Retrieval check answer. Start from the confident null-pointer stack frame and grow outward — check what that module received as input, and whether the ambiguous timing-flavored log lines are actually downstream symptoms of that same null dereference — rather than starting at the first ambiguous log line and working forward. A confident island anywhere constrains its neighbors more than grinding through noise in signal order does; the module name is the strongest evidence in the whole trace, so it should redirect where you look first, exactly like Hearsay-II redirecting effort toward a confident mid-utterance word hypothesis instead of insisting on left-to-right processing.
The design comparison
Design C (recall-heavy, compacted context) wins on durability past a reset. Here’s the mechanism, stated precisely rather than just asserted:
Every time Design C’s agent needs an early fact and it’s not sitting verbatim in context, the agent must reconstruct it — from a summary, from its own prior reasoning, or by re-deriving it from a memory store. Each such reconstruction is a genuine retrieval attempt: effortful, cue-poor, exactly the kind of “use” that last lesson’s formula sums into base-level activation. If that reconstructed fact gets written into any persistent store (a memory file, a re-derived summary line, a re-confirmed belief), that store now reflects a chunk that has accumulated multiple real use-events — the same mechanism that made chunk B (three effortful retrievals) out-rank chunk A (one fresh exposure) in the last lesson.
Design R’s agent, by contrast, never has to reconstruct anything — the fact is always sitting in context, a cheap recognition-style lookup every time. Nothing about repeatedly seeing an already-present fact routes through the effortful retrieval pathway; it’s closer to re-encoding the same content over and over than to retrieving it. When the context resets, whatever persisted (if anything) never went through a use-event in the activation sense — there’s no accumulated sum to fall back on, because nothing was ever actually retrieved under conditions that would count as a “use.”
Concretely: ask both agents to reproduce the fact from scratch after the reset. Design C’s agent has, along the way, effectively re-derived and re-confirmed that fact multiple times under real effort — its durable memory (whatever it wrote down, whatever pattern of re-derivation it learned) should reproduce it far more reliably. Design R’s agent has only ever read the fact, never retrieved it in the effortful sense, so there’s no reason its post-reset performance should be any better than its performance on a fact it saw exactly once.
Why this is the design principle behind the whole track
This lesson is where the pattern gets named explicitly: information sitting directly in context is recognition (cheap, always-available, doesn’t strengthen anything by being looked at again); information that must be reconstructed from a separate store or from your own generation is recall (effortful, but the effort is what builds durability). A system — human or agent — that always keeps everything trivially accessible never builds the equivalent of a strengthened retrieval pathway, because nothing is ever actually being reconstructed under conditions that count as a use.
This is also, concretely, why this puzzle path asks you to produce or select an answer before showing you one, rather than handing you an explanation to read first: reading the explanation first would make every one of these lessons a Design-R experience. Answering first, even with effort and even with some wrong guesses, is a Design-C experience — and per this lesson’s own mechanism, that’s the version that compounds.
Where this goes: if effortful retrieval is what compounds, when you’re forced to retrieve matters as much as whether you do — retrieving something you haven’t fully forgotten yet barely counts as effortful at all. Next lesson (immediately) takes on timing: why spacing retrievals further apart, rather than clustering them right after first exposure, changes how much each one is worth.