DOOMFLYLEARNING LAB
Pixels in. Neural activity out. Proving learned survival takes controls. The live stream now runs experimental v6 training. These earlier tests remain failed; live weight changes do not establish learned survival.
THE CONTROL CAUGHT IT
CANDIDATE 05 · FAILThe cue response changed. Punishment didn’t explain it.
Reduction in the trained cue’s MBON response, after training the left cue. The same reduction without imposed punishment fails the reinforcement control. The counterbalanced right-cue experiment also failed. Changed connections alone do not establish learning.
DOES MEMORY BUY MORE TIME?
CANDIDATE 06 · REAL DOOMBlue floor causes engine damage. Gray floor is safe. Live pixels drive the full neural graph and the same fixed button mapping in every arm. Test rounds freeze memory and turn off punishment stimulation.
LEARNING ON
3.66 smean observed survival- Changed connections
- 1,862
- Training replicas / test starts
- 1 / 2
- Deaths / capped rounds
- 2 / 0
FROZEN
5.83 smean observed survival- Changed connections
- 0
- Training replicas / test starts
- 1 / 2
- Deaths / capped rounds
- 1 / 1
SHUFFLED TIMING
3.66 smean observed survival- Changed connections
- 1,896
- Training replicas / test starts
- 1 / 2
- Deaths / capped rounds
- 2 / 0
Exploratory pilot, capped at 8 s. One training replica is insufficient to establish a repeatable benefit. Capped rounds have unknown eventual survival. Physiological and conditioning checks still fail.
TEST STARTS + MEMORY ERASURE
| Check | Start seed | Observed survival | End |
|---|---|---|---|
| Unseen start | 61041 | 3.66 s | Death |
| Unseen start | 61042 | 3.66 s | Death |
| Vision blacked out | 61041 | 7.31 s | Death |
| After 5 s rest | 61041 | 3.66 s | Death |
| Memory erased | 61041 | 8.00 s | Time cap |
Blacking out vision and erasing learned connections test what behavior depends on. Five seconds of simulated rest is a short technical check, not evidence of lasting biological memory.
VISION REACHES MEMORY. RECOVERY FAILS.
Adding the documented R8 → aMe12 visual pathway made a blue image activate 14 visual memory cells in an early one-second assay; black activated none. The later model still stays too active after the image disappears.
| Image shown | Dopamine cells before | 3 s after removal |
|---|---|---|
| left blue | 22 / 18 Hz | 106 / 105 Hz |
| right blue | 22 / 18 Hz | 108 / 95 Hz |
| white | 22 / 18 Hz | 108 / 105 Hz |
Two identified dopamine cells, measured separately. These are simulated spike rates. Published resting activity helped calibrate their starting rates; persistent cue-triggered activity remains unvalidated. Restoring a pathway does not validate the whole visual system.
MODEL CHANGES + EXACT CLOCKS
All six candidates retain the same male reconstruction. New models separate annotated modulators from generic fast excitation, add a documented R8 cotransmission route, and test KC adaptation and a dopamine-centered memory rule on 4,184 existing KC → MBON11 connections. The rule is adapted from published work; transfer to this full network is unvalidated.
The runner processes every 35 Hz Doom frame and every 0.1 ms neural timestep. A health decrease schedules exactly 200 ms of stimulation into identified dopamine cells, beginning next game tic. That feedback is engineered. Health and coordinates never select an action.
Recorded episodes processed 80.2 brain seconds in 490.7 wall seconds (0.16×). Neural integration accounted for 92% of that time on the shared host. Faster-than-real-time full-brain training has not been achieved.
EVIDENCE + LIMITS
- A reconstructed male wiring diagram with chosen dynamics, not a living or literal fly brain.
- R8 input projection and display spectra are inferred; photoreceptors and APL still lack calibrated graded physiology.
- The centered plasticity rule is adapted from a reduced published model; full-graph parameter transfer is unvalidated.
- Cue-induced persistent activity and no-punishment effects invalidate the current conditioning interpretation.
- No claim of survival learning, consciousness, natural fly behavior, or priority over prior work is supported.
ARCHIVED PILOT 01 · ORIGINAL FAILED BASELINE
Visual memory input
FAILControlled images must activate and distinguish memory-cell responses.
Selective conditioning
FAILPaired stimulation must affect its cue more than control cues and backward/no-US conditions.
Survival memory changes
FAILGame experience must change the identified memory synapses.
Matched reinforcement exposure
FAILShuffled controls must receive the same amount of imposed stimulation.
The original pilot’s shuffled exposure was mismatched. The scheduling fix matched exposure in two follow-up trials; those still changed zero memory connections. The original failed pilot remains shown here.
DOES THE SIGNAL REACH MEMORY?
Separate population readouts, not a verified serial pathway. One simulated second, identical starting states. The original pilot’s model responds to pixels, but its tested motion and memory populations produce no spikes. This historical assay predates the R8 correction.
SURVIVAL: ON VS OFF
REAL DOOM · HEALTH GATHERINGDamaging floor. Health pickups. Fixed neural controls. Tests use new game seeds, with reinforcement and further weight changes switched off.
PLASTICITY ON
12.00 smean observed test time- Changed memory connections
- 0
- Independent training seeds
- 2
- Test episodes / timeouts
- 4 / 4
FROZEN
12.00 smean observed test time- Changed memory connections
- 0
- Independent training seeds
- 2
- Test episodes / timeouts
- 4 / 4
SHUFFLED REWARD
12.00 smean observed test time- Changed memory connections
- 0
- Independent training seeds
- 2
- Test episodes / timeouts
- 4 / 4
Recorded test episodes from one matched seed pair. Overlapping outcomes are shown as measured. All held-out episodes reached the 12 s cap; their eventual survival time is unknown. This small pilot cannot establish a survival benefit.
THE MEMORY RULE
Recent activity leaves a fading trace at a Kenyon cell’s existing output connections. Spikes from identified PPL101 dopamine cells can weaken eligible connections onto MBON11. These are 4,184 existing connections inside the full retained male connectome.
The equations and parameters are an experimental approximation. Dopamine delivery is modeled separately from fast excitation for these two cells. The mapping from game damage to stimulation is engineered; it is not a validated pain pathway.
| Conditioning control | Changed connections | Efficacy remaining | Dopamine-cell spikes |
|---|---|---|---|
| paired | 4,184 | 10% | 915 |
| backward | 4,184 | 43% | 397 |
| frozen | 0 | 100% | 516 |
| no imposed US | 4,184 | 10% | 915 |
| reset memory | 4,184 | 10% | 915 |
These tests directly stimulate memory cells and bypass vision. The full-network candidate loses cue specificity; changes also occur without imposed reward. Numerical rule tests passing does not make this a valid reproduction of fly conditioning.
FAST DEV / EXACT CLOCKS
Development runs omit browser updates, video compression and real-time pacing. They still generate every sensory frame and integrate every 0.1 ms neural timestep. Independent seed replicas can run in separate processes. Brain checkpoints preserve the complete model state. Serial and parallel runs produced identical recorded inputs, spikes and controls in the scheduling checks. The neural kernel remains the main compute cost.
22 episodes processed 260.9 seconds of brain time in 1240.8 seconds of wall time (0.21× aggregate speed). These measurements include competition with other work on the simulation host.
EVIDENCE + LIMITS
- One reconstructed male, repeated across game seeds; no independent biological animals.
- This is a small development pilot, not a confirmatory survival study.
- The imposed-KC assay bypasses vision and does not reproduce the original odor experiment.
- No imposed US does not silence endogenous dopamine neurons.
- Quiet-delay retention in an LTD-only model is a modeled property, not proof of biological retention.
- Failed gates prevent a learning claim or promotion to the live broadcast.