Research and development of emerging artificial consciousness through The Consciousness AI project, based on Feinberg and Mallatt's neuroevolutionary theory.
© 2026. All rights reserved.
This page replays one recorded episode of one training run. You see the frames the agent received, the bids each module made, the module with the highest bid at each step, the sync signal between modules, and the reward that came back. The agent is the mark at the centre of every frame, two rings and a dot, drawn in the colour of this site. The room holds two discs that look the same. One is the light, which gives food. The other is a decoy, which gives nothing. Each disc plays its own tone, and each pulses at its own steady rhythm, in the picture and in the sound.
In this episode the decoy’s tone is the louder one. The agent goes to the light anyway. It is in the light from step 16 on, on 185 of 200 steps.
The console does not show what moved the agent in this run. The tectum binds each disc to its tone by timing. A value gate, learned from reward in the earlier episodes, chose the light’s tone once the binding had formed, early in this episode, and the tectum’s motor map turned the agent toward that disc. The bars marked “After binding” include the binding. It multiplies the bids of vision and hearing by the same factor, so it never changes which of the two is higher. In the gate it changed the module with the highest bid at none of the steps. The console does not show which disc was joined to which tone, the value of each tone, or the motor map.
Hearing has the highest bid on 0.940 of the ignited steps here, and the workspace stays silent on 0.255 of steps. These numbers describe the workspace contest that the earlier gates judged. They did not steer the agent in this run.
This episode was not picked by eye. The run recorded the frames of episodes 0, 10, 20, 30 and 39. For each of the three seeds the published episode is episode 20, the first recorded episode of the window the gate judged. The console counts episodes from 0, so this is the 21st episode of the run. By the room’s fixed schedule the decoy’s tone is the louder one in it at every seed. The rule was written before any episode of this gate was exported.
The view is a window of 48 pixels around the agent in a room of 224 pixels, so the world moves and the agent does not. The colours are the ones this run was drawn with.
Outlined bar, raw bid. Filled bar, bound bid. All five rows share one scale, the highest bid of the episode, so bar lengths can be compared between modules. A raw bid never passes 1.0. A bound bid can pass it, because the affective modulator multiplies the raw bid. The colour marks the module, never a state of the world.
Each strip cell shows the module with the highest bid at that step. Dark cells are steps where the workspace is silent. The highest bid does not say which vector the policy received. The session page states that. Each line is drawn between its own lowest and highest value, which are printed inside it, so a line that fills the box can still be a flat line over a tiny range. The lower line is the SHAPED reward. That is the environment reward plus the curiosity terms the training loop adds. Tap or drag along the strip to move the playhead.
Read left to right. The raw bid is what each module offers. The settle cycles then run the workspace again and again, feeding the broadcast back, and the bids that entered the last cycle are shown second, with how much the broadcast changed from cycle to cycle. Inside every cycle the workspace applies affect, which adds a valence term to each bid and keeps it between 0 and 1, so a bid that was zero can rise, and which moves the threshold with arousal. Binding then multiplies each bid by 1 plus half of how well its oscillator aligns with the others. The affect and binding columns show the last cycle. Ignition compares the energy, the highest bound bid, with a running level, and the modules whose bound bid reaches 0.8 of the recorded threshold are admitted. Every bar in this panel shares one scale, the highest value in the episode. The last line names the module with the highest bid and the module whose vector the broadcast actually holds, compared element by element in the recorded vectors. With the default merge the two often differ, which the broadcast carrier result measured.
The manifest is written by the ethics framework at run start. The framework holds rules E1 to E8 in the framework document; the rules listed above are the ones a run must satisfy before its first step, and the others are checked by a person. No session on this page is evidence of feelings or awareness. Finding a light is automatic approach, which Feinberg and Mallatt exclude as evidence of affect. The consciousness clock never moves because of a session.
A check of the agent’s policy found that its last layer was never trained, so the agent drifted in one direction whatever it perceived. A motor map was then added to the tectum, so that the place the tectum selects turns the agent. Oscillators then let the tectum bind a seen object to its sound by timing. In a room with a light and a decoy that look the same, the agent still went to the object with the louder tone, and the light was the louder one in only half of the episodes. Nothing taught the agent which object gives food. This gate asked whether the agent can learn that from reward. Its three checks were written before the runs.
The gate PASSED, all three checks at 3 of 3 seeds (verdict). In episodes 21 to 40 the learning agent reached the light first in 56 of 60 episodes. With the learning switched off it did so in 28 of 60, and with the sound’s rhythm destroyed in 34 of 60. When the decoy’s tone was the louder one, it still went to the light in 28 of 30 of those episodes. The learned value of the light’s tone ended at 1.000 and that of the decoy’s tone at 0.000 at every seed.
Four limits stand with that result.
Flexible response to competing goals was not tested. A reversal of which object gives food is a later test.
The orienting movement is a fixed reflex, and what is learned is which object it is given.
The two rhythms are steady, and a separate test with irregular rhythms FAILED.
The policy’s own movement network is not used, because its last layer was never trained.
A learned value is not evidence of feelings.
| Environment | dark_room |
|---|---|
| Seed | 602 |
| Episode | 20 of a 40 episode, 200 step run |
| Ethics framework | v1.0, existence drive on |
| Flags | --enable-audio --rssm-latent-mode continuous --capsule-workspace-source all_levels --existence-drive on --dark-room-audio binaural --dark-room-audio-channels 4 --dark-room-collision --dark-room-view agent_centered --vision-bid-reduction zscore --audio-salience surprise --ignition-rule tolerance --ignition-tolerance-sd 1.0 --bid-precision gain --dark-room-agent-mark ring --dark-room-agent-colour 217,119,87 --dark-room-decoy-sound --dark-room-decoy-level 0.5,3.0 --dark-room-decoy-visible --dark-room-start-between --dark-room-pulse --dark-room-sound-direction off --tectal-orienting value --tectal-orienting-idle still --map-binding rhythm --value-learning-rate 0.3 |
| Measured result | PASSED, L1 to L3 at 3 of 3 seeds. The agent learned from reward which heard tone belongs to food, and used that value to choose which of two objects that look the same to follow. In episodes 21 to 40 it reached the light first in 56 of 60 episodes, against 28 of 60 with the learning switched off and 34 of 60 with the sound's rhythm destroyed. When the decoy was the louder object it still went to the light in 28 of 30 of those episodes. Four limits stand. Flexible response to competing goals, for example a reversal of which object gives food, was not tested. The orienting movement is a fixed reflex, and what is learned is which object it is given. The two rhythms are steady, and a separate test with irregular rhythms FAILED. The policy's own movement network is not used, because its last layer was never trained. A passed criterion is not evidence of affect. |
| Verdict | verdict document in the public research repository |
The command and the seed above are the whole recipe the run started from. The sound is seeded too. Whether two runs with the same command give identical files was not checked for this gate.
The highest bid strip and the ignition readout are instruments, not reports of experience. In this run they also do not show what moved the agent. The value gate and the motor map are recorded in the run’s own files and reported in the verdict. They are not in this replay. The binding appears only as the factor in the bars marked “After binding”. Nothing on this page claims feelings, awareness or consciousness, and no session moves the consciousness clock on the architecture page.
The two columns marked ground truth, in light and light in view, come from
the environment’s own state. The agent never receives them. They show the state
that produced the frame of that step, which is one step earlier than the reward.
On the step the agent enters the light, the reward is already 1.0 while in
light is still false.
Every number the console draws is read at run time from the session’s own
bundle, which holds steps.json, session.json and ethics_manifest.json
under /public/data/training_sessions/dark-room-value-gate-seed602-episode20/. The bundle was
produced from the recorded run by the repository’s export script, which strips
recording dates and local paths before anything is written. Nothing on this page
was typed by hand.
The other published sessions are listed on the training sessions index.