Alloyra

Research preview — live at alloyra.fly.dev.

An alloy-design workbench for metallurgists. Captures the duty (application, environment, loads), ranks candidate alloys, tunes composition off a base metal, and audits candidates against interaction failure modes (SCC, hydrogen embrittlement, creep, galvanic) that no single-property filter can catch.

Requirements: see the Alloyra Blueprint artifact (§ numbers referenced throughout the code).

Layout

• packages/core — pure TS domain library: composition handling, calculators (PREN, WRC-1992, CE(IIW), Andrews Ms, Larson-Miller), microstructure descriptor model + search (strengthening mechanisms, twinning, grain-boundary character), provenance types. Zero framework imports; validation cases in test/.
• packages/data — versioned seed dataset (spec-min values with citations) and the Postgres-first Drizzle schema (not yet wired; M0 reads seeds).
• apps/web — Next.js workbench UI. Fixed-viewport shell, ⌘K palette, dense grids, provenance chips, in-house SVG plot kit (property charts, calculator sweep sparklines) — blueprint § 8.1: an instrument, not a website.

Backlog: BACKLOG.md. Competitive analysis and design v2: docs/competitive-analysis.md, docs/design-v2.md.

• packages/calphad — the in-browser CALPHAD engine (backlog B-501, cross-checked): TDB parser, compound-energy-formalism Gibbs energies, multicomponent tangent-plane equilibrium, temperature sweeps, Scheil solidification, and sampled isopleth maps — all in a web worker. Validated against pycalphad on a 52-equilibrium battery across all four shipped databases; see docs/engine-validation.md.
• services/calphad — Python pycalphad service. Not deployed and not required by the product: it serves as the offline validation oracle for the engine (services/calphad/scripts/) and as an optional self-host for anyone who wants a second opinion from pycalphad. Ships license-vetted assessed databases; see services/calphad/databases/SOURCES.md.

Validation oracle (optional, local)

All phase computation in the product runs in the visitor's browser — there is no calculation server. To regenerate the engine-validation report against pycalphad locally:

cd services/calphad
uv venv .venv && uv pip install -p .venv/bin/python -e .
cd ../../packages/calphad
node scripts/gen-cases.ts
(cd ../../services/calphad && .venv/bin/python scripts/crosscheck_oracle.py \
  ../../packages/calphad/scripts/crosscheck-cases.json \
  ../../packages/calphad/scripts/crosscheck-oracle.json)
node scripts/crosscheck.ts   # writes docs/engine-validation.md

Rule authoring

Seed rules ship in @alloyra/data and are never mutated. Experts edit through a local overlay on the Failure rules page — edit/add/disable with structural validation (citations required) — and every comparison records the effective ruleset label (e.g. 2026.08.0+local(2)). Overlays export/import as JSON for sharing.

Run

pnpm install
pnpm dev      # workbench at http://localhost:3000
pnpm test     # calculator validation cases

Release boundary

Alloyra v0.1.0 is a research preview: appropriate for demonstration, materials education, preliminary screening, and failure-rule development. It is not appropriate for material qualification, procurement specification, code-compliance decisions, or safety-critical design approval. All 15 seed failure rules are drafts — excluded from audits by default, and promotable only through a named, dated review record. Outputs are screening guidance for expert judgment, never design approval.

Deployment (fly.io)

One Fly app; visitors install nothing and no compute runs server-side:

• alloyra — the workbench. pnpm build emits apps/web/out/ (~1.5 MB), served by nginx from a shared-cpu-1x / 256 MB machine that suspends when idle. No server-side code; all user state lives in the visitor's browser, and all CALPHAD computation runs in the visitor's tab. There is no hosted calculation endpoint to abuse or meter — the former alloyra-calphad bridge is retired (its code stays in services/calphad as the validation oracle / optional self-host).

fly deploy --ha=false --strategy bluegreen                    # workbench

Full details — files, assumptions, CSP configuration, local smoke tests — are in DEPLOY.md.

License and contributions

MIT — see LICENSE. Pull requests are welcome; changes to the failure-rule content get a careful review, so cite the standard or source behind every value. Forking is welcome too.

Data policy

Only redistributable values: standards' published minimums (spec-min) or clearly flagged estimated typicals. Nothing from licensed databases. Every value carries provenance and a citation; computed values always show their formula and validity window.

Study workflow and evidence coverage

The active study name, selected duty and shortlist follow the user between workbench pages. Saving a duty selects it for the active study; the home page links directly to screening and the seawater example. The Saved studies page creates, duplicates and switches independent studies. Existing single-study browser data migrates without removing the legacy recovery keys.

Comparison performance is a weighted summary of available criteria. Input coverage is reported separately and is not a confidence or validation percentage. Partial results retain their shortlist order; only candidates with complete requested criteria are automatically ranked. PREN is limited to stainless-family candidates across screening, comparison and studio. A single-temperature yield record contributes to the duty score or stress elimination only when its test temperature matches the duty temperature exactly. Other temperatures are evidence gaps; no tolerance, interpolation or extrapolation is assumed. Rule flags remain independent.

Strengthening fit and particle inputs start empty. Older stored parameters are preserved as unverified; entering or explicitly reviewing them enables applicable calculations. Changing the base grade requires renewed review. Matrix presets are labeled assumed until reviewed, and incompatible presets block Orowan output. Composition and LMP scenario input origins are shown in the studio and carried into portable bundles and readable reports.

Condition screening, trade-offs and portable records

Screening operates on all 29 alloy conditions rather than 27 grades. Property lookups, chart identities, rationale reports and comparison handoff preserve the exact condition; a missing property never falls back to another condition. The comparison includes reference-relative trade-offs and a sampled 0–2 weight sensitivity plot. Incomplete evidence suppresses preferred-candidate claims. Density controls, collapsible navigation, a keyboard-accessible evidence drawer, and three studio sections make detailed work easier to read.

A versioned alloyra-study JSON bundle contains study inputs, reference snapshots, rule overlays/history, score and screening snapshots, empirical outputs and recorded worker requests/responses. Import validates before an atomic storage write and creates a new study. A different reference snapshot remains available for report/export, with live calculation panes disabled to avoid silently mixing reference versions. Historical outputs are reviewed in Saved studies; live panes recompute on demand. Each result snapshot carries its own inputs.

Worker runs retain outputs on completion/error/cancellation, with periodic progress saves and the originating study ID. Browser termination can leave an incomplete record. Runs include solver/worker and TDB SHA-256 fingerprints; apps/web/scripts/engine-manifest.mjs refreshes them at build time. Validation coverage matches the archived 2026-08-29 report by database, normalized composition and exact temperature. Changed fingerprints suppress coverage. The baseline fingerprint associates the repository implementation with the archived report; it is not a new execution of the 52-case oracle battery. Do not refresh the baseline after solver changes without reviewing/regenerating validation evidence. The point battery does not establish trajectory or isopleth-boundary coverage.

Measured material records and verification plans

Material records accept manually entered measurements or UTF-8 CSV (2 MB). CSV columns are heat_id,uns,condition_id,form,thickness_mm,source,kind,name,value,unit,test_temp_c. Each row is one chemistry value (kind=chemistry, element symbol, unit=wt%, blank temperature) or test (kind=test, property ID, explicit temperature). Supported tests: yield_strength and tensile_strength in MPa, elongation in %, and density in g/cm³. Metadata repeats for each heat/condition. Quoted commas, newlines, escaped quotes, CRLF and BOM are supported. Blank values, inequalities, duplicate element/test points, inconsistent metadata and totals above 100% fail validation before anything is saved. Thickness may be explicitly unknown. Review the parsed record before saving; source labels are supplied by the user, not authenticated certificates. Corrections create a new immutable record.

Comparison selects one record per reference condition. Measured chemistry and tests replace the reference values for that candidate; missing values are not borrowed from the specification. Yield chooses an exact duty-temperature test when present. Measured PREN requires explicit Cr, Mo, N and W (zeros permitted); missing terms suppress the index and its score. Content-based rules preserve unknown elements, while predicates about specification limits retain the original reference specification separately. The existing mid-spec Scheil comparison is disabled for measured records: it must not infer missing chemistry or balance.

Verification actions derive from the active duty, shortlist, selected evidence and active rules. They request relevant tests/inputs, flag out-of-range measured chemistry, and identify mechanisms or draft assumptions needing review. Owners, status and resolution evidence are stored as append-only updates. A resolved item requires a note; changed underlying inputs reopen it for review. Checking an action does not alter data, scores, rule review status or qualification. Current plan snapshots, measured records and update history travel with the study bundle and readable report. Prior result/plan snapshots retain the inputs used when generated; visit the plan after changing inputs to refresh its snapshot.

Engineering decision records

The Decision record page saves a draft and records immutable revisions. Outcomes are a preferred candidate for further work, insufficient evidence to choose, or no suitable candidate in the current shortlist. Preference requires an included condition; every alternative (including exclusions) needs a disposition. Reviewer, rationale and unresolved questions are required. Reviewer identity is self-reported, not an authenticated signature or material approval.

Each record freezes study inputs, saved results, source references, measured records and verification history. Saved results retain their own calculation inputs; recording does not rerun them. Snapshots exclude decision drafts and history to avoid recursive growth. Later evidence changes produce a review notice; new revisions supersede earlier records without rewriting them. The acknowledgement must correspond to the evidence still present at save time. Portable imports validate frozen studies and reject recursive snapshots, dangling selections and incomplete alternative dispositions before writing storage. Records have individual HTML reports and are also included in the full study bundle/report.

Traceable 316L physical properties

Dataset 2026.09.1 adds five physical-property references for Supra 316L/4404 from Outokumpu's Supra datasheet, Table 7 (page 8), accessed 2026-09-07. The source replaces the generic 193 GPa modulus with its documented 200 GPa reference at 20 °C. Electrical resistivity is converted to the catalog unit; mean thermal expansion retains its 20–100 °C interval.

Expand a scalar property in the alloy database to inspect its source, temperature, conditions and limitations. Structured citations include the publication link, table locator, access date and pending expert-review status; they travel with study and decision snapshots. These published references remain ESTIMATED, not measured heat evidence or design allowables. Independent metallurgical review is still needed before this becomes an expert-reviewed dataset.