hea-bench 2.6.2__tar.gz → 2.7.2__tar.gz

{hea_bench-2.6.2 → hea_bench-2.7.2}/CHANGELOG.md

@@ -10,6 +10,162 @@ The format is loosely based on
 
 ## [Unreleased]
 
+## [2.7.2] — 2026-10-06
+
+### Fixed
+
+- The landing page said the Dataset tab holds 7,783 alloys, the size of
+  corpus v0.1.0. The tab has opened on v0.2.0, 10,290 alloys, since
+  2.7.0. A test now ties that number to the corpus the tab opens.
+- Documentation that still called corpus v0.2.0 opt-in, or called the
+  Chizhevskiy rows in group notation unparseable, now matches 2.7.1.
+  Those rows parse, and the loader keeps them out because v0.2.0 is
+  frozen.
+
+## [2.7.1] — 2026-10-06
+
+2.7.1 is the first published 2.7 release. GitHub refused the v2.7.0 tag
+push, so 2.7.0 reached only the web site. 2.7.1 contains everything
+listed under 2.7.0 plus the changes below, which answer an outside
+review of the app and the library.
+
+### Added
+
+- **Every function that takes a composition also takes a formula
+  string.** `hb.smix("CoCrFeMnNi")` raised `AttributeError` before and
+  now equals `hb.smix(hb.parse_formula("CoCrFeMnNi"))`. A test fails for
+  any public composition function that does not accept a formula.
+- **Group notation**, in all four parts. `(CoCrFeNi)95Al5` is 95 parts
+  equimolar CoCrFeNi plus 5 parts Al, the usual HEA convention, and
+  groups nest. Repeated elements add up, so `CoCrFeNiNi` is Ni at
+  40 at.%. The app now uses the library's parser, and a shared case list
+  keeps the Python and browser parsers identical.
+- **Rule notes.** Guo VEC flags a VEC within 0.2 of its 8.0 or 6.87
+  cutoff (`guo_vec.boundary_note`), where the published data scatter.
+  The verdict itself is unchanged, because the benchmark measures it.
+  Senkov κ names the pair it takes ΔH_IM from (`im_pair`) and adds a
+  note when that pair has an element under 10 at.%, because κ does not
+  weight the pair by amount.
+- **Composition search shows each candidate's predicted phase**, the
+  four-class prediction set, in the library, the MCP `design_search`
+  tool and the app. The app's example search now caps Al at 20 at.% and
+  uses 4 or 5 elements, so it no longer opens on alloys of 40 to 70%
+  Al.
+- `predict_phase_set` reports the coverage the coverage study measured
+  for corpus v0.1.0 at the chosen confidence, inside and outside the
+  dataset's range (`measured_coverage`).
+
+### Changed
+
+- **Sheikh ductility returns `not_applicable` for any alloy with an
+  element outside Ti, Zr, Hf, V, Nb, Ta, Cr, Mo and W**, the refractory
+  alloys the screen was built on. It used to give a verdict for alloys
+  such as CoCrFeMnNi.
+- The domain flag reads "Inside the dataset's range?" instead of
+  "Resembles the dataset?". Inside the range it gives the measured
+  coverage of the 90% sets on unseen alloy systems (87.5% for single
+  versus multi-phase, 82.9% for the structure), so a yes is not read as
+  a promise of accuracy.
+- **The desktop app starts its engine by itself.** The engine has always
+  been inside the exe. The app now also fetches the Peivaste source file
+  by itself the first time it is online, checks it and keeps it, so the
+  whole app works offline after that. That file cannot ship inside the
+  exe until its authors license it for redistribution. Until the app
+  has been online once, the hardness prediction still works, because it
+  needs only data inside the app, and the phase and dataset cards say
+  what they are waiting for. The website's Start button behaves the
+  same way.
+- The Benchmark tab leads with each model's two scores and the gap.
+- The app's footer says "Runs on this device" instead of "Offline",
+  which read like a warning.
+
+### Fixed
+
+- A permalink always opens in alloy mode, even when the page was last
+  in Oxides mode.
+- A formula that fails to parse clears the previous results instead of
+  leaving them up beside the error.
+- The app merges repeated elements (`CoCrFeNiNi`) instead of rejecting
+  them as duplicates.
+- Amounts with more than two decimals reach the result, its result ID
+  and the permalink unrounded. The table's 0.01 at.% display rounding
+  used to change them.
+- The calculator shows one melting-temperature card, in K and °C,
+  instead of two.
+- The release bot retries the tag push, which GitHub refused once for
+  v2.7.0.
+
+The experimental corpus is unchanged. The Chizhevskiy loader keeps rows
+written in group notation out, so corpus v0.2.0 stays byte-identical,
+and adding them is left for a future corpus version.
+
+## [2.7.0] — 2026-10-06
+
+### Added
+
+- **Every feature is now in all four parts**: the Python package, the
+  MCP server, the website and the desktop app. A person only needs to
+  pick one of them. This is a standing project rule from now on, and CI
+  enforces it (see the last item below).
+- **The MCP server has every feature of the package: 24 tools, up from
+  13.** New tools: `ceramic_report`, `element_data`, `corpus_build`,
+  `corpus_export`, `measured_properties`, `predict_phase_set`,
+  `benchmark_summary`, `benchmark_run`, `benchmark_folds`,
+  `benchmark_score` and `coverage_study`. `alloy_descriptors` now
+  returns the Miedema formation enthalpies, and `alloy_descriptors`,
+  `alloy_rules` and `omega_sensitivity` accept custom elements and pair
+  enthalpies. `campaign_suggest` also accepts a campaign inline. The
+  server moves to the official MCP Python SDK 2.x (`mcp>=2.3,<3`), and
+  the `mcp` extra now includes scikit-learn, so hardness, phase sets and
+  campaigns work from a registry install.
+- **The dataset works from a plain pip install.** The wheel ships the
+  openly licensed source datasets, and `hea_bench.corpus.build_corpus()`
+  downloads the one file that declares no license (Peivaste), checks its
+  SHA-256 and builds every corpus version in a per-user folder. The MCP
+  `corpus_build` tool does the same.
+- **Library functions for what only the apps had:**
+  `hea_bench.custom_data()` (your own elements and pair enthalpies, for
+  any descriptor or rule), `hea_bench.omega_sensitivity()`,
+  `hea_bench.descriptors.miedema_decomposition` and
+  `hea_bench.descriptors.elements.element_data()`.
+- `pip install "hea-bench[all]"` installs everything. Python 3.13 and
+  3.14 are now tested and declared.
+- New checks that keep the four parts equal.
+  `tests/test_feature_parity.py` (formerly `test_app_parity.py`) fails CI
+  unless every feature has its library function, its MCP tool, a working
+  app surface and tests. The CI `package` job builds the wheel, installs
+  it fresh and calls every MCP tool over stdio
+  (`tests/test_installed_package.py`, also run by the pre-push
+  preflight). The release `desktop-build` job uses every tab inside the
+  built exe before attaching it (`tests/test_desktop_smoke.py`). The
+  preflight also checks that the MCP container image receives every file
+  the wheel bundles.
+
+### Changed
+
+- **The app is much smoother.** It keeps its fitted models (hardness,
+  phase sets, the dataset check) in the browser between visits, so a
+  returning visitor gets predictions about a second after the engine
+  starts instead of waiting about 18 seconds for refits. The engine
+  prepares models in the background while the page is idle, starts by
+  itself for returning visitors, and long tables no longer stall tab
+  switches.
+- **The dataset opens on corpus v0.2.0, the largest (10,290 alloys)**,
+  in `load_corpus()`, the app's Dataset tab and the MCP corpus tools.
+  The benchmark, the phase predictions and the domain flag stay on the
+  reference corpus v0.1.0, so every published number is unchanged.
+- The in-app engine runs Pyodide 0.29.5 (was 0.29.3), the desktop app
+  uses Tauri 2.12.1 (was 2.11.2) without the unused logging plugin, the
+  MCP container image uses Python 3.13, and every GitHub Action is on
+  its current major version.
+
+### Fixed
+
+- The data-file fingerprints the app shows now match the ones the
+  library reports on every platform. They were computed from a Windows
+  checkout's line endings; they now use the bytes git stores. The data
+  itself is unchanged.
+
 ## [2.6.2] — 2026-10-06
 
 ### Added

{hea_bench-2.6.2 → hea_bench-2.7.2}/CITATION.cff

@@ -50,7 +50,7 @@ keywords:
   - alloy design
 license: MIT
 repository-code: 'https://github.com/dfieser/hea-bench'
-version: 2.6.2
+version: 2.7.2
 date-released: '2026-10-06'
 doi: 10.5281/zenodo.20346287
 identifiers:

{hea_bench-2.6.2 → hea_bench-2.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: hea-bench
-Version: 2.6.2
+Version: 2.7.2
 Summary: An open, interpretable calculator of high-entropy-alloy and high-entropy-oxide thermodynamic and geometric descriptors plus the canonical empirical phase-prediction rules.
 Project-URL: Homepage, https://dfieser.github.io/hea-bench/
 Project-URL: Repository, https://github.com/dfieser/hea-bench
@@ -19,19 +19,26 @@ Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
 Classifier: Topic :: Scientific/Engineering :: Chemistry
 Classifier: Topic :: Scientific/Engineering :: Physics
 Requires-Python: >=3.10
+Provides-Extra: all
+Requires-Dist: mcp<3,>=2.3; extra == 'all'
+Requires-Dist: scikit-learn~=1.7.2; extra == 'all'
 Provides-Extra: benchmark
 Requires-Dist: scikit-learn~=1.7.2; extra == 'benchmark'
 Provides-Extra: dev
+Requires-Dist: build>=1.2; extra == 'dev'
 Requires-Dist: pytest-cov>=4.0; extra == 'dev'
 Requires-Dist: pytest>=7.0; extra == 'dev'
 Requires-Dist: ruff~=0.15.0; extra == 'dev'
 Provides-Extra: interop
 Requires-Dist: heacalculator~=2.0.1; extra == 'interop'
 Provides-Extra: mcp
-Requires-Dist: mcp<2,>=1.9.4; extra == 'mcp'
+Requires-Dist: mcp<3,>=2.3; extra == 'mcp'
+Requires-Dist: scikit-learn~=1.7.2; extra == 'mcp'
 Provides-Extra: properties
 Requires-Dist: scikit-learn~=1.7.2; extra == 'properties'
 Description-Content-Type: text/markdown
@@ -90,10 +97,13 @@ For any composition it reports:
   Yang–Zhang Ω, Pauling electronegativity mismatch Δχ, Mansoori excess
   entropy S<sub>E</sub>, ΔG<sub>ss</sub>, ΔG<sub>max</sub>, King Φ, Ye φ.
 - **Phase-prediction rules:** Yeh entropy, Zhang δ, Guo–Liu VEC,
-  Yang–Zhang Ω, King Φ, Ye φ.
-- **Miedema formation enthalpies** (browser/desktop apps): compound /
-  solid-solution / amorphous, decomposed into chemical, elastic,
-  structural, and topological terms.
+  Yang–Zhang Ω, King Φ, Ye φ, Senkov–Miracle κ, the Tsai σ-phase
+  window and Sheikh ductility.
+- **Miedema formation enthalpies:** compound / solid-solution /
+  amorphous, decomposed into chemical, elastic, structural, and
+  topological terms.
+- **Your own elements and pair values:** any descriptor or rule, for a
+  composition with an element or pair enthalpy that is not tabulated.
 - **High-entropy oxides** (`hea_bench.oxides` + the apps' Oxides mode):
   rock-salt, perovskite, fluorite, and pyrochlore formability
   descriptors over Shannon ionic radii with automatic charge-balance
@@ -125,16 +135,27 @@ flowchart LR
     D --> E
 ```
 
-## Four ways to run it
+## Four ways to run it, each with every feature
 
-| Surface | Where | Status |
+| Part | Where | Best for |
 |---|---|---|
-| **Python library + CLI** | `pip install hea-bench` | done, tested |
-| **Zero-install browser app** | <https://dfieser.github.io/hea-bench/> · `web/index.html` | done, Python-parity-tested |
-| **Native desktop app** | a single portable `.exe`, [download (no install)](https://github.com/dfieser/hea-bench/releases/latest/download/HEA-Bench.exe) (Tauri wrapper of the same page) | done, built from the same parity-tested core |
-| **MCP server for AI agents** | `pip install "hea-bench[mcp]"`, then `hea-bench-mcp` | done, thirteen tools over the same core |
-
-The three surfaces share **one calculation core**. The browser/desktop
+| **Web app** | <https://dfieser.github.io/hea-bench/> | anyone, no install |
+| **Desktop app** | one portable `.exe`, [download (no install)](https://github.com/dfieser/hea-bench/releases/latest/download/HEA-Bench.exe), the same app in its own window, engine inside, works offline | Windows users who want it local |
+| **Python library + CLI** | `pip install "hea-bench[all]"` | scripts and notebooks |
+| **MCP server for AI agents** | `pip install "hea-bench[mcp]"`, then `hea-bench-mcp` | Claude, Cursor and other MCP clients, as 24 tools |
+
+Every feature is in all four parts, so pick whichever suits you and you
+lose nothing. That is a standing project rule, and CI enforces it:
+`tests/test_feature_parity.py` fails whenever a public feature lacks
+its library function, its MCP tool or a working app surface, a headless
+browser uses every tab of the built site, the built desktop exe goes
+through the same steps before it is attached to a release, and a
+freshly installed wheel answers every MCP tool. `pip install
+hea-bench` alone has no dependencies and covers everything except the
+fitted models and the MCP server; the `[all]` extra adds scikit-learn
+and the MCP SDK for those.
+
+All four share **one calculation core**. The browser/desktop
 core (`web/hea-calculator-core.js`) is a pure-JS port of the Python
 library, and `tests/test_web_parity.py` guarantees the two match on all
 1484 binary pairs and the canonical multi-element fixtures, while
@@ -143,10 +164,7 @@ down to identical warning messages. The app's Dataset, Design and
 Benchmark tabs and its predictions panel run the Python package itself,
 unchanged, inside the page (Pyodide, `web/hea-engine-worker.js`), and
 `tests/test_web_engine.py` checks that it returns exactly what CPython
-returns. `tests/test_app_parity.py` fails CI whenever a public library
-feature has no working app surface, so the app never falls behind the
-library, and before every release a headless browser uses every tab of
-the built site the way a person would.
+returns.
 
 ## Quick start (Python)
 
@@ -260,18 +278,22 @@ ceramics datasets: [docs/ceramics.md](docs/ceramics.md).
 ## Quick start (AI agents, MCP)
 
 LLM agents hallucinate descriptor values; this server grounds them.
-`hea_bench.mcp_server` exposes the whole workflow over the
-[Model Context Protocol](https://modelcontextprotocol.io/) as thirteen
-deterministic tools: the original calculator seven
-(`parse_composition`, batch `alloy_descriptors` and `alloy_rules`,
-`omega_sensitivity`, `oxide_report`, `element_coverage`, `about`) plus
-the capability layers (`corpus_query` and `corpus_describe` over the
-provenance-tracked experimental corpus, `predict_properties` with
-intervals and domain flags at the top level of every payload,
-`check_applicability` for the novelty components, `design_search` with
-hard caps on palette, step, and candidate count, and
-`campaign_suggest` operating on a campaign file the user supplies).
-Every response carries units or uncertainty fields, citation keys
+`hea_bench.mcp_server` exposes every feature of the package over the
+[Model Context Protocol](https://modelcontextprotocol.io/) as 24
+deterministic tools. The calculator: `parse_composition`, batch
+`alloy_descriptors` (with the Miedema enthalpies) and `alloy_rules`,
+both taking custom elements and pair values, `omega_sensitivity`,
+`oxide_report`, `ceramic_report`, `element_data` and
+`element_coverage`. The dataset: `corpus_build` (once, to build the
+corpus locally), `corpus_query`, `corpus_describe`, `corpus_export` and
+`measured_properties`. Predictions: `predict_properties` with
+intervals and `predict_phase_set` with prediction sets, both carrying
+the domain flag, and `check_applicability`. Design:
+`design_search`, with hard caps on palette, step and candidate count,
+and `campaign_suggest`, on a campaign file or an inline campaign. The
+benchmark: `benchmark_summary`, `benchmark_run`, `benchmark_folds`,
+`benchmark_score` and `coverage_study`. And `about`. Every response
+carries units or uncertainty fields, citation keys
 where a parametrization is involved, and the library version, so an
 agent's reasoning trace contains auditable receipts rather than bare
 floats; `about()` reports which capabilities are available in the
@@ -358,10 +380,13 @@ Li et al., *Commun. Mater.* **6**:9 (2025),
 Baselines, split digests, and full provenance:
 [`docs/benchmark-baselines.md`](./docs/benchmark-baselines.md).
 
-This surface currently works from a repository checkout only: the
-corpus's largest source dataset declares no license, so the corpus is
-rebuilt locally from a fetch script and pinned hashes rather than
-redistributed (see [`data/raw/README.md`](./data/raw/README.md)).
+The corpus is built on your machine, once, because its largest source
+dataset declares no license and is never redistributed (see
+[`data/raw/README.md`](./data/raw/README.md)).
+`hea_bench.corpus.build_corpus()` downloads that file from its authors'
+repository, accepts it only if its pinned SHA-256 matches, and builds
+every corpus version. The apps and the MCP server's `corpus_build` tool
+do the same.
 
 ## The corpus as a standalone product
 
@@ -371,7 +396,7 @@ addressable directly, with no task or split machinery involved:
 ```python
 from hea_bench.corpus import load_corpus
 
-corpus = load_corpus()                # v0.1.0, every row, full provenance
+corpus = load_corpus()                # v0.2.0 (10,290 alloys), every row, full provenance
 corpus.describe()                     # counts, families, agreement rate
 al_bcc = corpus.query(contains=["Al"], phase="BCC", descriptor_ready=True)
 al_bcc.rows[0].raw_labels             # each source's verbatim reported phase
@@ -383,8 +408,11 @@ processing route and primary-literature DOI where available, and
 upstream record identifiers, so a label can be audited without leaving
 the package. Provenance chains, per-source license status,
 harmonization rules, and known limitations are documented in the
-[corpus card](docs/corpus-card.md). The corpus data is still built
-locally from the recipe above, for the same licensing reason.
+[corpus card](docs/corpus-card.md). `load_corpus()` opens the largest
+version, v0.2.0, which adds the LLM-extracted Chizhevskiy database.
+`load_corpus(version="0.1.0")` is the hand-curated reference (7,783
+alloys) that the benchmark, the phase predictions and the domain flag
+always use, so every published number stays put.
 
 ## Uncertainty and domain of applicability
 
@@ -398,7 +426,7 @@ guarantee's exchangeability assumption plausibly holds for your query:
 from hea_bench.corpus import load_corpus
 from hea_bench.uncertainty import ConformalClassifier, fit_domain
 
-domain = fit_domain(load_corpus())
+domain = fit_domain(load_corpus(version="0.1.0"))
 domain.novelty({"Hf": 0.2, "Nb": 0.2, "Ta": 0.2, "Ti": 0.2, "Zr": 0.2})
 # {'element_set_seen': True, 'family_count': ..., 'nearest_family_distance': 0.0,
 #  'descriptor_distance': ..., 'element_coverage': True, 'in_domain': True, ...}
@@ -517,11 +545,18 @@ the number does not, so read Ω qualitatively in that regime.
 hea-bench/
 ├── src/hea_bench/
 │   ├── descriptors/     ΔS_mix, δ, VEC, T_m, ΔH_mix, Ω, S_E, φ + data tables
-│   ├── rules/           the six empirical phase-prediction rules
+│   ├── rules/           the nine empirical phase-prediction rules
 │   ├── oxides/          HEO module: families, oxidation-state solver,
 │   │                    Shannon radii (94 elements, vendored from pymatgen)
+│   ├── ceramics/        carbides, nitrides, diborides
+│   ├── corpus/          the experimental corpus: build_corpus, load_corpus
 │   ├── benchmark/       frozen family-grouped + random paired splits and evaluation
-│   │                    (repo-only; corpus is built locally, see data/raw/)
+│   ├── properties/      density, melting point, cost, hardness
+│   ├── uncertainty/     conformal sets and intervals, phase sets, domain flag
+│   ├── design/          composition search and campaigns
+│   ├── mcp_server.py    the MCP server (hea-bench-mcp)
+│   ├── webapp.py        the bridge the in-page engine calls
+│   ├── custom.py        custom elements and pair values
 │   ├── composition.py   formula parser, normalizer
 │   ├── constants.py     R = 8.314
 │   └── cli.py           command-line entry point

{hea_bench-2.6.2 → hea_bench-2.7.2}/README.md

@@ -52,10 +52,13 @@ For any composition it reports:
   Yang–Zhang Ω, Pauling electronegativity mismatch Δχ, Mansoori excess
   entropy S<sub>E</sub>, ΔG<sub>ss</sub>, ΔG<sub>max</sub>, King Φ, Ye φ.
 - **Phase-prediction rules:** Yeh entropy, Zhang δ, Guo–Liu VEC,
-  Yang–Zhang Ω, King Φ, Ye φ.
-- **Miedema formation enthalpies** (browser/desktop apps): compound /
-  solid-solution / amorphous, decomposed into chemical, elastic,
-  structural, and topological terms.
+  Yang–Zhang Ω, King Φ, Ye φ, Senkov–Miracle κ, the Tsai σ-phase
+  window and Sheikh ductility.
+- **Miedema formation enthalpies:** compound / solid-solution /
+  amorphous, decomposed into chemical, elastic, structural, and
+  topological terms.
+- **Your own elements and pair values:** any descriptor or rule, for a
+  composition with an element or pair enthalpy that is not tabulated.
 - **High-entropy oxides** (`hea_bench.oxides` + the apps' Oxides mode):
   rock-salt, perovskite, fluorite, and pyrochlore formability
   descriptors over Shannon ionic radii with automatic charge-balance
@@ -87,16 +90,27 @@ flowchart LR
     D --> E
 ```
 
-## Four ways to run it
+## Four ways to run it, each with every feature
 
-| Surface | Where | Status |
+| Part | Where | Best for |
 |---|---|---|
-| **Python library + CLI** | `pip install hea-bench` | done, tested |
-| **Zero-install browser app** | <https://dfieser.github.io/hea-bench/> · `web/index.html` | done, Python-parity-tested |
-| **Native desktop app** | a single portable `.exe`, [download (no install)](https://github.com/dfieser/hea-bench/releases/latest/download/HEA-Bench.exe) (Tauri wrapper of the same page) | done, built from the same parity-tested core |
-| **MCP server for AI agents** | `pip install "hea-bench[mcp]"`, then `hea-bench-mcp` | done, thirteen tools over the same core |
-
-The three surfaces share **one calculation core**. The browser/desktop
+| **Web app** | <https://dfieser.github.io/hea-bench/> | anyone, no install |
+| **Desktop app** | one portable `.exe`, [download (no install)](https://github.com/dfieser/hea-bench/releases/latest/download/HEA-Bench.exe), the same app in its own window, engine inside, works offline | Windows users who want it local |
+| **Python library + CLI** | `pip install "hea-bench[all]"` | scripts and notebooks |
+| **MCP server for AI agents** | `pip install "hea-bench[mcp]"`, then `hea-bench-mcp` | Claude, Cursor and other MCP clients, as 24 tools |
+
+Every feature is in all four parts, so pick whichever suits you and you
+lose nothing. That is a standing project rule, and CI enforces it:
+`tests/test_feature_parity.py` fails whenever a public feature lacks
+its library function, its MCP tool or a working app surface, a headless
+browser uses every tab of the built site, the built desktop exe goes
+through the same steps before it is attached to a release, and a
+freshly installed wheel answers every MCP tool. `pip install
+hea-bench` alone has no dependencies and covers everything except the
+fitted models and the MCP server; the `[all]` extra adds scikit-learn
+and the MCP SDK for those.
+
+All four share **one calculation core**. The browser/desktop
 core (`web/hea-calculator-core.js`) is a pure-JS port of the Python
 library, and `tests/test_web_parity.py` guarantees the two match on all
 1484 binary pairs and the canonical multi-element fixtures, while
@@ -105,10 +119,7 @@ down to identical warning messages. The app's Dataset, Design and
 Benchmark tabs and its predictions panel run the Python package itself,
 unchanged, inside the page (Pyodide, `web/hea-engine-worker.js`), and
 `tests/test_web_engine.py` checks that it returns exactly what CPython
-returns. `tests/test_app_parity.py` fails CI whenever a public library
-feature has no working app surface, so the app never falls behind the
-library, and before every release a headless browser uses every tab of
-the built site the way a person would.
+returns.
 
 ## Quick start (Python)
 
@@ -222,18 +233,22 @@ ceramics datasets: [docs/ceramics.md](docs/ceramics.md).
 ## Quick start (AI agents, MCP)
 
 LLM agents hallucinate descriptor values; this server grounds them.
-`hea_bench.mcp_server` exposes the whole workflow over the
-[Model Context Protocol](https://modelcontextprotocol.io/) as thirteen
-deterministic tools: the original calculator seven
-(`parse_composition`, batch `alloy_descriptors` and `alloy_rules`,
-`omega_sensitivity`, `oxide_report`, `element_coverage`, `about`) plus
-the capability layers (`corpus_query` and `corpus_describe` over the
-provenance-tracked experimental corpus, `predict_properties` with
-intervals and domain flags at the top level of every payload,
-`check_applicability` for the novelty components, `design_search` with
-hard caps on palette, step, and candidate count, and
-`campaign_suggest` operating on a campaign file the user supplies).
-Every response carries units or uncertainty fields, citation keys
+`hea_bench.mcp_server` exposes every feature of the package over the
+[Model Context Protocol](https://modelcontextprotocol.io/) as 24
+deterministic tools. The calculator: `parse_composition`, batch
+`alloy_descriptors` (with the Miedema enthalpies) and `alloy_rules`,
+both taking custom elements and pair values, `omega_sensitivity`,
+`oxide_report`, `ceramic_report`, `element_data` and
+`element_coverage`. The dataset: `corpus_build` (once, to build the
+corpus locally), `corpus_query`, `corpus_describe`, `corpus_export` and
+`measured_properties`. Predictions: `predict_properties` with
+intervals and `predict_phase_set` with prediction sets, both carrying
+the domain flag, and `check_applicability`. Design:
+`design_search`, with hard caps on palette, step and candidate count,
+and `campaign_suggest`, on a campaign file or an inline campaign. The
+benchmark: `benchmark_summary`, `benchmark_run`, `benchmark_folds`,
+`benchmark_score` and `coverage_study`. And `about`. Every response
+carries units or uncertainty fields, citation keys
 where a parametrization is involved, and the library version, so an
 agent's reasoning trace contains auditable receipts rather than bare
 floats; `about()` reports which capabilities are available in the
@@ -320,10 +335,13 @@ Li et al., *Commun. Mater.* **6**:9 (2025),
 Baselines, split digests, and full provenance:
 [`docs/benchmark-baselines.md`](./docs/benchmark-baselines.md).
 
-This surface currently works from a repository checkout only: the
-corpus's largest source dataset declares no license, so the corpus is
-rebuilt locally from a fetch script and pinned hashes rather than
-redistributed (see [`data/raw/README.md`](./data/raw/README.md)).
+The corpus is built on your machine, once, because its largest source
+dataset declares no license and is never redistributed (see
+[`data/raw/README.md`](./data/raw/README.md)).
+`hea_bench.corpus.build_corpus()` downloads that file from its authors'
+repository, accepts it only if its pinned SHA-256 matches, and builds
+every corpus version. The apps and the MCP server's `corpus_build` tool
+do the same.
 
 ## The corpus as a standalone product
 
@@ -333,7 +351,7 @@ addressable directly, with no task or split machinery involved:
 ```python
 from hea_bench.corpus import load_corpus
 
-corpus = load_corpus()                # v0.1.0, every row, full provenance
+corpus = load_corpus()                # v0.2.0 (10,290 alloys), every row, full provenance
 corpus.describe()                     # counts, families, agreement rate
 al_bcc = corpus.query(contains=["Al"], phase="BCC", descriptor_ready=True)
 al_bcc.rows[0].raw_labels             # each source's verbatim reported phase
@@ -345,8 +363,11 @@ processing route and primary-literature DOI where available, and
 upstream record identifiers, so a label can be audited without leaving
 the package. Provenance chains, per-source license status,
 harmonization rules, and known limitations are documented in the
-[corpus card](docs/corpus-card.md). The corpus data is still built
-locally from the recipe above, for the same licensing reason.
+[corpus card](docs/corpus-card.md). `load_corpus()` opens the largest
+version, v0.2.0, which adds the LLM-extracted Chizhevskiy database.
+`load_corpus(version="0.1.0")` is the hand-curated reference (7,783
+alloys) that the benchmark, the phase predictions and the domain flag
+always use, so every published number stays put.
 
 ## Uncertainty and domain of applicability
 
@@ -360,7 +381,7 @@ guarantee's exchangeability assumption plausibly holds for your query:
 from hea_bench.corpus import load_corpus
 from hea_bench.uncertainty import ConformalClassifier, fit_domain
 
-domain = fit_domain(load_corpus())
+domain = fit_domain(load_corpus(version="0.1.0"))
 domain.novelty({"Hf": 0.2, "Nb": 0.2, "Ta": 0.2, "Ti": 0.2, "Zr": 0.2})
 # {'element_set_seen': True, 'family_count': ..., 'nearest_family_distance': 0.0,
 #  'descriptor_distance': ..., 'element_coverage': True, 'in_domain': True, ...}
@@ -479,11 +500,18 @@ the number does not, so read Ω qualitatively in that regime.
 hea-bench/
 ├── src/hea_bench/
 │   ├── descriptors/     ΔS_mix, δ, VEC, T_m, ΔH_mix, Ω, S_E, φ + data tables
-│   ├── rules/           the six empirical phase-prediction rules
+│   ├── rules/           the nine empirical phase-prediction rules
 │   ├── oxides/          HEO module: families, oxidation-state solver,
 │   │                    Shannon radii (94 elements, vendored from pymatgen)
+│   ├── ceramics/        carbides, nitrides, diborides
+│   ├── corpus/          the experimental corpus: build_corpus, load_corpus
 │   ├── benchmark/       frozen family-grouped + random paired splits and evaluation
-│   │                    (repo-only; corpus is built locally, see data/raw/)
+│   ├── properties/      density, melting point, cost, hardness
+│   ├── uncertainty/     conformal sets and intervals, phase sets, domain flag
+│   ├── design/          composition search and campaigns
+│   ├── mcp_server.py    the MCP server (hea-bench-mcp)
+│   ├── webapp.py        the bridge the in-page engine calls
+│   ├── custom.py        custom elements and pair values
 │   ├── composition.py   formula parser, normalizer
 │   ├── constants.py     R = 8.314
 │   └── cli.py           command-line entry point

{hea_bench-2.6.2 → hea_bench-2.7.2}/data/raw/README.md

@@ -36,7 +36,7 @@ date and neither repository declared a license.
 | [`borg2020/`](borg2020/) | Borg *et al.* 2020 (*Sci. Data* 7, 430) | **CC-BY-4.0** (figshare deposit explicitly licensed) | ✓ **Mirrored 2026-05-20.** 1,545 measurement rows / 740 unique (formula, processing) alloys / 23 columns including phase, processing, mechanical properties, and per-row source DOI. The canonical backbone of the benchmark. |
 | [`pei2020/`](pei2020/) | Pei *et al.* 2020 (*npj Comput. Mater.* 6, 50) | **CC-BY-4.0** (Crossref-confirmed, `tdm` and `vor`, zero delay) | ✓ **Mirrored 2026-05-20.** 1,252 rows, 2 columns (Alloy, Phase), four phase labels (bcc/fcc/hcp/multi-phase). 1,208 unique formulas load after deduplicating 43 repeated formulas and rejecting the malformed `Ta_Th` entry at row 1251. |
 | [`peivaste/`](peivaste/) | Iman-Peivaste/ML_HEAs_Phase_Dataset (GitHub) + companion Sci. Rep. 13, 22556 (2023) by same authors | **None.** No LICENSE file on the data repository, re-confirmed 2026-08-10. The *Sci. Rep.* article is CC-BY, the GitHub dataset is not covered by it. | **Pointer-only.** `fetch.py` provided, CSV gitignored. Largest single contributor to the corpus, which is why the corpus as a whole cannot be redistributed. |
-| [`chizhevskiy2026/`](chizhevskiy2026/) | Chizhevskiy *et al.* 2026 (*Sci. Data* 13, 612), repo Vladimirchizh/hea_database | **CC-BY-4.0**, resolved 2026-08-10: the maintainer merged [hea_database#2](https://github.com/Vladimirchizh/hea_database/pull/2) adding a LICENSE file (GitHub license detection confirms CC-BY-4.0). | ✓ **Mirrored 2026-08-10**, byte-identical to upstream at the pinned SHA-256. 12,427 LLM-extracted records, 2,974 unique parseable alloys. Feeds corpus **v0.2.0 only**; v0.1.0 is frozen without it. Its labels agree with the v0.1.0 consensus on only ~70% of overlapping alloys, so disagreements are quarantined as conflicts and corpus v0.2.0 is opt-in by design; v0.1.0 stays the reference corpus. |
+| [`chizhevskiy2026/`](chizhevskiy2026/) | Chizhevskiy *et al.* 2026 (*Sci. Data* 13, 612), repo Vladimirchizh/hea_database | **CC-BY-4.0**, resolved 2026-08-10: the maintainer merged [hea_database#2](https://github.com/Vladimirchizh/hea_database/pull/2) adding a LICENSE file (GitHub license detection confirms CC-BY-4.0). | ✓ **Mirrored 2026-08-10**, byte-identical to upstream at the pinned SHA-256. 12,427 LLM-extracted records, 2,974 unique alloys in plain formulas (rows in group notation stay out, because v0.2.0 is frozen). Feeds corpus **v0.2.0 only**; v0.1.0 is frozen without it. Its labels agree with the v0.1.0 consensus on only ~70% of overlapping alloys, so disagreements are quarantined as conflicts. Browsing opens on v0.2.0 since 2.7.0, and v0.1.0 stays the reference corpus for the benchmark, the phase predictions and the domain flag. |
 | [`couzinie2018/`](couzinie2018/) | Couzinié *et al.* 2018 (*Data in Brief* 21, 1622) | CC-BY-4.0 | Acquisition stalled. Elsevier CDN supplementary files are mis-attributed, and the content is mostly subsumed by Borg anyway. **Deprioritized.** See README for log. |
 | (planned) `miracle_senkov2017/` | Miracle & Senkov 2017 (*Acta Mater.* 122, 448) | Elsevier, to be determined | Not yet attempted |
 | (planned) `witman/` | mwitman1/HEAphaseML (GitHub) | To be determined | Not yet attempted |