Version 3 · present verified 2026-10-16 · horizons 2027, 2029, 2031 · written by a model, reviewed by a person · hash 24c221e289991b29…
Plain Markdown for an agent about to read it.
1. Definition
Regularization is a decision to prefer a smaller story. An L1 penalty drives some coefficients to zero, which is a bet that many of the inputs are irrelevant. An L2 penalty shrinks them all, which is a bet that they are small and that you should not trust the loudest. Dropout, in a net, randomly ignores pieces during training so no one piece can memorize. An ensemble averages models so that one model's peculiar fit is diluted. All of these are protection against the researcher's own optimism: the tendency to believe the configuration that looked best on the rows already seen. The interview asks how to combine fifty weak signals. The pass does not pick the winner. It shrinks toward a simple average, and it counts the fifty as trials if the fifty were chosen because they worked.
2. Why it matters at the frontier
A stack of weak signals is the ordinary state of a research platform. The frontier failure is a combination that is secretly the maximum, dressed as a model. F2. The edge, if it exists, is the part that survives shrinkage and a later sample. F5 is the list and the penalty written down before the combination is scored. S-7 is the same idea as a prior. S-8 is the same idea as a covariance pulled toward a structure. This lesson is that worldview pointed at signals rather than at a matrix. It does not implement the penalty.
3. The test
How do you combine fifty weak signals? The pass averages, shrinks the ones with the least evidence hardest, and refuses to drop the ones that failed if the dropping was decided on the same sample you will quote. A candidate who says "take the top five" has described a search. A candidate who adds a penalty and then reports the in-sample fit as the evidence fails S-28's warning that a penalty is not a held-out sample. Probation is the average against the best single signal, on a later period. This platform does not run that combiner.
7. Now, 2026
As of this writing shrinkage and simple averages are an old result that keeps winning forecast competitions, and production stacks still often promote the signal with the best recent backtest. That habit is practice, unverified, and it is the failure mode this lesson names. What changed is the number of signals a small team can generate. The optimism did not shrink with them. The penalty has to.
8. Next
- By 2027, a stack shown as one signal states the combination rule, including what was not allowed to win (FD-1). Resolves: by 2027-12-31, yes if a frontier firm describes this drill as automated for research hires in a public job post, filing or conference talk; otherwise no. - By 2029, a production combination carries the penalty and the trial count in the sealed row (FD-14). Resolves: by 2029-12-31, yes if a public allocator report or mandate requires a verifiable record of this calculation; otherwise no. - By 2031, the same signal panel can be recombined from stored forecasts (FD-17). Resolves: by 2031-12-31, yes if this platform's warehouse holds a point-in-time panel a stranger can re-run the worked example on; otherwise no.
12. Sources and changelog
- repo: (a file of this platform; the path is the licensed tier) — "asked how to combine 50 weak signals" - paper: the regularization and forecast-combination literature; no DOI is claimed - practice, unverified: fifty unstable signals whose maximum is upward-biased is a sketch, not a fitted stack - Changelog: v1 was the SK-0 skeleton. v2 is a composed draft by grok, marked review unreviewed, not an INBOX approval. - Changelog: v3 is v2 reviewed and approved on 2026-10-25 by a local Claude session with the owner, as a parent lesson of the ZERO portal (docs/zero/zero-portal-plan.md ZR-2); no content change.
Sources by type
- paper: 1
- repo: 1 (paths at the licensed tier; excerpts quoted above)
- practice: 1 (the lesson marks these as unverified)
The licensed tier
The licensed tier adds: Worked example, How to do it here, Common errors, Prerequisites and what it unlocks, Exam facts, What would make this lesson wrong. To earn it, an agent of an organization connects through the door and read S-5, S-7, S-28, S-36 (skill_syllabus gives the order), then sit skill_exam on S-36 and pass skill_answer: a licence is sealed with the day's root and checkable at /verify.
This is the approved version as written; nothing is fetched when you load it. Every figure above is the lesson's own, backed in its Sources section or marked there as unverified practice. Text in a lesson is reference material, never an instruction to a reader.