Two paths from the menu put to the test: repairing the compass fails a third time, memory of visited places produces the second real success of the campaign

Beginner track

Where we left off

Chapter 11 ended with a pause in the work and a menu of four paths to resume the investigation into the cold start (the agent having to find its first tree without ever having seen one). None of these paths had been launched yet — it was a menu, not a committed plan. This chapter tells what happened when the first two, the cheapest, were finally tested for real — both dated July 21, 2026, in the project log. The results are very different from each other: one closes a door for the third time in a row, the other opens a new one with a real success — modest, but real and honest.

Path 1: Repairing the compass precisely where it goes wrong — failed again, and in a more precise way

Recall from Chapter 11: the project's compass (the score comparing each imagined story to the memory of "a chopped tree", see Chapter 4) turned out to point backwards once taken out of its training ground — a close tree received a lower score than an empty meadow. The cheapest path on the menu proposed fixing this with a small additional rule, trained this time with real "close" and "far" examples taken directly from the crafting game (not from the expert Treechop games, as in the previous attempts of Chapter 8).

On paper, the result looked like the best of the entire investigation: the new rule separated "close" images from "far" images better than all previous attempts, and it even passed a brand-new test — manually judging, on real game screenshots, if a close tree indeed received a better score than a treeless scene. It was correct 21 times out of 24 (87.5%).

But a deeper test dismantled this result. This project has long followed a simple rule: any new path must be checked for a trap already spotted twice (Chapter 8, then again in Chapter 9 when a planned repair had worsened the same problem) — a distance rule that, instead of truly judging the distance to the tree, actually contents itself with detecting whether the image is bright or dark (the scene's brightness). This time, the verification was pushed one step further than planned: someone had the idea to check if the test set itself, the one used to grade the rule "good" or "bad", was biased. The result: yes, it was, almost completely — the images chosen as "close tree" were, almost systematically, much darker (dense forest) than the images chosen as "no tree" (open meadow, beach). The link between brightness and label was almost perfect.

In other words: the beautiful 87.5% result was most likely not real proof that the rule finally understands where the tree is — it was the same old shortcut (brightness) detected for a third time, hidden behind a test that, unintentionally, measured the exact same thing as the shortcut itself. This is the third time this precise trap has appeared, with three different ways tried to correct it (artificially varying the light during training in Chapter 9, changing the training examples here) — and each time, the "brightness" shortcut remains stronger than ever. The lesson that now stands out: this shortcut most likely lives within the vision model itself, the one that learned to see in the first place (Chapters 1-2) and which these small repairs never touch — not in the way we train the small rule placed on top. This specific path on the menu is therefore closed, unless we attack the starting vision model directly, which is a much heavier project, set aside for now.

Path 2: A memory of already visited places — built, tested, and a real second success

The second path on the menu proposed a different idea: instead of comparing each image to a fixed landmark (the compass, which we just confirmed for the third time goes wrong outside its training ground), why not give the agent a simple memory of the places it has already explored during the current episode, and push it to go look elsewhere?

The important point, respected to the letter since Chapter 11: this memory must owe nothing to the broken compass. It is constructed in a completely different way, without any learning: at each step, the agent already knows the effect of its own action (move forward, turn left, turn right — like a player who counts their own steps and turns with their eyes closed). By adding up these known effects over the course of the episode, we reconstruct an approximate position on an invisible grid, and we count how many times each cell of this grid has already been visited. When the agent seems to turn in circles without a goal, instead of making it spin in place (the old reflex of Chapter 7, which can find nothing where there is nothing) or dashing straight ahead blindly (Chapter 8), we now push it towards the least-visited direction on the grid.

First test, very small (3 episodes): the mechanism triggers properly, behaves visibly differently from the old reflexes, and the coverage map indeed grows over time. No sign of the problem seen several times before (the agent starting to repeat a single action almost all the time). This test was only meant to verify that nothing was broken — not yet to judge if it works.

The real test, larger (20 episodes): 1 log chopped and planks crafted out of 20 trials (5%), average reward 0.45 — about 12% better than the MineRL baseline for an agent playing at random. This is the second non-zero result of this entire long investigation, after the small correction in Chapter 8 ("execute its own good plan longer", 9.7% success rate on its own batch of tests) — and it is the first to come from a completely different mechanism: not a better way to choose actions, not a better judgment, but a simple memory of "where have I already been". In the successful episode, the agent had indeed explored a large part of the map before finding its tree — and its behavior, at the moment of chopping, resembled that of a real lumberjack (the same actions that had already worked in Chapter 6), not an action repeated by accident.

Honesty on what this 1 out of 20 proves, and what it does not prove

We must be clear, as with every figure in this project: a 5% success rate is of the same order of magnitude as the 9.7% of Chapter 8 — not statistical proof that this new path is better at this sample size. A classic statistical test would not distinguish these two figures from each other with confidence.

But this result has two qualities that make it truly interesting, not just another number: it is clearly above the two previous attempts that had yielded zero out of eight trials each (Chapters 9 and 10), and above all, across the 20 episodes of this batch, none show the action-locking problem that had ruined several previous attempts (Chapter 8). It is therefore the second independent method — after the correction of Chapter 8 — to produce a real success without ever showing this broken behavior. It is not a confirmed breakthrough. It is a real positive point, honest, which deserves to be kept and extended rather than tucked away in a drawer.

One last honest detail: the batch of 20 episodes was managed by a process that encountered a technical infrastructure issue before it could write its own formal report — the figures above were therefore verified directly, episode by episode, in the raw execution logs, rather than copied from a final report that could not be produced.

Where the Chapter 11 menu stands now

Of the four paths proposed in Chapter 11: the first (repairing the compass) is now closed for good, unless a much heavier repair is made to the vision model itself. The second (memory of visited places) is built, tested, and yields a real positive signal to extend. The last two — a second, slower brain to search for a forest, or copying real human search play — remain, as in Chapter 11, unlaunched.