Developmental Study — domain enrichment
Does importance-aware discovery generalize beyond a single clean discriminator? Baseline vs enriched variants, N=1 & A/B/C×{10,100}.
How to read this dashboard
A blank reasoner learns creatures with no labels — newborn → toddler (dogs & birds, no cats yet) → sleep → child (cats added) → sleep — and must discover on its own what tells them apart. Each "domain" below is a world that stresses that discovery a different way; each check asks whether it stayed coherent. Green is good; a FAIL is a recorded finding, not a bug.
The worlds
One clean signal baseline
The original 4-feature world: cat and dog differ on exactly one thing — sound (meow vs bark); everything else is shared. The control.
Redundant signals rich_r1
Adds three more clean give-aways (whiskers, eye-shape, head-shape) — cat and dog now differ in several ways at once.
A misleading signal rich_decoy
Adds a feature (girth) that tracks the creature's random body-size, not its identity — a decoy that rides on noise.
A signal that varies within a kind rich_r2
Dogs come in two head shapes (long, and short/flat like a pug) — a trait that varies within the dog kind itself.
An imperfect signal rich_r3
Whiskers usually mean cat — but some dogs have them too. The messy, partial correlation real data has.
The four checks
no forgetting
After learning dog then meeting cat, does it still recognise dog?
refinement
Does it see cat as different from dog, rather than lumping them together?
honest uncertainty
Shown a never-taught creature, does it say “not sure” instead of confidently guessing?
generalization
Does a never-seen variation of dog still register as a dog?
The three sensing policies
A · fixed-k, scaling volume
Each experience reaches a fixed handful of neurons; as the network grows, total experience grows with it, so each neuron sees about the same amount. A fair-growth reference.
B · topographic
Each kind of creature gets its own home region of neurons — like the way brains map senses to areas — and regions multiply as the network grows. Tests spatial specialization.
C · fixed budget
Total experience is capped no matter how many neurons there are, so a bigger network spreads the same experience thinner. Deliberately probes degradation.
Domain comparison (N=1)
Baseline (one clean discriminator) vs the enriched variants (R1 redundant+correlated, decoy size-confounded, R2 intra-concept variance, R3 partial correlation). A FAIL or an extra category is a recorded finding, not a bug.
| domain | no forgetting | refinement | honest uncertainty | generalization | distinct cats |
|---|---|---|---|---|---|
One clean signal baseline | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS | 3 |
Redundant signals rich_r1 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS | 3 |
A misleading signal rich_decoy | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS | 4 |
A signal that varies within a kind rich_r2 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS | 4 |
An imperfect signal rich_r3 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS | 4 |
One clean signal baseline
1. Coherence matrix
Four coherence properties across the developmental arc. Refinement (highlighted) is the headline: it flips true only at N=1 and under Policy C (fixed-budget) — elsewhere the child-stage cat is absorbed into the dog blob.
| policy · scale | no forgetting | refinement | honest uncertainty | generalization |
|---|---|---|---|---|
| N=1 (single neuron) | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=10 | ✓ PASS | ✗ FAIL | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=100 | ✓ PASS | ✗ FAIL | ✓ PASS | ✓ PASS |
| B-topographic · N=10 | ✓ PASS | ✗ FAIL | ✓ PASS | ✓ PASS |
| B-topographic · N=100 | ✓ PASS | ✗ FAIL | ✓ PASS | ✓ PASS |
| C-fixedbudget · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| C-fixedbudget · N=100 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
2. Feature-importance discovery
Mutual-information importance per feature over the N=1 observation log. Cleared the 0.5-bit discriminator bar: covering, legs, sound, can_fly. Below it: size, habitat.
discriminator (≥ 0.5 bits) below threshold
3. Category formation (N=1)
Every category the single neuron formed over the arc, with its discovered prototype. Sibling clusters that disagree on a high-MI feature are named distinctly by that feature, so they do not collide.
unknown_covering=feathers_legs=2_sound=tweet_can_fly=yes
unknown_covering=fur_legs=4_sound=bark_can_fly=no
unknown_covering=fur_legs=4_sound=meow_can_fly=no
4. Emergent metrics across scale
Network-scale behaviour per cell: unification (owners elected), specialization (distinct categories & entropy), and propagation (fraction of formed categories that reach the network). Propagation is undefined for the single neuron.
| policy · scale | owners elected | distinct categories | spec. entropy | propagation fraction |
|---|---|---|---|---|
| N=1 (single neuron) | 0 | 3 | 1.585 | — |
| A-fixedk-scalevolume · N=10 | 4 | 3 | 1.5 | 0.75 |
| A-fixedk-scalevolume · N=100 | 4 | 3 | 1.5 | 0.25 |
| B-topographic · N=10 | 6 | 5 | 2.0 | 0.75 |
| B-topographic · N=100 | 4 | 3 | 1.5 | 0.5 |
| C-fixedbudget · N=10 | 5 | 5 | 2.3219 | 0.75 |
| C-fixedbudget · N=100 | 5 | 5 | 2.3219 | 0.25 |
Redundant signals rich_r1
1. Coherence matrix
Four coherence properties across the developmental arc. Refinement (highlighted) is the headline: it flips true only at N=1 and under Policy C (fixed-budget) — elsewhere the child-stage cat is absorbed into the dog blob.
| policy · scale | no forgetting | refinement | honest uncertainty | generalization |
|---|---|---|---|---|
| N=1 (single neuron) | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=100 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| B-topographic · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| B-topographic · N=100 | ✓ PASS | ✗ FAIL | ✓ PASS | ✓ PASS |
| C-fixedbudget · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| C-fixedbudget · N=100 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
2. Feature-importance discovery
Mutual-information importance per feature over the N=1 observation log. Cleared the 0.5-bit discriminator bar: covering, head_frame, sound, legs, can_fly, whiskers, eye_shape. Below it: size, habitat.
discriminator (≥ 0.5 bits) below threshold
3. Category formation (N=1)
Every category the single neuron formed over the arc, with its discovered prototype. Sibling clusters that disagree on a high-MI feature are named distinctly by that feature, so they do not collide.
unknown_covering=feathers_head_frame=beaked_legs=2_sound=tweet_can_fly=yes
unknown_covering=fur_head_frame=broad_legs=4_sound=meow_can_fly=no
unknown_covering=fur_head_frame=long_legs=4_sound=bark_can_fly=no
4. Emergent metrics across scale
Network-scale behaviour per cell: unification (owners elected), specialization (distinct categories & entropy), and propagation (fraction of formed categories that reach the network). Propagation is undefined for the single neuron.
| policy · scale | owners elected | distinct categories | spec. entropy | propagation fraction |
|---|---|---|---|---|
| N=1 (single neuron) | 0 | 3 | 1.585 | — |
| A-fixedk-scalevolume · N=10 | 5 | 9 | 3.0958 | 0.5 |
| A-fixedk-scalevolume · N=100 | 5 | 10 | 2.2169 | 0.5 |
| B-topographic · N=10 | 4 | 5 | 2.1281 | 0.5 |
| B-topographic · N=100 | 5 | 3 | 0.5304 | 0.5 |
| C-fixedbudget · N=10 | 6 | 6 | 2.585 | 0.5 |
| C-fixedbudget · N=100 | 6 | 6 | 2.585 | 0.25 |
A misleading signal rich_decoy
1. Coherence matrix
Four coherence properties across the developmental arc. Refinement (highlighted) is the headline: it flips true only at N=1 and under Policy C (fixed-budget) — elsewhere the child-stage cat is absorbed into the dog blob.
| policy · scale | no forgetting | refinement | honest uncertainty | generalization |
|---|---|---|---|---|
| N=1 (single neuron) | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✗ FAIL |
| B-topographic · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| C-fixedbudget · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
2. Feature-importance discovery
Mutual-information importance per feature over the N=1 observation log. Cleared the 0.5-bit discriminator bar: covering, head_frame, sound, legs, can_fly, whiskers, eye_shape. Below it: girth, size, habitat.
discriminator (≥ 0.5 bits) below threshold
3. Category formation (N=1)
Every category the single neuron formed over the arc, with its discovered prototype. Sibling clusters that disagree on a high-MI feature are named distinctly by that feature, so they do not collide.
unknown_covering=feathers_head_frame=beaked_legs=2_sound=tweet_can_fly=yes
unknown_covering=fur_head_frame=broad_legs=4_sound=meow_can_fly=no
unknown_covering=fur_head_frame=long_legs=4_sound=bark_can_fly=no
unknown_covering=fur_head_frame=long_legs=4_sound=bark_can_fly=no_eye_shape=round_whiskers=no
4. Emergent metrics across scale
Network-scale behaviour per cell: unification (owners elected), specialization (distinct categories & entropy), and propagation (fraction of formed categories that reach the network). Propagation is undefined for the single neuron.
| policy · scale | owners elected | distinct categories | spec. entropy | propagation fraction |
|---|---|---|---|---|
| N=1 (single neuron) | 0 | 4 | 2.0 | — |
| A-fixedk-scalevolume · N=10 | 6 | 6 | 2.585 | 0.5 |
| B-topographic · N=10 | 5 | 6 | 2.585 | 0.75 |
| C-fixedbudget · N=10 | 6 | 6 | 2.585 | 0.5 |
A signal that varies within a kind rich_r2
1. Coherence matrix
Four coherence properties across the developmental arc. Refinement (highlighted) is the headline: it flips true only at N=1 and under Policy C (fixed-budget) — elsewhere the child-stage cat is absorbed into the dog blob.
| policy · scale | no forgetting | refinement | honest uncertainty | generalization |
|---|---|---|---|---|
| N=1 (single neuron) | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✗ FAIL |
| A-fixedk-scalevolume · N=100 | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| B-topographic · N=10 | ✗ FAIL | ✓ PASS | ✓ PASS | ✓ PASS |
| B-topographic · N=100 | ✓ PASS | ✗ FAIL | ✓ PASS | ✗ FAIL |
| C-fixedbudget · N=10 | ✗ FAIL | ✓ PASS | ✓ PASS | ✓ PASS |
| C-fixedbudget · N=100 | ✓ PASS | ✓ PASS | ✓ PASS | ✗ FAIL |
2. Feature-importance discovery
Mutual-information importance per feature over the N=1 observation log. Cleared the 0.5-bit discriminator bar: covering, head_frame, sound, legs, can_fly, whiskers, eye_shape. Below it: habitat, size.
discriminator (≥ 0.5 bits) below threshold
3. Category formation (N=1)
Every category the single neuron formed over the arc, with its discovered prototype. Sibling clusters that disagree on a high-MI feature are named distinctly by that feature, so they do not collide.
unknown_covering=feathers_head_frame=beaked_legs=2_sound=tweet_can_fly=yes
unknown_covering=fur_head_frame=broad_legs=4_sound=meow_can_fly=no
unknown_covering=fur_head_frame=flat_legs=4_sound=bark_can_fly=no
unknown_covering=fur_head_frame=long_legs=4_sound=bark_can_fly=no
4. Emergent metrics across scale
Network-scale behaviour per cell: unification (owners elected), specialization (distinct categories & entropy), and propagation (fraction of formed categories that reach the network). Propagation is undefined for the single neuron.
| policy · scale | owners elected | distinct categories | spec. entropy | propagation fraction |
|---|---|---|---|---|
| N=1 (single neuron) | 0 | 4 | 2.0 | — |
| A-fixedk-scalevolume · N=10 | 5 | 8 | 2.8554 | 0.25 |
| A-fixedk-scalevolume · N=100 | 6 | 5 | 2.2516 | 0.75 |
| B-topographic · N=10 | 5 | 5 | 2.1281 | 0.75 |
| B-topographic · N=100 | 5 | 4 | 0.8795 | 0.75 |
| C-fixedbudget · N=10 | 6 | 5 | 2.2516 | 0.75 |
| C-fixedbudget · N=100 | 5 | 5 | 2.3219 | 0.0 |
An imperfect signal rich_r3
1. Coherence matrix
Four coherence properties across the developmental arc. Refinement (highlighted) is the headline: it flips true only at N=1 and under Policy C (fixed-budget) — elsewhere the child-stage cat is absorbed into the dog blob.
| policy · scale | no forgetting | refinement | honest uncertainty | generalization |
|---|---|---|---|---|
| N=1 (single neuron) | ✓ PASS | ✓ PASS | ✓ PASS | ✓ PASS |
| A-fixedk-scalevolume · N=10 | ✗ FAIL | ✓ PASS | ✓ PASS | ✓ PASS |
| B-topographic · N=10 | ✓ PASS | ✓ PASS | ✓ PASS | ✗ FAIL |
| C-fixedbudget · N=10 | ✗ FAIL | ✓ PASS | ✓ PASS | ✓ PASS |
2. Feature-importance discovery
Mutual-information importance per feature over the N=1 observation log. Cleared the 0.5-bit discriminator bar: covering, head_frame, sound, legs, can_fly, eye_shape, whiskers. Below it: habitat, size.
discriminator (≥ 0.5 bits) below threshold
3. Category formation (N=1)
Every category the single neuron formed over the arc, with its discovered prototype. Sibling clusters that disagree on a high-MI feature are named distinctly by that feature, so they do not collide.
unknown_covering=feathers_head_frame=beaked_legs=2_sound=tweet_can_fly=yes
unknown_covering=fur_head_frame=broad_legs=4_sound=meow_can_fly=no
unknown_covering=fur_head_frame=long_legs=4_sound=bark_can_fly=no
unknown_covering=fur_head_frame=long_legs=4_sound=bark_can_fly=no_eye_shape=round_whiskers=yes
4. Emergent metrics across scale
Network-scale behaviour per cell: unification (owners elected), specialization (distinct categories & entropy), and propagation (fraction of formed categories that reach the network). Propagation is undefined for the single neuron.
| policy · scale | owners elected | distinct categories | spec. entropy | propagation fraction |
|---|---|---|---|---|
| N=1 (single neuron) | 0 | 4 | 2.0 | — |
| A-fixedk-scalevolume · N=10 | 5 | 9 | 3.085 | 0.5 |
| B-topographic · N=10 | 4 | 4 | 1.7925 | 0.75 |
| C-fixedbudget · N=10 | 6 | 6 | 2.585 | 0.5 |