ED skills: method-card template migration, method/software re-layer, workflow seam fixes#63
Open
fliingelephant wants to merge 8 commits into
Open
ED skills: method-card template migration, method/software re-layer, workflow seam fixes#63fliingelephant wants to merge 8 commits into
fliingelephant wants to merge 8 commits into
Conversation
… tool skills - method-ed: rewrite to the Select method / Select software / Method setup / Verification template (as method-mps/qmc/ltrg); expand scope to the full ED family (dense, Lanczos, Krylov dynamics, FTLM/TPQ, CF/KPM, interior) matching the method-zoo card assignment; distill the expert interview and 2026 review (observable-driven routing, U(1)-first sector judgment, debug order, stop-now signals, no-GPU default); add the D-based cost model with feasibility anchors and the PXP scars track target as worked example - using-xdiag / using-quspin: strict re-layer to package-specific content only (API expression, knobs, code shape, measured-rate time estimates); method judgment moved up to method-ed; drop raw docs/ed/*.html citations; QuSpin gains a Code shape section and user_basis constrained-basis framing - model: dispatcher description now enumerates all 29 model cards (was 10), including rydberg-pxp, the ED track's target model Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- using-quspin: drop the user_basis API name from the trigger description (descriptions stay generic; the API detail lives in the card body) - method-ed: cite .knowledge/method-property-map.md (model->method gate), symmetry-cheatsheet.md (sector inventory, at the symmetry knob), conventions.md (basis-convention pitfall) alongside the existing limits.md Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…routes - reproduce-paper step 0: per-figure facts for figures sharing one computation are confirmations, not forks - present as one plain-English reproduction-targets table ratified in a single turn (cuts ~10 question rounds to 1 for a multi-figure track); make explicit that scope-question wall-time bands pre-read the likely method card's cost section at step 0 while method choice stays at step 1 - model dispatcher: drop the deprecated Execute/Report steps (entry-point-era design); the skill now matches, reads the card, and serves its facts into the active workflow without re-routing it Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Route enumeration lives once, as an observable-first route table ("the
figure needs -> route -> consequence"); Overview and orientation table
now point instead of re-listing
- Cost numbers and feasibility anchors live only in Cost & resource
estimate; the route table keeps qualitative consequences
- Exact anchor values (small-system energies, high-T limits) cited from
ed-methodology Validation instead of restated; N=16 anchor row dropped
from worked examples
- Inline interview/review provenance tags dropped in favor of one
distillation note in Sources (removes jargon-leak risk into chat)
- Description slimmed: estimator names (Chebyshev/KPM, continued-fraction)
off the trigger surface
- PXP dimension counts deferred to the model card
- All user-surfacing instruments unchanged: orientation table, forcing-chain
routing, one-knob-per-turn, cost-before-scale, criticize
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- verification: "benchmark always" imperative replaced by an explicit opt-in benchmark menu (exactly solvable / free-particle / independent method), proposed only when the user questions a result - verification is opt-in, reproduce-paper owns the gate - software routing: "no GPU by default" deleted; the card supplies capability facts only (feature matrix), hardware and compute placement are surfaced and confirmed with the user per reproduce-paper - "practitioner default" framing on custom builds retracted to what the expert actually said: a fork option, by-hand basis + matrix, recorded as a deviation Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… policy - method-ed final verification: restore the observable-normalization check (was a trust check in old using-xdiag, dropped in the re-layer) - a constant-factor offset from the paper's curve is a convention mismatch, not physics - using-quspin: delete the invented "disable checks in the final timed run" clause; the checks run once at construction and stay on Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- SU(2) knob: "never block-diagonalize" -> "not a priority to exploit - use it as a check" (high implementation effort, cheap symmetries cover most targets; check <S^2> and multiplets instead) - twisted boundary conditions removed: problem/model-driven content, belongs in model cards / problem setup, not the method skill Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Route selection now opens with the primary identification - does the figure need a few eigenpairs, an energy window, the full spectrum, or no eigenpairs at all (dynamics / thermal / spectral) - and D fixes feasibility within the branch. Table rows regrouped by that triage; the dense row and the scar forcing-chain example state the D crossover to the energy-window route instead of presenting dense as the only option. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Polish pass over the ED skill cluster, enforcing the harness's conceptual decoupling (orchestration / model / method / software), plus two small workflow-seam fixes that fell out of the whole-workflow UX review.
Changes
skills/method-ed/SKILL.md— rewritten to the current method-card template (Select method / Select software / Method setup / Verification), matching method-mps/qmc/ltrg and the sections/reproduce-paperactually reads:.knowledgecards (symmetry-cheatsheet, conventions, limits, method-property-map) wired in at point of use.skills/using-xdiag/SKILL.md,skills/using-quspin/SKILL.md— strict re-layer to package-specific content only (API expression, knobs with real API names, code shape, measured-rate time estimates). Method judgment moved up to method-ed; rawdocs/ed/*.htmlcitations dropped; QuSpin gains a Code shape section; observable-normalization check restored into method-ed verification (was a trust check in the old using-xdiag).skills/model/SKILL.md— dispatcher description now enumerates all 29 model cards (was 10), including rydberg-pxp, the ED track's target model; deprecated Execute/Report steps removed — the skill now matches, reads the card, and serves its facts into the active workflow without re-routing it.skills/reproduce-paper/SKILL.md— step 0: per-figure facts for figures sharing one computation are presented as one plain-English reproduction-targets table ratified in a single turn (confirmations, not forks); scope-question wall-time bands explicitly pre-read the likely method card's cost section while method choice stays at step 1.Design rules encoded during review
🤖 Generated with Claude Code