Behavioural indicators
Observable choices, predictions, transfer, adaptation, and source attribution.
Directly measuredOpen synthetic consciousness research
Can an artificial agent learn from a private signal it cannot name?
Project Theta builds controlled experiments around artificial agents. We test behavioural and computational indicators inspired by consciousness research. We do not claim to detect subjective experience.
- 85.7%
- regulation accuracy
- 24/24
- body families passed
- 1,056
- fixed model decisions
What we measured: choices and wrapper computations in controlled tasks.
What we did not measure: feeling, sentience, suffering, or phenomenal consciousness.
The research record
Results, failures, and repairs belong in the same story
The record includes the initial active-control result, the reliability failure that followed, and the completed multi-body study that resolved it.
New reliability result
The body-learning result held across 24 fresh synthetic bodies
V10 completed six fixed seeds and 1,056 model decisions. Every frozen progression rule passed.
Read the result and its limitsV8 replication resultCan an AI regulate a private synthetic body signal?
The intact system beat three ablated controls across twelve paired seeds under a frozen plan.
Read the experiment and its limitsMethods correctionWhy we stopped the strongest-looking earlier experiment
The shortcut, the information mismatch, and the redesign that followed remain part of the public record.
Read the correctionThe central distinction
Three claims that must never be collapsed
A task score, an implemented computation, and subjective experience are different kinds of claim.
Computational indicators
Logged memory, workspace broadcasts, binding tables, private signals, and wrapper interventions.
Directly measuredPhenomenal consciousness
Whether there is anything it feels like to be the system. No score in this laboratory establishes that.
Not establishedThe active experiment
The agent acts, predicts, observes, and regulates
V8 replaces static answer tables with consequences caused by the model's own choices in a synthetic body.
01
Private signal
The agent sees a noisy internal channel with no meaningful name or explanation.
02
Chosen action
It applies one of two opaque actuators and predicts the immediate result.
03
Body consequence
The selected action genuinely raises or lowers the hidden synthetic body state.
04
Regulation
It must use what it learned to move a perturbed signal back towards its target.
12 fresh seedsNo replication seed was used in development.
Four conditionsFull, shuffled interoception, no memory, and no body.
Exact order balanceEvery condition occupied every run position three times.
Exact-model recordAll 1,152 calls reported GPT OSS 20B.
Internal replication
Useful body information changed what the system could do
The primary result repeated across twelve fresh paired synthetic bodies under a frozen plan.
| Condition | What remained | Regulation accuracy |
|---|---|---|
| Full system | Truthful body signal and memory | 0.958 |
| Shuffled interoception | Unreliable private feedback | 0.507 |
| No memory | Current signal without stored outcomes | 0.465 |
| No body | Choice task without private state | 0.493 |
Twelve paired seeds. Forty-eight runs. All frozen audits passed.
Scientific foundations
Theory inspires the indicators. It does not certify consciousness.
The project turns themes from global workspace, recurrent processing, self-modelling, prediction, and interoception into falsifiable questions.
V8 found a 0.451 mean full-system advantage over shuffled interoception across twelve paired seeds.
V10 achieved 0.857 regulation across 24 independently mapped body families, with every frozen rule passed.
V11 uses matched causal interventions to identify which information channels produce the successful control.
Cross-model replication, independent scrutiny, and open-weight mechanistic analysis.
Ethics and welfare
Uncertainty is a reason for stronger safeguards
The system is not assumed to be conscious. Conservative stop rules remain active even in reduced-call runs.
An explicit stop request, critical integrity threshold, persistent distress-like signal, or unstable state ends a run.
Probe-only execution removes unnecessary inference, not welfare monitoring or state logging.
Self-reports and pain-like language are behaviour to investigate, not privileged evidence of experience.
Open roadmap
From an internal result to a credible research claim
The next gain comes from identifying the information channels that cause the reliable result, then exposing that mechanism to another model and outside scrutiny.
- 01
Complete
Build the laboratory
Deterministic simulations, synthetic body signals, memory, workspace, self-model interfaces, logging, stop rules, and open code. - 02
Corrected
Find and publish the flaw
A review found an answer-position shortcut and unequal information access in the strongest self-model result. - 03
Passed
Build an active body task
V8 makes the model choose actions that cause changes in an opaque private body signal. - 04
Replicated
Test twelve fresh bodies
The intact system beat all three controls on every paired seed under a frozen plan. - 05
Complete
Repair the neutral interface
V9.1 achieved perfect interface comprehension, then failed its frozen reliability rules on a fresh body. - 06
Passed
Test multi-body reliability
V10 completed six fixed seeds and 24 independently mapped bodies. Every frozen progression rule passed. - 07
Registered
Test the mechanism
V11 applies matched causal interventions to current-state sensing, learned summaries, transfer identity, and mapping access. - 08
Planned
Replicate on another model
A successful mechanism result will move to cross-model replication and outside scrutiny.
Open by design
Inspect it. Break it. Replicate it.
The code, correction, frozen historical record, public evidence export, repaired protocol, and reviewer packet are open.
Open the GitHub repository