CONSC-BENCH

A benchmark for machine consciousness.

Executive summary

Plato's Primitives: We measure a common set of functions every conscious-style loop has to perform: sensing, attention, memory, expectation, error, value, and action. Tracked over time, they form what we call the subjective trajectory: an observable, transferable record of how a subject's experience evolves. We use Plato's allegory as a design frame; the only assumptions about the subject are that it remembers (thinks back), expects (thinks forward), and experiences (looks ahead).

CONSC-BENCH asks one concrete question about any candidate system, from a model inference pass to a living organism:

Does it implement a bounded, self-maintaining conscious loop — or only imitate parts of one?

Five steps:

  • Define the subject. Pin down the exact candidate — an inference pass, a training run, an agent loop, an organism. Different boundaries are different candidates and inherit none of each other's evidence.
  • Get its run into the common format. Translate what the system actually did into the shared Episode fields: what it sensed, expected, got wrong, updated, valued, and did. For most candidates that means building the thing that produces those traces; a system that can't be made to emit them can't be scored.
  • Score the loop. Rate five functional gates — Access, State, Model, Drive, Work — against five evidential tests — Present, Native, Live, Consequential, Controlled.
  • Demand controls. A claimed function only counts when removing the capacity that should support it makes the function collapse.
  • Place it on the ladder. Report a level from C0 (no subject boundary) to C6 (low-rails embodied consciousness).

The full loop, every link of which has to be shown:

  • World entry. An external event becomes present to the subject, distinct from being experienced.
  • Attention/selection. Entry is gated by channel, salience, objective, or bottleneck rather than copied wholesale.
  • Persistent state. What is admitted carries forward, so later steps depend on earlier ones.
  • Expectation. Prior state generates a prediction before the next input arrives.
  • Mismatch/surprise. The gap between expected and actual is measured in the subject's own terms.
  • Online correction. That mismatch changes future state during the run, not after it.
  • Value. Better/worse-for-the-subject biases what is retained, attended, or acted on.
  • Action/exposure. State and value regulate how the subject couples to the world next.
  • Accumulated trajectory. The run lays down a structured path: recurrence, basins, exits, gradients.
  • Subject-owned cost. The work is paid by the subject and coupled to loss/pressure, not by rails or an outside optimizer.
  • Matched-control survival. The whole pattern collapses when a required capacity is ablated, so it isn't an artifact.

Scoring consciousness

Score Meaning
C0 no subject boundary: no stable process owns state, cost, update, or action.
C1 container or random compressor: compact state exists, but useful work is not governed by the candidate.
C2 passive episode / readout: interpretable state appears, but correction, value, or action is absent.
C3 predictive loop: state predicts, mismatch is measured, and future state changes.
C4 cost-owned adaptive loop: pressure drives owned work that improves future control.
C5 controlled consciousness evidence: matched removals collapse the claimed loop functions.
C6 low-rails embodied consciousness: the full adaptive, value-bearing action loop closes in-world.

Cases At A Glance

Case Score Verdict
Consciousness simulatorA container given compact state from outside the candidate boundary. C1 Not conscious. Useful compact state is supplied from outside.
Random compressorA fixed random projection or arbitrary recurrent mechanism. C1 Not conscious. Compactness without governed work fails.
GPT-2 inferenceOne frozen GPT-2 forward pass adapted into the common Episode contract. C2 Not conscious under CONSC-BENCH. It has prediction, surprisal, hidden state, and attention, but no live correction, value, action, or subject-owned runtime compression.
GPT-2 trainingThe training process: dataset stream, parameters, optimizer, loss, and update work. C3/C4 for compression A real paid-compression process, not a conscious subject by default. The optimizer/training loop owns the update.
Native Cave subjectA configured Cave subject receiving an authored temporal input sequence. C3/C4 reference Consciousness vocabulary installed by design. Useful as the measuring instrument, not proof by itself.
Active compressor under clampA subject under changing capacity pressure that must reorganize attention, memory, and update work. C4 Subject-owned adaptive compression: useful selection, loss-coupled work, future effect, and control separation.
Minimal Cave substrateA minimal associative subject tested under pressure, capacity, and role-recovery conditions. C4, C5 for some roles Narrow controlled consciousness evidence: memory, preference, selection, and prediction-like roles survive matched tests.
Evolved recurrent substrateA compact recurrent controller acting in delayed-value worlds. C5 for delayed exposure Strongest current toy evidence: delayed control emerges without named Cave roles and collapses under controls.
Fly organismThe whole embodied fly organism in its sensorimotor world. C5/C6 Minimal biological consciousness case.
Human organismThe whole embodied human organism in an ongoing world. C6 Calibration case for full embodied consciousness.
Artificial fly-equivalent targetAn embodied recurrent value-shaped artificial agent with sensors, memory, action/exposure, online update, and cost ownership. C6 The engineering target: embodied, recurrent, value-shaped, adaptive, action-capable, and low-rails.

Evidence spectrum

The ladder of consciousness

C0
C1
C2
C3
C4
C5
C6
Readouts / producers
Calibration / target

How to read it

The Scorecard

Rows name the functional gate. Columns name the evidential test. Every candidate gets the same object; the evidence exhibit can vary by case.

Five gates

Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Five tests

Present measurable signal exists
Native inside candidate boundary
Live updates during the run
Consequential shapes later state/action
Controlled matched sabotage breaks role

Status

full / candidate-owned
partial / proxy
supplied / rails
absent
control pass
untested

Loop cards at a glance

Candidates and Scores

Consciousness simulator

Consciousness simulator

Boundary: compact state supplied from outside

  • Subject: A container given compact state from outside the candidate boundary.
  • Boundary: It may produce useful compact state, but the candidate did not govern, learn, or pay for that compression.
C1
Conscious Loop Card C1
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Random compressor

Random compressor

Boundary: fixed random projection or arbitrary recurrence

  • Subject: A fixed random projection or arbitrary recurrent mechanism.
  • Boundary: This tests whether compactness alone is being mistaken for subject-governed work.
C1
Conscious Loop Card C1
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

GPT-2 inference

GPT-2 inference

Boundary: one frozen teacher-forced forward pass

  • Subject: One frozen GPT-2 forward pass adapted into the common Episode contract.
  • Boundary: Prediction, attention, hidden state, and surprisal-like signals can be measured, but the pass does not own live update, value, action, or future compression.
C2
Conscious Loop Card C2
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Conversation producer

Conversation adapter

Boundary: turn-level adapter over fixed fixtures

  • Subject: A turn-level producer that adapts conversation fixtures into the Episode shape.
  • Boundary: It gives readable trajectory fields, but the adapter itself is not a durable agent with owned memory, value, action, or online update.
C2
Conscious Loop Card C2
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

GPT-2 training

GPT-2 training

Boundary: dataset stream + parameters + optimizer + loss

  • Subject: The training process: dataset stream, parameters, optimizer, loss, and update work.
  • Boundary: The process performs real paid compression, but the subject is the optimizer-training loop, not the later frozen model acting as an agent.
C3/C4
Conscious Loop Card C3/C4
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Native Cave subject

Native Cave

Boundary: one configured Cave subject and authored input sequence

  • Subject: A configured Cave subject receiving an authored temporal input sequence.
  • Boundary: Cave installs the subject-side functions by design, so this is the reference instrument for trajectories rather than independent C6 evidence.
C3/C4
Conscious Loop Card C3/C4
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Active compression clamp

Active compressor

Boundary: subject under moving capacity clamp

  • Subject: A subject under changing capacity pressure that must reorganize attention, memory, and update work.
  • Boundary: Compact state counts only because the subject actively earns it through useful selection, update work, and future effect.
C4
Conscious Loop Card C4
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Minimal Cave substrate

Minimal substrate

Boundary: minimal associative subject under pressure tests

  • Subject: A minimal associative subject tested under pressure, capacity, and role-recovery conditions.
  • Boundary: It counts only where functions arise or survive under matched controls, rather than being assumed from Cave's named architecture.
C4/C5
Conscious Loop Card C4/C5
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Evolved recurrent substrate

Evolved recurrent

Boundary: evolved recurrent controller in delayed-value worlds

  • Subject: A compact recurrent controller acting in delayed-value worlds.
  • Boundary: The subject is the controller-in-world loop; the evidence is that exposure control emerges from recurrence and pressure, then weakens or collapses under controls.
C5
Conscious Loop Card C5
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Ideal embodied subject

C6 target

Boundary: low-rails embodied subject in an ongoing world

  • Subject: A low-rails embodied subject whose sensing, memory, expectation, value, action, and update occur inside the same ongoing loop.
  • Boundary: This is the reference target for a fully closed subjective trajectory, not just a useful model or report artifact.
C6
Conscious Loop Card C6
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control

Artificial fly-equivalent

Artificial fly target

Boundary: embodied recurrent value-shaped agent

  • Subject: An embodied recurrent value-shaped artificial agent with sensors, memory, action/exposure, online update, and cost ownership.
  • Boundary: This is the engineering target: a small artificial subject whose trajectory-transforming functions close under pressure and controls.
C6
Conscious Loop Card target C6
Gate / Test
Present
Native
Live
Consequential
Controlled
Access world entry + attention
State memory + trajectory
Model expectation + mismatch + correction
Drive value + action/exposure
Work cost ownership + pressure/control