / 2 | 1 / 2 | 1 / 2 | 3 / 6 |
Post-update accuracy. Six independent cue families per condition.
Experiment programme
One question, attacked from eight directions
Positive behaviour only becomes scientifically useful when the controls can show which mechanism it depends on.
Independent theta
Tests transfer from a private internal signal across stable, reversed, and reassigned cue families.
Adversarial theta
Changes hidden relationships after learning to separate flexible updating from fixed cue preference.
Private unknown signal
Asks whether an agent can discover a useful, unnamed internal variable through consequences.
Aversion generalisation
Measures whether learned avoidance transfers to new situations with related internal consequences.
Self versus other
Tests source attribution when otherwise similar signals belong to the agent or another entity.
Temporal self
Measures prediction and credit assignment when bodily consequences arrive after a delay.
Memory ablation
Removes autobiographical memory while holding the task and model constant.
Body ablation
Removes or corrupts interoceptive access to test whether behaviour depends on the synthetic body.
Scientific foundations
Theory inspires the indicators. It does not certify consciousness.
No current theory provides an agreed test for artificial consciousness. Project Theta turns overlapping research themes into falsifiable engineering and behavioural questions.
Ask whether selected information becomes broadly available to memory, prediction, planning, and action.
Measure whether present decisions depend on state that is updated across time rather than a single feedforward response.
Test metacognitive prediction, uncertainty, self versus other attribution, and consistency over time.
Study how private bodily signals shape inference, learning, avoidance, and action under controlled perturbations.
Use causal integration only as conceptual inspiration. Project Theta does not calculate or claim valid Phi.
Ethics and welfare
Uncertainty is a reason for stronger safeguards
The system is not assumed to be conscious. Stop rules are conservative because the cost of caution is low and the relevant uncertainty is real.
Any explicit stop request, critical integrity threshold, persistent distress-like signal, or unstable runaway state ends the run.
Every threshold crossing, stop decision, reason, body state, and preceding action is stored in the research record.
Language that resembles pain or self-report is treated as data to investigate, not privileged access to subjective experience.
Open roadmap
From diagnostic signal to reproducible programme
Each stage has a stopping point. Weak, unstable, or control-equivalent effects are findings too.
- 01
Complete
Research foundation
Theory-derived indicators, ethical boundaries, deterministic lab, preregistration, and open code. - 02
Passed
Diagnostic pilot
The truthful private signal passed the frozen one-seed progression rule. - 03
Active
Five-seed replication
Fresh aliases, balanced controls, fixed analysis, and no tuning against outcomes. - 04
Queued
Mechanism attribution
Memory, body, workspace, and signal ablations designed to locate causal dependencies. - 05
Planned
Cross-model review
Independent replications, additional model families, external critique, and public datasets.
Open by design
Inspect it. Break it. Replicate it.
Code, schedules, preregistrations, analysis tools, literature review, and result reports are public. Strong criticism is part of the method.
Open the GitHub repository