One question per lab. Play until you can predict the fix. Harness research is the study the labs cite — keep it on the path, not in a drawer.
Same tokens, four costs. Click a box. Labels stay short on purpose. The first box splits further — look vs store: DSA cuts softmax, KDA cuts occupancy, MLA only thins each stored token.
No tabs. No packed files. If two names shared a page, they were not the same topic.
What grows when I add one token?
K and V, every layer, every KV head — then that whole cache is re-read on the next decode step. A 128k window is a larger tax, not a discount.
If decode got cheaper, why is the GPU still full?
DSA stops looking at most tokens and still stores them. KDA stops storing them in 3/4 of the layers. MLA only thins each stored token.
Advertised window, or effective window?
Attention is a budget. More tokens thin every key; the middle of the prompt is where the budget runs out. Pin at the edges; don’t keep the haystack.
What does the provider key its cache on?
Exact byte-prefix. Change one character, prefill is billed again. Keep the key stable — that is the fix, not a cleverer cache.
What is allowed to pass through a lossy summarizer?
Unpinned rules die when their turn is dropped (38% if dropped, 0% if they survive). Pin them in the system prompt so they never enter the summarizer. Constraint survival is this topic, not a second lab.
The objective is still in context. Why did the run stop serving it?
Completes against a different target. Detect-only is theater. Isolate noisy work, echo goal and rules as two channels, gate completion against the original instruction. P25 T1–T6 are in the lab, not a tab next door.
The objective is still unmet. Why did the run call itself done?
63% of failures are agent-initiated self-stops: its own smoke test passed, the real bar didn't. Goal drift's mirror image — this stops early against a weaker target instead of wandering to a different one.
What fraction of traffic actually needs the expensive model?
Pay frontier only for the hard slice. The bill is that fraction, not “which model is best.”
What does retrieval serve after the world changed?
The stale pattern, until v2 replaces v1. Contradiction is not an update. Lazy staging; consolidate on recurrence.
Should every token rewrite memory?
No. Write on surprise, forget on a schedule. Train-on-every-token is the bug Titans exist to fix.
21 areas · 35 problems · ledger. The labs cite this corpus; they do not replace it. Open the atlas, then walk the path. Serving is area 09; P17 is the byte-key; 19–21 are the later territories.
What must a harness actually solve?
Dashboard, ⌘K, lifecycle map, 21 syntheses, 35 problem deep-dives, paper notes, ledger. Suggested path is on the dashboard — including serving, prospective memory, identity, and metrics.
Which buffer may you write to at inference, and when?
Survey + verdict. Play the simulator first; this page is the evidence table and the serving-tier map.
Where do the 35 problems actually fire?
A catalog tape, not a mechanism. Useful after you have felt compaction and drift.
What actually drives hypertrophy — timing tricks, or tension + daily protein?
3D map and tables. Different domain. Start at MPS if you want the one transferable timing fact.