A blank reasoner was taught to tell animals apart on its own. In a simple world it succeeded — but when the world got richer, it broke. This is the journey from there to here: what the study tested, what broke, and how each fix turned a red column green.
Before the results make sense, here's the setup — in plain terms.
A blank reasoner learns creatures the way a child might — no labels, no teacher. It has to discover the categories from raw experience, fed in developmental stages:
The hard part isn't memorising — it's working out on its own what tells two look-alike animals apart (a cat and a dog share four legs and fur; only their sound differs), and refining "dog" into "dog vs cat" without forgetting either — all while staying honest about creatures it hasn't been taught.
What it targeted. An earlier result showed this discovery worked — but only in a hand-built world with one clean give-away. This study asked the harder question: does the discovery still work when the world is richer and messier? Each of the five worlds below makes it harder in a different, deliberate way.
Every world holds the same five creatures (dog, cat, bird, fish, reptile). What
changes is how a cat can be told from a dog — each world stresses one way reality is
harder than a single clean signal. The short codes (rich_r1 …) are just the labels
used in the data and the dashboard.
The original 4-feature world. Cat and dog differ on exactly one thing — sound (meow vs bark); everything else (four legs, fur) is shared.
Adds three more clean give-aways — whiskers, eye-shape, head-shape — so cat and dog now differ in several ways at once.
Adds a feature (girth) that tracks the creature's random body size, not its identity — a decoy that rides on noise.
Dogs come in two head shapes — long, and short/flat like a pug — a real trait that differs within the dog kind itself.
Whiskers usually mean cat — but some dogs have them too. The messy, partial correlation real data actually has.
Isolating one difficulty per world is what let the study prove the two defects were independent — watch the misleading-signal world stay broken after the first fix, in the matrix below.
Each row is one world; each column is a moment in time. A cell shows that world's four coherence checks as coloured dots ( pass, fail), in the fixed order 1·2·3·4 explained below. Read left-to-right and watch the red turn green as each fix lands.
| world | ① Original studyunfixed mechanismwhere we started | ② After fix #1record-time merge cap · PR #18over-merged worlds separate again | ③ After fix #22nd-highest-MI importance · PR #19the decoy stops masquerading |
|---|---|---|---|
| One clean signalbaseline · the control | 4/4 | 4/4 | 4/4 |
| Redundant signalsrich_r1 | 2/4 | ✓ recovered | 4/4 |
| Varies within a kindrich_r2 | 3/4 | ✓ recovered | 4/4 |
| Imperfect signalrich_r3 · never broke | 4/4 | 4/4 | 4/4 |
| A misleading signalrich_decoy | 2/4 | 2/4 | ✓ recovered |
Note: this is the single-neuron discovery point, where the enrichment defects live. Scattered failures at network scale (many neurons learning together) are a separate, earlier scaling result — not what these two fixes address. The full per-scale grid is in the live dashboard.
Measure first, understand the cause, then make the smallest change that fixes it — and re-measure honestly at every step.
the honest, surprising finding
Adding clean, redundant signals (Redundant signals) and a size-confounded decoy (A misleading signal) broke coherence. The mechanism didn't survive richer worlds — inverting the hypothesis. Two distinct root causes were diagnosed, not asserted.
record-time merge cap
As it filed away new sightings, the reasoner merged near-identical ones. With more features, two same-kind animals that differed only on random traits got merged into a single blob — before discovery could run. A cap on how different a merge tolerates fixed it: Redundant signals and Varies-within-a-kind recovered. The decoy stayed broken.
2nd-highest-MI importance
The reasoner rated a feature "important" if it correlated with any other feature — so the size-tracking decoy, which mirrored the random body-size trait, was promoted like a real cue. Requiring a second corroborating witness demoted it — A misleading signal recovered. All green.
Both fixes are small, surgical, and monotone in spirit — each designed with diagnostics on the live code and shipped with a measured blast radius (the control world stayed 4/4, every asserting test green).
Two sightings merge only when they differ on at most one feature — AND-ed onto the existing similarity gate, so the change can only ever remove a merge.
The old rule tolerated more differing features as the feature count grew, so enrichment let same-kind animals differing only on random traits collapse into one entry, starving the part that discovers boundaries. An absolute cap doesn't grow with feature count.
A feature's importance is now its second-strongest correlation with any other feature, not its strongest.
The strongest-only rule promoted a feature for correlating with anything — including a random trait. The size-tracking decoy correlated strongly with body size and cleared the bar. A genuine identity cue corroborates with a whole block of features, so its second-strongest stays high; a lone-partner decoy's collapses.
What the journey taught — beyond the two patches.