wai-quantum: a person's taste, learned from their own choices between equals #497

Merged

dcharlot merged 1 commit from wai/quantum-taste-r2 into main 2026-10-26 16:41:32 -04:00

Conversation 0 Commits 1 Files changed 21 +1349 -22

dcharlot commented 2026-10-26 15:50:03 -04:00

Owner

Copy link

Feature 10 of the co-design report: a personal, plural taste model.

Replays #496 on main after #495 merged, as one commit with the same tree. #496's checks were green except one: the macOS energy job's eoc-meter test a_child_processs_work_is_not_attributed. That is a process-energy test in another workspace, and it failed while the shared macOS host was loaded. The same job passed on #495, and the forge has no rerun, so this is pushed once, fresh. #496 is closed.

Engine (engine::taste)

• Only between equals. A choice is recorded only when the engine proves the two programs the same circuit, or the same state. The proof must be a tableau, exact or state check, never random inputs. So a choice is about taste, never correctness.
• The model. A Bradley–Terry logistic over differences of measured features: gates, two-qubit gates, depth, two-qubit depth, T gates and the modelled error, plus the text's statements, gate kinds, repetition and comments. It is fitted by ridge-penalised Newton's method, so the same choices give the same weights everywhere. A floor on each feature's spread stops a feature that barely varied from extrapolating to certainty.
• Decomposed. Each of economy, clarity, insight, precision and utility is learned from its own choices. A ranking names where a part, or the cost, puts another first.
• Plural and personal. Profiles are kept by name and never pooled.
• Typicality is shown, never learned. Typicality is 1 − novelty (#495). The novelty dial blends the standardised order by taste with the order by novelty.
• Measured.
  • Choices predicted as made (each from the ones before it) and held out (over the last 100).
  • Choices that nothing measured separates.
  • Whether a ranking's first three are built more alike than the first three by cost: the collapse check.
• next: the least certain pair that the engine proves equal.

CLI and MCP

• wai-quantum taste choose | show | rank | next, over files or a specification's elites (--spec). Profiles are kept in quantum-taste.json beside the user's key.
• MCP: taste_rank and taste_next read the person's profile. No tool records a choice. At initialize, an agent is told so, and told to call a program new only when novelty_qasm says so.

Studio

• "Your taste" in the Workbench:
  • Compare two, with the part the choice was about;
  • Rank by my taste, with each program's typicality and the dial;
  • profiles by name.
• In Alternatives, an alternative that is the same circuit as the draft, taken whole or refused whole, counts as a choice. Improve orders its candidates by the profile.
• Bindings: engine_taste in the page, taste in the desktop app.

CI: wai-web's native engine test count goes from 10 to 11.

Checked on the build host:

• engine engine:: 105 pass, 7 of them new;
• CLI 15 pass, clippy clean with default and no default features;
• wai-web native engine tests: 11 pass;
• desktop: 5 pass, clippy clean;
• Studio e2e: 121 checks pass, 10 of them new;
• the public-names check is clean.

🤖 Generated with Claude Code

Feature 10 of the co-design report: a personal, plural taste model. **Replays #496** on main after #495 merged, as one commit with the same tree. #496's checks were green except one: the macOS energy job's `eoc-meter` test `a_child_processs_work_is_not_attributed`. That is a process-energy test in another workspace, and it failed while the shared macOS host was loaded. The same job passed on #495, and the forge has no rerun, so this is pushed once, fresh. #496 is closed. **Engine (`engine::taste`)** - **Only between equals.** A choice is recorded only when the engine proves the two programs the same circuit, or the same state. The proof must be a tableau, exact or state check, never random inputs. So a choice is about taste, never correctness. - **The model.** A Bradley–Terry logistic over differences of measured features: gates, two-qubit gates, depth, two-qubit depth, T gates and the modelled error, plus the text's statements, gate kinds, repetition and comments. It is fitted by ridge-penalised Newton's method, so the same choices give the same weights everywhere. A floor on each feature's spread stops a feature that barely varied from extrapolating to certainty. - **Decomposed.** Each of economy, clarity, insight, precision and utility is learned from its own choices. A ranking names where a part, or the cost, puts another first. - **Plural and personal.** Profiles are kept by name and never pooled. - **Typicality is shown, never learned.** Typicality is 1 − novelty (#495). The novelty dial blends the standardised order by taste with the order by novelty. - **Measured.** - Choices predicted as made (each from the ones before it) and held out (over the last 100). - Choices that nothing measured separates. - Whether a ranking's first three are built more alike than the first three by cost: the collapse check. - **`next`:** the least certain pair that the engine proves equal. **CLI and MCP** - `wai-quantum taste choose | show | rank | next`, over files or a specification's elites (`--spec`). Profiles are kept in `quantum-taste.json` beside the user's key. - MCP: `taste_rank` and `taste_next` read the person's profile. No tool records a choice. At initialize, an agent is told so, and told to call a program new only when `novelty_qasm` says so. **Studio** - "Your taste" in the Workbench: - Compare two, with the part the choice was about; - Rank by my taste, with each program's typicality and the dial; - profiles by name. - In Alternatives, an alternative that is the same circuit as the draft, taken whole or refused whole, counts as a choice. Improve orders its candidates by the profile. - Bindings: `engine_taste` in the page, `taste` in the desktop app. **CI:** wai-web's native engine test count goes from 10 to 11. **Checked on the build host:** - engine `engine::` 105 pass, 7 of them new; - CLI 15 pass, clippy clean with default and no default features; - wai-web native engine tests: 11 pass; - desktop: 5 pass, clippy clean; - Studio e2e: 121 checks pass, 10 of them new; - the public-names check is clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 1 commit 2026-10-26 15:50:03 -04:00

wai-quantum: a person's taste, learned from their own choices between equals ... dcb5391f90

Feature 10 of the co-design report: a personal, plural taste model.

- engine::taste:
  - a choice is recorded only when the engine proves the two programs the
    same circuit or the same state, by a proof (tableau, exact, state),
    never by random inputs: it is about taste, never correctness;
  - a Bradley-Terry logistic over differences of measured features (gates,
    two-qubit gates, depth, two-qubit depth, T gates, modelled error) and
    of the text (statements, kinds of gate, repetition, comments), ridge
    penalised and fitted by Newton's method: deterministic everywhere;
  - the rubric's parts (economy, clarity, insight, precision, utility),
    each learned from its own choices; a ranking names where a part or the
    cost puts another first;
  - profiles personal and plural, by name, never pooled;
  - typicality (one less engine::novelty's novelty) shown and never
    learned; the novelty dial blends the standardised order by taste with
    the order by novelty; a floor on each feature's spread keeps a feature
    that barely varied from turning into certainty;
  - measured: choices predicted as made and held out (last 100), the ones
    nothing measured separates, and whether a ranking's first three are
    built more alike than the cost's; scores about the candidates' mean;
  - next: the least certain pair the engine proves equal; call for hosts.
- novelty::token is shared with it.
- The command line: taste choose | show | rank | next (files or --spec),
  profiles in quantum-taste.json beside the user's key.
- MCP: taste_rank and taste_next read the person's profile; no tool
  records a choice. initialize tells an agent so, and to call a program
  new only when novelty_qasm says so.
- The Studio: "Your taste" in the Workbench (Compare two with the part a
  choice was about, Rank by my taste with typicality and the dial,
  profiles by name); an alternative that is the same circuit as the draft,
  taken whole or refused whole, is a choice; Improve orders by the profile.
- Bindings: engine_taste in the page, taste in the desktop app.
- CI: wai-web's native engine tests count 11.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot referenced this pull request 2026-10-26 16:11:36 -04:00

wai-quantum: rewrites, convergence, and the 0.6.16 / 0.1.4 release #498

dcharlot merged commit 907c46889f into main 2026-10-26 16:41:32 -04:00

dcharlot referenced this pull request from a commit 2026-10-26 16:41:33 -04:00

Merge pull request 'wai-quantum: a person's taste, learned from their own choices between equals' (#497) from wai/quantum-taste-r2 into main

dcharlot referenced this pull request from a commit 2026-10-26 17:15:27 -04:00

Merge origin/main (907c46889): #497, a person's taste in wai-quantum

dcharlot referenced this pull request 2026-10-26 17:15:43 -04:00

wai merge train E: twenty-five PRs onto main, the relay's upstream window bounded, and three fixes the merged tree needed #499

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!497

No description provided.