[auto] Add principle-no-lookahead, principle-report-the-disqualifier, report-rendering

[EdbertChan]

Summary

Two principles from a backtesting session that spent roughly three hours producing results from a signal that could not have been known at the time it was credited to.

principle-no-lookahead — score a decision only against inputs that existed before that decision was made. Names the one defect across its four vocabularies (leakage, look-ahead bias, point-in-time correctness, repainting), and carries the cheap mechanical test: recompute from only the data available at time t and compare against the full-history computation; any difference is a leak.

principle-report-the-disqualifier — a report is finished when a reader who disagrees knows where to look. Requires a falsification observation, a measured/inferred/conjecture grade per claim, ruin reporting where the path can cross an absorbing barrier, and the count of configurations tried.

Review Claim

Both skills state a general principle, name the established prior art with DOIs, and carry a mechanical test rather than only prose.

Safety Invariant

Where no prior art exists it is recorded as such rather than invented: "repainting" is marked vendor and forum vocabulary with no peer-reviewed source.

Motivation

The incident: a study's chart view placed a pivot on the bar it occurred; the live view only from its confirmation bar, 15 to 50 minutes later. A truncation test showed the signal at 00:05 does not exist when the data ends at 00:05, 00:30 or 00:45, and appears when it ends at 00:55. Backtesting the chart view produced 23 green days out of 23 and 38,970% on a 10,000 USD account; the live view liquidated the same account. Neither the too-good result nor the unvalidated code path had a name in the corpus.

Non-goals

No existing skill, hook, or install path is modified. No always-on file is touched.

Test Plan

Manual review. Both files carry valid frontmatter; every claim of prior art resolves to a DOI or a stated absence.

Revert Plan

git revert the commit; two new directories, no existing file changed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01F43CBUnsDEs6J2zEC1r8a8

---

Note

Low Risk
Adds markdown skill docs under corpus/skills/ only; no runtime, auth, or data-path changes.

Overview
Adds three new corpus skills (documentation-only; no hooks or install paths changed).

principle-no-lookahead codifies point-in-time evaluation: separate display vs decision artifacts, require cutoffs and a truncation test (prefix recompute vs full history), treat implausibly good metrics as bugs, and cite leakage/rolling-origin literature—with explicit “no peer-reviewed prior art” for repainting vocabulary.

principle-report-the-disqualifier defines what causal or outcome reports must expose: falsifying observations, measured / inferred / conjecture grades, absorbing-state (“did the path survive”) columns, before/after fix headlines, and sensitivity to frozen constants; it cross-links to existing assert-invariants and no-lookahead skills.

report-rendering requires delivering analyses as one self-contained HTML (embedded data: figures, print CSS) plus PDF from the same HTML via headless Chrome, with a minimal Python + shell recipe and a gate to not render unvetted numbers.

^{Reviewed by Cursor Bugbot for commit bff693d. Bugbot is set up for automated code reviews on this repo. Configure here.}

---

Added: report-rendering

Third skill on this branch, from the same session. A three-section analysis was
called finished while it sat as separate markdown files in a repo; asked to see
the report, the only answer was a list of paths.

Ship one self-contained HTML file with figures embedded as data: URIs, plus a
PDF printed from that same HTML so the two cannot disagree. The recipe needs
nothing beyond a browser: markdown to HTML, then
google-chrome --headless --print-to-pdf. Print styles are called out
explicitly because their absence is invisible in the browser and only breaks in
the PDF.

Verified by producing reports/master-report.html (885,056 bytes, figure
inline) and reports/master-report.pdf (4,477,296 bytes).

Grounding: RFC 2397 for data: URIs, Knuth 1984 and Peng 2011 for the
document-and-analysis-as-one-artifact lineage, and an explicit "no known prior
art" for the delivery-format principle itself.

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_460a78d6-0366-4542-b122-182afa359448)

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_c05823a5-0592-482b-9cf1-b516cb3b7784)