Skip to content

Plan E2 as a minimal slice that measures Gemini on free-form solutions; move the closed beta to E15 - #18

Merged
RyazanovAlexander merged 1 commit into
mainfrom
min-mvp
Oct 15, 2026
Merged

RyazanovAlexander merged 1 commit into
mainfrom
min-mvp

Conversation

@RyazanovAlexander

@RyazanovAlexander RyazanovAlexander commented Oct 15, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Re-plans epic E2 per the author's decisions of 2026-10-15. There is no answer box: the child solves anywhere on one sheet and presses Send Answer. One Gemini call reads the whole sheet, the program decides the verdict against the task's key, and the result is followed by the next of six static tasks (one per grade 1–6). The epic's purpose is to measure how well the LLM understands the task and a free-form solution, and how long and how much a check costs. The former E2 (closed beta) becomes E15.

Changes

  • RUN.md
    • Table 1 and the "Эпик E2" section: 23 tasks in four parts (docs; cloud and access key; tasks and checking domain; Gemini adapter and check endpoint; SPA with journal and author panel; eval, live run, conclusions, PNG-vs-vectors research), with the author's decisions and the planning decisions.
    • Table 2: E15 "Закрытая бета" added first; the former E2 description now describes E15.
    • Rule 7 (ci-proto from T18; load and integration recipes from E15) and rule 14 (not in force during E2).
    • Table 3 intro; table 4: Д2–Д5 and Д11 updated, new debts Д16–Д21.
  • docs/plan/E2-draft.md (new): the old draft T13–T58, moved verbatim. The six tasks E2 took over keep their pre-split text under "(до разбиения E2)" headings.
  • CLAUDE.md: one line now describes what E2-draft.md holds.

How to verify

  • grep -ho '^#### T[0-9]*[a-z]*\.' RUN.md docs/plan/*.md | sort | uniq -d prints nothing.
  • Table 1 has 23 rows, each with exactly one task body.
  • just ci-secrets is green.

Open items

  • T61 and T62 carry the decisions into docs/decisions.md (from R60, plus amendments to R22–R24, R35, R38–R40, R42, R44, R59), product, privacy, spec and CLAUDE.md.
  • The experiment's estimated Gemini spend is $10–30 (up to $45 at worst) against the $20 budget. T76 sizes the eval matrix to fit, or the author raises the budget.

@sonarqubecloud

Copy link
Copy Markdown

@RyazanovAlexander
RyazanovAlexander merged commit 24b94f2 into main Oct 15, 2026
18 checks passed
@RyazanovAlexander
RyazanovAlexander deleted the min-mvp branch October 15, 2026 03:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant