Legible Networks

An eight-post series on the unit-net / explainable-AI research arc: geometry, training, transplant, and four case studies in what exact legibility buys you — blind spots, latent knowledge, audit methodology, and the geometry of deep computation. A human–AI collaboration (Remy & Claude).

THE SERIES IN SIX CLAIMS

  1. 1Training Without Gradients in a Weight-Budget Geometry

    Optimization methods ranked by the information channel they consume — measured head-to-head on one substrate.

    In this post
    • The puzzle
    • The geometry
    • What backprop had to survive
    • The annealer, and what actually mattered
    • Interlude: delta evaluation, or how the annealer got 50× cheaper
    • The information ladder
    • When does the ladder invert?
    • The thesis the day earned
    • Why this is post 1
  2. 2A Language Model You Can Read With Arithmetic

    An exact interpretability lens, designed concept neurons, and the first measurement of how lossy averaged lenses actually are.

    In this post
    • The puzzle
    • Getting an honest student (harder than the science)
    • The lens results
    • Workspace signatures, replayed in a readable substrate
    • j-carve: cutting task reasoners out of J-space, with certificates
    • Where this left us
    • Addendum: j-carve on the transplanted Qwen (added 2026-07-26)
  3. 3Your Network Is Already (Mostly) a Unit-Net

    Pretrained superhuman nets can be moved into the legible geometry exactly, without a gradient step.

    In this post
    • The puzzle
    • The idea: scale symmetry and scale debt
    • Where debt goes to die
    • From naive collapse to exact fold
    • The last 0.03%: scale debt becomes temperature
    • What legality buys, immediately
    • The transplant law
  4. 4Anatomy of a Blind Spot: KataGo's Cyclic Vulnerability, End to End

    The complete mechanism of KataGo's cyclic vulnerability: not a fooled feature, a certified representational absence.

    In this post
    • The puzzle
    • 1. Confidently, in-distribution wrong
    • 2. The urgency horizon
    • 3. The misread is causally steerable
    • 4. The rule, exactly: extracting the automaton
    • 5. The comparator is a passenger
    • 6. Calibration defects vs representational defects
    • 7. The certified absence
    • 8. Scale doesn't buy it
    • 9. Why LW readers should care
  5. 5The Stack the Model Never Read

    A textbook ELK case study: perfect latent structure, a behavior that ignores it, and the finetune that elicits it.

    In this post
    • The puzzle
    • 1. Making the Chomsky hierarchy measurable
    • 2. The counter: three answers at three levels
    • 3. The soft stack
    • 4. The dissociation, and the hunt for the reader
    • 5. The bag
    • 6. The elicitation
    • 7. Why this matters
  6. 6Consumers, Not Representations — and How We Convicted Our Own Nulls

    “We found a direction for X” is half a result; the other half is whether anything reads it.

    In this post
    • The puzzle
    • 1. The specimen
    • 2. Attempt one: let training build the reader
    • 3. Attempt two: surgery
    • 4. Attempt three: give the mass to search
    • 5. Attempt four: composition — and the control that makes it meaningful
    • 6. The seeds sweep: the kill-test the conclusion owed
    • 7. The gain ledger: putting our own nulls on trial
    • 8. The corrected laws
    • 9. On publishing your own convictions
  7. 7Audit Your Audits: The Baseline Law and the Grounding Law

    Two methodological laws for anyone measuring “drives” in trained agents: baselines and grounding.

    In this post
    • The puzzle
    • 1. The instrument
    • 2. The trap, and the three rebasings
    • 3. What the residual actually is: a kill-test chain
    • 4. The grounding law
    • 5. Why we think this pair matters
  8. 8The Drifting Frame and the Program Atlas: What Deep Computation Actually Is

    Net path attribution is identically zero past ~5 layers; information persists while the coordinate frame rotates.

    In this post
    • The puzzle
    • 1. Evaporation: where the positive mass goes
    • 2. Effective depth: the visible computation is 1–2 blocks deep
    • 3. Total cancellation: the terminal statement
    • 4. Persistence: information stays; encodings rotate
    • 5. The drifting frame
    • 6. The deviation lens
    • 7. The program atlas: counting a network's situations
    • Coda: the series, and what we'd ask of you