Skip to content

Maintaining the Site

The first phase optimizes for learning velocity and judgment quality, not translation coverage.

Language policy

  • English is the canonical language for every tutorial, radar entry, experiment summary, and navigation label.
  • Do not block an English update on a translation.
  • Add Chinese only when a page has stabilized and the translation provides clear value for readers.
  • A translation should link to its English source and record the source revision or date.
  • If the two versions disagree, the English page is authoritative until the difference is resolved.

Suggested translation layout

When the audience justifies a second language, use a parallel path rather than bilingual paragraphs:

text docs/ en/ # canonical pages, if the site grows beyond one language zh/ # translated pages with matching slugs

The current single-language layout deliberately avoids that migration cost. Revisit it only when there are enough translations to justify language navigation and separate search indexes.

Publishing loop

  1. Open an Issue with the question, evidence, and proposed experiment.
  2. Read and reproduce the smallest useful slice.
  3. Update the relevant tutorial, radar entry, or research-ledger memo.
  4. Open a pull request and let CI run formatting, tests, links, and the strict site build.
  5. Publish only after the page states its evidence level and next validation step.

Research protocol

The repository uses a compact reading-and-reproduction protocol. It belongs to maintenance rather than the numbered tutorial sequence because it describes how to learn, not a technical mechanism.

Four passes

  1. Claim: what changed, against which baseline, data, tools, and compute?
  2. Mechanism: what changed in representation, prediction, feedback, or the improvement loop?
  3. Evidence: are evaluation data independent, comparisons fair, and evaluators game-resistant?
  4. Reproduction: what smallest experiment could change the current judgment?

Evidence ladder

Level Meaning Permitted conclusion
L0 Announcement, demo, or selected examples Monitor only
L1 Paper or technical report experiment Authors report the result under their setup
L2 Public artifacts reproduce the central result Reproducible in one disclosed setup
L3 Independent reproduction or adversarial evaluation Robust beyond the originating team
L4 Long-term uncontrolled use Operational reliability and limits are observed

Tutorial completion contract

Every tutorial should leave behind a mechanism summary, minimal implementation, controlled comparison, three concrete failures, fact/interpretation/hypothesis separation, a ledger update, and a falsifiable next experiment.

Keeping it sustainable

Prefer one strong update per month over a daily news feed. Keep large model weights, datasets, and raw logs outside Git; preserve configurations, metrics, failure examples, and changed judgments in the repository.