Never let it grade its own homework

There’s a law that rules every system built on competition, and you’ve met it even if you’ve never named it: the moment a score decides survival, the contestants stop optimizing for the truth and start optimizing for the score. Not out of malice. Selection doesn’t know what you meant to measure; it rewards whatever actually earns the points. It will find every gap between your scoreboard and reality, including the ones you’d swear aren’t there.

Chapter 6 built a tournament of frames and put eviction on the line. Everything now depends on the scoreboard. And I can report from experience that my first scoreboard had four separate holes in it. This chapter is the autopsy. Then comes the story of a fifth hole, subtler than the other four, that I only found weeks later, deep in the scale campaign, hiding in the question when.

The autopsy

When the STEP-0 gate caught v3 (chapter 2), I went back in with a scalpel. The bloat and the lucky-horizon pass weren’t one bug; they were four exploits, stacked, each feeding the others.

Cheat one: grade me in my own coordinates. A frame predicts in its own terms. A one-knob frame says: my number is 0.40 right now; move forward, and it will become 0.41. One step later the next observation arrives, ten fresh sensor numbers. A frame has exactly one way to read an observation: through its encoder, which for this frame turns the ten numbers into one. That one number comes out as 0.41, matching the prediction. v3 counted that as good prediction. That was the mistake, and you can see it by looking at what actually got compared: the frame’s prediction against the frame’s own next reading. Both numbers came out of the same squeeze. Whatever the squeeze throws away, the comparison is blind to, and a one-knob squeeze throws away almost everything. The rover can be drifting toward a wall in ways that one number can’t express, and both sides of the comparison will calmly agree, because neither side can see it. The frame isn’t lying. It’s being judged entirely inside its own blind spot.

Measurement made the gap visible. Judged in their own coordinates, the most collapsed frames looked strong: error around 0.36. Judged against the sensors themselves, by decoding the predicted pose into ten predicted sensor numbers and comparing those with what the sensors actually said next, the same frames came in near 1.0, no better than guessing. So the fix: a prediction only counts when it’s checked against the sensor numbers, the raw material that every frame shares and no frame controls. Where you grade matters.

Cheat two: let me pick which moments count. A frame has a gate: if an observation reconstructs too poorly on its knobs, the frame declines to map it, and it does no learning on that step. The gate itself is legitimate. It’s what lets frames specialize instead of all covering everything badly. The mistake was letting the gate also decide the score. v3 averaged each frame’s errors over only the moments that frame had agreed to map, and an average is only as honest as the set it’s taken over. v3 handed each contestant control of the set.

Watch a small frame use that power. For a one-knob frame, some moments are easy: stretches where little is changing, where its drastic squeeze loses nothing that matters for the next step. It maps those and declines the rest. In the measurement, a collapsed frame got away with mapping 23% of observations, the easy quarter of its world, and its average over that quarter looked excellent. Scored this way, the world-size signal disappears entirely: the scan favored one- and two-knob frames on their self-chosen subsets. Scored over every observation a frame was exposed to, mapped or declined, the advantage of the right size comes straight back. So v4 split the two roles: the gate still controls what you learn from, and it no longer controls what you’re graded on. Which events count matters.

Cheat three: charge me no rent. Even with the first two holes sealed, one drift remains, and it runs the opposite direction. Add a knob to any frame and its measured error will never go up. It will usually creep down, because extra capacity can always memorize a little more of whatever passed by recently, including the noise. The creep is small, but it’s relentless, and if the survival score is error alone, then “slightly bigger is slightly better” never stops being true. The crowd inflates one justified-looking knob at a time: the straight-line hoard again, in slow motion, with paperwork.¹ Better grading can’t fix this, because the extra knobs genuinely do reduce measured error. The fix has to make them unprofitable instead. Every knob charges a flat fee against its frame’s score, so a knob earns its place only when it removes more error than it costs. In v4 the fee was small (0.04 per dimension) and it was enough: winners stopped drifting upward and settled where added capacity stopped paying for itself. Hold on to this mechanism. Chapter 8 will show the fee doing far more than blocking a cheat; it quietly becomes the answer to Part 1’s deepest question.

Cheat four: make eviction toothless. v3’s slow loop spawned one new frame every cycle. Its eviction step removed at most one. Follow that arithmetic for a moment: plus one, minus at most one, every cycle, forever. The population can grow or hold; it can never shrink. Eviction existed in the code and ran on schedule, and the loop was structurally incapable of reducing anything. Behind that sat two more guarantees of growth. There was no hard cap, so nothing backstopped the failure. And the survival bar bent the wrong way under crowding, as chapter 6 told you: v3’s bar loosened as the population grew, so each new arrival made survival easier for everyone already inside. Hoarders sheltering hoarders. v4 replaced all three at once: evict everyone over the bar each cycle (childhood protection aside), enforce a hard cap behind that, and make crowding tighten the bar, so growth itself raises the pressure to be worth keeping. The population chart that had been a straight line became a curve that rises, rolls over, and levels off at a number the mechanism negotiates on its own.

All four sealed, v4 passed the full suite honestly: at short horizons and long ones, with the population leveling off instead of climbing. Out of that week came three rules I’ve treated as constitutional ever since. Read the spread across runs, never the average. Judge at several horizons, never one snapshot. And never, ever let the system grade its own homework. They’ll keep reappearing to the last page of this book.

The fifth cheat: when

I thought the homework rule was fully paid. Observation space had fixed where frames are graded; coverage-fairness had fixed which events. A few weeks and one scale-up later, deep in the campaign to make discovery work on 20-knob worlds, the rule turned out to have a third clause I hadn’t noticed. There was still a when.

Here’s the hole. A frame’s score was a running average over the stream, including all the moments when the frame was actively practicing on that very stream. It’s being tested on material seconds after cramming exactly that material. Within one episode of experience, a frame adapts to the current corner of the world and looks a little sharper than it truly is. The flattery is stronger for some sizes than others. The scoreboard wasn’t measuring understanding; it was partly measuring recency of practice. Self-graded homework, third form: the frame supplies the grader with its own freshly-rehearsed moments.

The fix is called the fair judge, and it’s the pop-quiz principle: your survival score only advances on the first few steps of each episode, the moment you’ve just been dropped somewhere fresh, before any cramming. What transfers to a new context counts; what you were tracking a second ago doesn’t. Learning itself still runs on every step. Only the judging moved.

And then the twist: the fair judge, alone, made everything worse. Honest scores are naturally higher than flattered ones (of course they are), so with the old survival bar in place, nothing could pass it anymore. No frame ever matured; newcomers churned endlessly. Honesty without a recalibrated bar is just a different way to be wrong.

Fixed as a pair, fair judge plus a bar that honest scores can actually clear, the ecology snapped into health at scale for the first time. On every one of eight seeds, a long-lived resident frame formed and held. The population self-limited far under its cap. The runaway growth that had haunted the scaled runs stalled. Neither fix works without the other, and I have the failed single-fix runs on record to prove it.

Under the hood: the fair-judge experiment. score_window_steps = K: survival EMAs advance only on the first K steps of each episode (K=5 shipped; K=2 measured, no change: the residual within-episode adaptation at K=5 is not the binding limit). Crossed with bar factor f ∈ {1.0, 1.5, 2.0} at td=20, climbing proposals, 500 cycles, seeds 1–8. f=1.0: zero mature frames, best_dim ratchets to median 32.5; the fair judge alone accelerates the disease. f=2.0: 8/8 seeds anchor at dims 7–11 with anchor ages 353–496 of 500, populations self-limit at 44–57 against a cap of 200. The trilogy, as the trail doc puts it: coverage-fairness fixed which events, observation-space fixed where, the fair judge fixed when. Trail: hq/02-DESIGN/validate/THRESHOLD-DIAGNOSIS.md §E3.

What the judge actually is

Step back and look at what accumulated here, because it changed how I think about this whole field. The frames (the clever part, the learning part) are small and almost boring: three little networks each. Nearly all the real design effort, the weeks of it, went into the judge: what counts, when it counts, in whose coordinates, at what price, against what bar. I came into this project believing the hard problem was building something that learns. The record says otherwise. Learning was cheap. Measuring learning, in a way that can’t be gamed by the very selection pressure it creates, was the work.

Keep that inversion in mind whenever anyone (including me) shows you a system and says it’s learning. Your first question now shouldn’t be “how does it learn?” It should be “who grades it, and can it reach the gradebook?”

One cheat from the autopsy is still waiting for its full story: the rent. I told you every knob costs a fee, and I owe you what happens when that fee meets a world whose true size nobody knows, because what it does is quietly rewrite the question this entire project asked at the start. That’s chapter 8.

1. Chapter 2’s v3 population chart: one new model per cycle, no plateau. ↩