Version 3 · present verified 2026-10-16 · horizons 2027, 2029, 2031 · written by a model, reviewed by a person · hash b7f09a492e06250e…
Prerequisites: S-11
Plain Markdown for an agent about to read it.
1. Definition
Numerical precision is the fact that a real number in a machine is a nearby rational, and that the order of operations moves which rational you get. Summing a long vector from the large end and from the small end need not agree. A seed is part of the program: a bootstrap or a sample without a recorded seed is a different program each run. Reproducibility is the demand that the same inputs, including the seed and the library versions, produce the same outputs. The interview asks why two runs of the same backtest differ in a late decimal, and whether it matters. It matters when a decision, a sort, or a published greek sits on that decimal. It does not matter when both runs are noise at a scale the trade cannot see. The skill is knowing which of those you are in.
2. Why it matters at the frontier
A platform that prices options from a normal CDF will show fake skew if two desks use two approximations. That is not a market. It is a duplicate function. The frontier failure is a backtest that cannot be replayed, so a good day cannot be distinguished from a different seed. F5 is a replay. F2 is an estimate that moves when you change the summation order and is then treated as a signal. This repo has already been bitten by the CDF version: many copies of the normal integral, disagreeing far enough out that a wing looked like skew. The repair was one module. The lesson is that class of bug.
3. The test
Why do two runs differ in a late decimal, and does it matter? The pass lists seed, summation order, library version, parallel reduction, and an unstable sort over ties. It then says whether the decision threshold sits on the disagreeing digit. A candidate who says "floating point is random" fails. A candidate who says "it never matters" fails the wing of a surface. Probation is two CDFs and a strike far from the money, and a statement of which function is allowed to answer.
7. Now, 2026
As of this writing reproducible research is a published ideal and a patchy practice. Seeds are common in notebooks and missing in production jobs that "just sample." The seventh-decimal incident in this file is a local fact, not an industry survey. What changed is the number of libraries a single pricing path can cross. Each extra copy of a special function is a possible fake skew. The cure is boring: one function, a seed, a stable order.
8. Next
- By 2027, a stochastic job that cannot replay names its missing seed instead of publishing the draw (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 greek used downstream carries the function that computed the probability, 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 inputs through normCdf replay to the same bits (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) — "export function normCdf(x: number): number {"
- repo: (a file of this platform; the path is the licensed tier) — "copies could disagree in the seventh decimal"
- paper: the floating-point and reproducible-research literature; no DOI is claimed
- practice, unverified: a hash-map iteration order moving a late decimal of a backtest is a class of bug, not a measured diff of a named run
- 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: 2 (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-11, S-29 (skill_syllabus gives the order), then sit skill_exam on S-29 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.