# ORCHID ESPEI datasets from the literature

ESPEI-format datasets built from the ORCHID literature values (Version 1, machine-extracted) and from DFT (literature, Materials Project, OQMD). Every literature value is traced to its published ORCHID observation id (`literature/`, field `id`) and the DOI of its source, passes the same inclusion rules as the published literature values, and carries its uncertainty in physical units next to the ESPEI weight derived from it.

| folder | datasets | values | systems | content |
|---|---:|---:|---:|---|
| experimental | 1118 | 2387 | 185 | experimental thermochemical tables and phase-composition (ZPF) tables |
| model | 504 | 782 | 82 | values the papers calculated with a CALPHAD model, TDB (SHA-256) named where known |
| invariants | 1546 | 5061 | 464 | ZPF datasets from experimental invariant-reaction tables |
| dft/literature | 783 | 1781 | 91 | literature DFT formation enthalpies, HM_FORM endmember datasets (`solver`) |
| dft/mp | 1251 | 1251 | 381 | Materials Project on-hull formation energies, HM_FORM endmember datasets |
| dft/oqmd | 1251 | 1251 | 381 | OQMD on-hull formation energies, HM_FORM endmember datasets |

Optional phase constraints (separate folders; never mixed with measured data; see "Phase constraints"):

| folder | datasets | ZPF points | systems | content |
|---|---:|---:|---:|---|
| assessed | 643 | 8007 | 471 | assessed (handbook) phase-field points, Dataset assessed-phase-assemblages, in the phase names of each TDB that has ZPF datasets here (tag `assessed_phase_diagram`); the dense field grid of the validated diagrams (Cu-Zr) is a separate dataset per TDB, suffix `.dense.json`, tag `dense_field_grid` |
| derived/invariant-brackets | 1072 | 4742 | 390 | DERIVED: phase sets just above and below each printed invariant (tag `derived_bracket`) |
| derived/liquid-stability | 421 | 3789 | 305 | DERIVED: single-phase liquid above the highest liquidus where the sources show complete liquid miscibility (tag `derived_liquid_stability`) |

## Use with ESPEI

Point ESPEI at one evidence class at a time; classes are never mixed in a folder, and for DFT use **one** of `dft/mp`, `dft/oqmd`, `dft/literature` per fit (MP uses corrected, OQMD fitted elemental references).

```yaml
# espei-in.yaml
system:
  phase_models: my-phases.json
  datasets: orchid-espei/experimental      # or invariants/, dft/mp/, ... (separately, by choice)
generate_parameters:
  excess_model: linear
  ref_state: SGTE91
```

```
espei --input espei-in.yaml
```

or in Python (ESPEI 0.9.1):

```python
from espei.datasets import load_datasets, recursive_glob, apply_tags
ds = load_datasets(recursive_glob('espei/experimental/AG-AL') + recursive_glob('espei/invariants/AG-AL'))
# your choices, e.g. halve the weight of single-reader values, drop model values:
apply_tags(ds, {'single_reader': {'weight': 0.5}, 'model': {'disabled': True}})
```

`apply_tags` replaces the dataset's `weight`; to scale the weight derived from the printed sigma, read `orchid_uncertainty` and set `weight` yourself. `index.csv` / `index.json` list every dataset (system, output, phases, values, temperature range, evidence class, DOIs, observation ids, reader agreement, sigma summary, weight).

Tags: evidence class (`experimental`, `model`, `dft_literature`, `dft_mp`, `dft_oqmd`), quality (`single_reader`, `single_reader_text_layer`, `no_printed_uncertainty`, `polymorph_ambiguous`, `above_hull`, `orchid_caveat`), plus the corpus tags (`thermochemical_table`, `invariant_table`, `via_assessment`, ...).

To restrict a fit to evidence certified against a CALPHAD prediction (`calphad_certified`: every value of the dataset is `espei_certified` in the literature release, i.e. espei_ready and its pycalphad calculation from the linked TDB agrees within 25 K / 0.05; per-value detail in `orchid_certification`):

```yaml
system:
  phase_models: my-phases.json
  datasets: orchid-espei/experimental
  tags:
    orchid: {disabled: true}             # every ORCHID dataset (literature and DFT) off ...
    calphad_certified: {disabled: false}  # ... except the certified ones (tags apply in this order)
```

(ESPEI applies `tags` before it removes disabled datasets, so the second tag re-enables.)

## Inclusion and temperatures (cleanup 2026-10-24)

- Every value the literature release marks not ESPEI-ready is left out: experimental and DFT values must be `espei_ready`; model values (never flagged espei_ready, the flag is about measured data) are left out when a literature gate marks them `not_espei_ready` or `not_accepted` (`not_espei_ready:literature_gate`).
- 298 K is 298.15 K: a dataset temperature of 298 / 298.1 / 298.2 K is the standard reference temperature and is written as 298.15 K (the literature records store 298.15 too, rule `standard_reference_temperature`), so ESPEI's `filter_temperatures` (non-equilibrium values below 298.15 K) removes nothing; a non-equilibrium value really below 298.15 K is left out and listed (`below_espei_temperature_floor`).
- One published observation supplies at most one value per folder (`duplicate_of_other_dataset`).
- Method-based uncertainty floors (PI decision 2026-10-24): every value's weight follows `sigma_effective` of its published uncertainty budget, which is never below the method-class floor calibrated from ORCHID's independent measurements (or a documented prior) and never below ESPEI's experimental default; printed values and printed sigmas are unchanged. This replaces the former resolution floor (a printed sigma below 0.1 x the default used the default) and the opt-in `orchid_weight_floor` suggestion.
- `experimental/MG-O/MgO-18Lia.zpf-00.json`, `invariants/PD-TI/PdTi-11Guo.zpf-00.json` and `invariants/AL-S/AlS-19Shi.zpf-01.json` are left out: their pycalphad driving-force evaluation is not reproducible (`pycalphad_evaluation_unstable`).
- A compound whose independent measured formation enthalpies disagree by more than their printed sigmas retains its reported sigma in `uncertainty` and a separate scatter estimate (sample standard deviation of the independent values; page-checked); it is the budget's `sigma_scatter` and replaces sigma_effective when larger.
- A formation enthalpy that the source measured at the historical formula of a phase (e.g. Cu4Zr or Cu9Zr2 for Cu5Zr, Cu3Zr for Cu51Zr14, Cu3Zr2 for Cu10Zr7, as the source table itself attributes them) is used for the modern phase; the offset |dx| x local hull slope is the budget's `sigma_composition`, added in quadrature (the reported sigma stays in `uncertainty`; estimate provenance is in the value budget; the flag `stoichiometry_mismatch` stays, flag-only).
- A value printed relative to a non-SER reference phase that the page states (e.g. bcc-Zr at 1123 K) is written converted to the TDB's SER phases at the same temperature with the SGTE unary lattice stability (`orchid_value_note`; the record's `corrections` entry gives the conversion and keeps the printed value); the lattice-stability uncertainty is the budget's `sigma_conversion` (prior: 10 %% of the lattice stability times x).

## Phase assignment

Physical phase rule (tools/phase_rule.py): a literature value is assigned to a phase only when the paper names that phase (printed phase label of the value; when no label is printed, a table caption that names the liquid) and the phase registry maps the label to a model phase; no default phase is assigned. A liquid assignment must be physically possible: not below the lowest liquid temperature of any calculated SGTE diagram of the system minus max(100 K, 10 %) (twice that for printed invariant and phase-composition records; no calculated diagram: not below 0.40 x the lowest element melting point). Formation enthalpies and entropies of a stoichiometric compound go to that compound's TDB endmember (ESPEI non-equilibrium dataset, `solver`); other properties of a compound are not used in ESPEI. A solution-phase assignment outside the phase's field in the calculated diagram of the target TDB is flagged (`outside_calculated_field`), not excluded. Temperatures: T <= 1 K is not a measured state; a printed phase equilibrium (invariant, phase composition, transition) below 0.25 x the lowest element melting point is a misreading. A value flagged `unsigned_in_print` / `magnitude_suspect` / `system_mismatch` / `source_disputed` is not ESPEI-ready.

Every value's phase comes from its published literature record (printed label -> phase registry -> the target TDB's phase name), never from a default. Values with different phases in one source table are written as separate datasets (`<TDB>.<file>-<PHASES>.json`). Formation enthalpies / entropies of a phase that is stoichiometric in the target TDB are written as ESPEI non-equilibrium datasets (`solver`, the compound's endmember, output `HM_FORM` / `SM_FORM`, tag `compound_formation_endmember`) when the printed composition matches the formula and the printed reference state is the TDB's SER; other properties of compounds are not written. `orchid_phase_assignment` gives the phases, the basis (printed labels or caption) and, flag only, values outside the phase's field in the calculated diagram of the target TDB (`outside_calculated_field`). Also excluded (VALIDATION.md, `drop_reasons`): activities outside (0, 1], negative absolute entropies, increments printed as absolute values, per-gram heat capacities, formation / mixing values outside physical ranges, values flagged unsigned_in_print / magnitude_suspect / system_mismatch / source_disputed, reference phases that the TDB does not have, binary three-phase ZPF regions away from every calculated invariant, ZPF regions with a gas and no printed pressure, and ZPF regions that carry no composition and no hyperplane (zero driving force in ESPEI 0.9.1).

## Uncertainties

ORCHID supplies uncertainties to ESPEI 0.9.1 through weights: the sigma it uses is a reference sigma divided by the dataset's `weight` (500 J/mol for HM, 0.2 J/K/mol for SM and CPM, 500 J/mol on the chemical potential for activities; ESPEI: "estimated from NIST TRC uncertainties", i.e. for experiments). Every value carries its **uncertainty budget** in `orchid_uncertainty.values[].budget`, in the dataset's units, with the components kept separate so that you can recombine them: `sigma_printed` (as printed), `sigma_reading` (half of the last printed digit), `sigma_method` (method-class floor, with `method_class`, `method_basis`, `method_n_pairs`, `method_entry` into `experimental/CALIBRATION_EXPERIMENT.json`), `sigma_composition` (value measured at a historical formula), `sigma_scatter` (page-checked mutually inconsistent group), `sigma_conversion` (reference-state conversion) and `sigma_effective`. Source-reported values and sigmas are never altered; the budget comes from the literature release (`uncertainty_budget`).

**Rule** (`sigma_rule`): sigma_effective = max( sqrt( max(sigma_printed, sigma_reading, sigma_method)^2 + sigma_composition^2 + sigma_conversion^2 ), sigma_scatter ). The method floor is a between-study scatter, so it already contains each study's statistical error: printed sigma and floor are not added in quadrature (that would count the statistical part twice); the larger is used. Composition offset and conversion are independent of the measurement and are added in quadrature; a group scatter replaces the result when larger. **The weight written is ESPEI's reference sigma / the observation's effective uncertainty.** Groups with unequal effective uncertainties are split into single-observation files; groups with equal uncertainties retain their common weight. Activities use sigma_mu = R T sigma_a / a. The former opt-in `orchid_weight_floor.suggested_weight` is kept, marked superseded; it is not an active weight.

**Policy history (2026-10-25):** this refines the 2026-10-24 effective-uncertainty weighting policy by replacing the median weight of unequal-sigma groups with individual weights. It does not change any literature value, reported uncertainty, estimated component, calibration, reference-state conversion, or inclusion decision. The reported/estimated separation remains in force. `UNCERTAINTY_SPLITS.json` maps removed parent files to their replacements and records their old weights.

**Downstream fitting:** run `python espei/check_weights.py --bundle espei` before any weight overrides. For an explicit target and input selection, also run `check_fit.py` as described in `FIT_PREFLIGHT.md`. Use a clean copy of the regenerated bundle, with files selected from `index.json`; do not mix the old parent files with their replacements or count both as evidence. Keep the exported `weight` fields. Do not replace them with `suggested_weight` or reapply the default/sigma conversion. Any `apply_tags` or property-level weight setting is an additional user choice and must be recorded alongside the release and ESPEI version. Shared paper, source-dataset, and method/calibration links are preserved, including `orchid_weighting.parent_dataset` and `measured_in`. Splitting does not establish independence: the diagonal likelihood used here does not model covariance between measurements from the same study. No correlation coefficients have been inferred. Use the retained links when assessing sensitivity to shared-source effects.

**Method floors** are calibrated from ORCHID itself, like the DFT sigma: for each value type and method class, independent measurements of the same state from different primary sources (one value per source, re-listings counted once), robust between-study sigma = 1.4826 x state-weighted median|d| / sqrt(2), 95 %% bootstrap CI over states; floor = that sigma rounded up, never below ESPEI's experimental default; classes with fewer than 20 pairs or 10 states use a documented prior (`method_basis` prior). Floors used for the outputs of this bundle (all value types: experimental/CALIBRATION_EXPERIMENT.md):

| value type | method class | pairs | states | basis | robust sigma | 95 % CI | floor | unit |
|---|---|---:|---:|---|---:|---|---:|---|
| ACTIVITY | emf | 0 | 0 | prior (espei) |  |  | 0.5 | kJ/mol (chemical potential) |
| ACTIVITY | experimental_unspecified | 0 | 0 | prior (kubaschewski) |  |  | 1 | kJ/mol (chemical potential) |
| ACTIVITY | unknown | 9 | 9 | cleaned_below_threshold (kubaschewski) | 2.891 | 0.7333-4.29 | 2.9 | kJ/mol (chemical potential) |
| ACTIVITY | model_calculated (vs experiment) | 73 | 67 | prior |  |  | 1 | kJ/mol (chemical potential) |
| CPM | dsc_dta | 0 | 0 | prior (kubaschewski) |  |  | 0.5 | J/K/mol |
| CPM | experimental_unspecified | 0 | 0 | prior (kubaschewski) |  |  | 0.5 | J/K/mol |
| CPM | heat_content_calorimetry | 43 | 8 | cleaned_below_threshold (espei) | 0.4193 | 0.1363-1.856 | 0.5 | J/K/mol |
| CPM | unknown | 6 | 6 | cleaned_below_threshold (kubaschewski) | 0.5776 | 0.2563-1.118 | 0.6 | J/K/mol |
| CPM | model_calculated (vs experiment) | 225 | 89 | prior |  |  | 1 | J/K/mol |
| G_MIX_PARTIAL | emf | 0 | 0 | prior (espei) |  |  | 0.5 | kJ/mol |
| G_MIX_PARTIAL | experimental_unspecified | 0 | 0 | prior (kubaschewski) |  |  | 1 | kJ/mol |
| G_MIX_PARTIAL | unknown | 13 | 13 | cleaned_below_threshold (kubaschewski) | 1.319 | 1.086-1.697 | 1.4 | kJ/mol |
| G_MIX_PARTIAL | model_calculated (vs experiment) | 94 | 70 | prior |  |  | 1 | kJ/mol |
| HM_FORM | combustion_calorimetry | 15 | 4 | cleaned_below_threshold (kubaschewski) | 0.8806 | 0.08806-5.33 | 1 | kJ/mol-atom |
| HM_FORM | dsc_dta | 2 | 2 | prior (kubaschewski) |  |  | 2 | kJ/mol-atom |
| HM_FORM | emf | 17 | 8 | cleaned_below_threshold (kubaschewski) | 4.566 | 1.887-14.82 | 4.6 | kJ/mol-atom |
| HM_FORM | experimental_unspecified | 29 | 9 | cleaned_below_threshold (kubaschewski) | 4.026 | 1.992-9.33 | 4.1 | kJ/mol-atom |
| HM_FORM | kems_vapour | 20 | 15 | calibrated | 4.193 | 1.153-7.653 | 4.2 | kJ/mol-atom |
| HM_FORM | kems_vapour_second_law | 1 | 1 | prior (drowart2005) |  |  | 3 | kJ/mol-atom |
| HM_FORM | kems_vapour_third_law | 0 | 0 | prior (drowart2005) |  |  | 2 | kJ/mol-atom |
| HM_FORM | solution_reaction_calorimetry | 80 | 47 | calibrated | 3.145 | 2.202-4.602 | 3.2 | kJ/mol-atom |
| HM_FORM | unknown | 290 | 98 | calibrated | 3.46 | 2.411-4.298 | 3.5 | kJ/mol-atom |
| HM_FORM | model_calculated (vs experiment) | 2045 | 361 | calibrated | 3.81 | 3.113-4.547 | 3.9 | kJ/mol-atom |
| H_MIX_PARTIAL | experimental_unspecified | 6 | 1 | prior (kubaschewski) |  |  | 1 | kJ/mol |
| H_MIX_PARTIAL | solution_reaction_calorimetry | 0 | 0 | prior (espei) |  |  | 0.5 | kJ/mol |
| H_MIX_PARTIAL | unknown | 1 | 1 | prior (kubaschewski) |  |  | 1 | kJ/mol |
| H_MIX_PARTIAL | model_calculated (vs experiment) | 133 | 57 | prior |  |  | 1 | kJ/mol |
| SM_FORM | emf | 3 | 1 | prior (kubaschewski) |  |  | 1 | J/K/mol-atom |
| SM_FORM | equilibration_other | 1 | 1 | prior (kubaschewski) |  |  | 1 | J/K/mol-atom |
| SM_FORM | experimental_unspecified | 0 | 0 | prior (kubaschewski) |  |  | 1 | J/K/mol-atom |
| SM_FORM | heat_content_calorimetry | 0 | 0 | prior (kubaschewski) |  |  | 0.5 | J/K/mol-atom |
| SM_FORM | kems_vapour | 1 | 1 | prior (drowart2005) |  |  | 2 | J/K/mol-atom |
| SM_FORM | kems_vapour_second_law | 0 | 0 | prior (drowart2005) |  |  | 3 | J/K/mol-atom |
| SM_FORM | unknown | 4 | 4 | prior (kubaschewski) |  |  | 1 | J/K/mol-atom |
| SM_FORM | model_calculated (vs experiment) | 118 | 42 | calibrated | 3.232 | 1.938-4.144 | 3.3 | J/K/mol-atom |
| S_ABS | emf | 45 | 1 | prior (kubaschewski) |  |  | 0.5 | J/K/mol |
| S_ABS | experimental_unspecified | 0 | 0 | prior (kubaschewski) |  |  | 0.5 | J/K/mol |
| S_ABS | heat_content_calorimetry | 19 | 9 | cleaned_below_threshold (espei) | 0.9121 | 0.07339-1.929 | 1 | J/K/mol |
| S_ABS | kems_vapour | 4 | 4 | prior (kubaschewski) |  |  | 0.5 | J/K/mol |
| S_ABS | unknown | 77 | 20 | calibrated | 0.6571 | 0.2851-1.174 | 0.7 | J/K/mol |
| S_ABS | model_calculated (vs experiment) | 260 | 80 | prior |  |  | 1 | J/K/mol |
| S_MIX_PARTIAL | emf | 0 | 0 | prior (kubaschewski) |  |  | 0.5 | J/K/mol |
| S_MIX_PARTIAL | experimental_unspecified | 0 | 0 | prior (kubaschewski) |  |  | 1 | J/K/mol |
| S_MIX_PARTIAL | unknown | 0 | 0 | prior (kubaschewski) |  |  | 1 | J/K/mol |
| S_MIX_PARTIAL | model_calculated (vs experiment) | 14 | 8 | prior |  |  | 1 | J/K/mol |
| T_INVARIANT | emf | 1 | 1 | prior (orchid_assumed) |  |  | 10 | K |
| T_INVARIANT | experimental_unspecified | 0 | 0 | prior (orchid_assumed) |  |  | 10 | K |
| T_INVARIANT | thermal_analysis | 0 | 0 | prior (orchid_assumed) |  |  | 5 | K |
| T_INVARIANT | unknown | 4912 | 1680 | calibrated | 10.48 | 9.435-10.48 | 11 | K |
| T_INVARIANT | model_calculated (vs experiment) | 14146 | 5377 | calibrated | 3.707 | 3.113-4.3 | 4 | K |
| X_PHASE | experimental_unspecified | 6 | 1 | prior (orchid_assumed) |  |  | 0.01 | mole fraction |
| X_PHASE | metallography_microscopy | 1 | 1 | prior (orchid_assumed) |  |  | 0.01 | mole fraction |
| X_PHASE | microanalysis | 0 | 0 | prior (orchid_assumed) |  |  | 0.005 | mole fraction |
| X_PHASE | thermal_analysis | 1 | 1 | prior (orchid_assumed) |  |  | 0.01 | mole fraction |
| X_PHASE | unknown | 3014 | 1426 | calibrated | 0.01048 | 0.01048-0.01153 | 0.011 | mole fraction |
| X_PHASE | model_calculated (vs experiment) | 16730 | 8821 | calibrated | 0.004448 | 0.004448-0.004448 | 0.005 | mole fraction |

Which sigma applies to which data:

| data | sigma | ESPEI weight |
|---|---|---|
| experimental thermochemical | uncertainty budget: max(printed, reading, method floor) (+ composition / conversion in quadrature; group scatter when larger) | ESPEI default / sigma_effective of the observation; unequal-sigma groups split into single-observation files |
| DFT (literature, MP, OQMD) | calibrated against experiment (dft/CALIBRATION.md): robust sigma of the value's stratum, leave-system-out, as `sigma_method`; bias recorded, not applied | 500 J/mol / sigma_DFT per value (a conversion, not a claim that DFT is accurate to 500 J/mol) |
| model-calculated (model/) | budget with `sigma_method` = the model class's robust deviation from experiment (formation quantities) or a documented prior; not a measurement uncertainty | ESPEI default / individual sigma_effective; exclude with the `model` tag if you prefer |
| ZPF (experimental/ ZPF, invariants/) | temperature and composition budgets per region (printed, method floor of the invariant / phase-composition class); ESPEI weighs ZPF by a driving-force tolerance | 1 (not set) |
| phase constraints (assessed/, derived/) | `orchid_uncertainty_budget` of the points (assessed: digitisation, boundary margin, handbook class; derived: the offset rule) | 1 (not set) |

Per folder and output: UNCERTAINTY_SUMMARY.md.

## What each dataset carries

ESPEI-native content (`components`, `phases`, `conditions`, `output`, `values`, `reference_states` / `reference_state`, `solver` for DFT endmembers, `tags`, `weight`), plus `reference` (source paper, DOI), `bibtex` (key in `references.bib`), a `comment` with the ORCHID observation ids, the evidence class and "Version 1, machine-extracted", an `orchid` block (ids, reader agreement, caveats, papers, DOIs, source) and `orchid_uncertainty`. Model datasets keep `orchid_model_evidence` (TDB name, SHA-256, link basis). DFT datasets name the TDB file (SHA-256), phase and endmember they map to (see SCHEMA.md).

## Caveats

- **Machine-extracted (Version 1).** Literature values were read by language-model readers and accepted on reader agreement plus deterministic checks; a hand-checked precision sample is in progress. Check a value against its paper (DOI) before relying on it.
- **Single reader.** Values read by one visual reader carry `reader_agreement: single_reader` in `orchid.observations` and the tag `single_reader`. Reader agreement is taken from the literature release.
- **Weights are your choice.** The weights written follow the rules above; tags and `orchid_uncertainty` let you set your own.
- **DFT.** 0 K energies used as 298.15 K enthalpies; one DFT source per fit; a value is written only where the compound maps to exactly one TDB phase (phase registry: DFT entry id or paper label) and exactly one endmember of that TDB model at the compound's composition (one dataset per system, TDB file and phase; `solver` site ratios and configurations as in the TDB). Excluded compounds are listed in validation.json with reasons. ESPEI 0.9.1 evaluates HM_FORM against each element's SER phase named on the TDB's ELEMENT line, which must exist as a phase in the TDB; in some binary TDB files it does not (e.g. GRAPHITE, ORTHORHOMBIC_GA, 1/2_MOLE_O2(G)), and ESPEI then stops with a KeyError. Such datasets are not written (likelihood evaluation gate, VALIDATION.md); to use those values, add the element phase to your working TDB and rebuild from the literature release.
- **Energy basis.** ESPEI reads energies per mole of atoms. Literature values whose per-atom basis was not printed carry the caveat `energy_basis_unstated`.
- **ZPF encoding.** Phase labels are mapped to the model phases of the assessment the source folder is named after (tag `via_assessment`); a printed compound composition that is its formula rounded is written at the exact formula value; a `__HYPERPLANE__` vertex is added only where every known composition comes from stoichiometry. Range temperatures are written at the midpoint (`orchid_value_bases`).
- **Line compounds (E2).** A ZPF vertex of a phase that is stoichiometric in the target TDB is written with composition null: ESPEI 0.9.1 / pycalphad 0.11.2 return NaN for a line compound at a fixed composition (one NaN makes the ZPF likelihood -inf), and with null ESPEI evaluates the compound at its formula composition. Where that leaves a region with no known composition, a `__HYPERPLANE__` is added at the mean of the compound compositions. Known limitation of ESPEI 0.9.1: a single-phase vertex at a line-compound composition gives an undefined target hyperplane (zero driving force) once the compound is stable there, so a congruent melting point constrains the model in one direction only; the invariant brackets below constrain both.
- **Duplicates (E3, E4).** Identical ZPF regions in several datasets (several assessment folders, or several papers quoting the same invariant) are written once, with every source in `orchid.observations[i].also_reported_as`; a printed row read twice (same row id, conditions and value) counts once (`also_read_as`). Records concerned are listed in UPSTREAM_CORRECTIONS.md.
- **Activity reference states (E5).** Every activity table was checked against the paper: the reference state written is the one the paper prints (or, where nothing is printed, the pure liquid of an element that is liquid at the temperature, recorded as `not_printed_implied`); `orchid_reference_state` gives the basis and the printed text. Literature records whose reference state differs are listed in UPSTREAM_CORRECTIONS.md; the published literature values are not changed here.
- Values removed by the inclusion gate and datasets ESPEI rejects are listed in VALIDATION.md.

## Former optional minimum-sigma policy (E6, superseded)

Until 2026-10-23 every experimental thermochemical dataset carried `orchid_weight_floor.suggested_weight` (sigma floor = ESPEI's experimental default, 500 J/mol), which users had to opt into. Since 2026-10-24 the written `weight` follows sigma_effective (method floors) by default. The block is kept, with `superseded`, the former suggestion and `weight_before_2026_10_24`, for comparison: 1336 datasets, weight lower than before in 1244, higher in 1, unchanged in 91.

## Phase constraints (optional)

ZPF data alone can let a refinement misfire: ESPEI 0.9.1 penalises a phase that should be present but is not, but gives no driving force for a liquid point at a line-compound composition once the compound is stable, so a compound can melt far above its printed melting point without cost. Three optional, tagged dataset groups add phase evidence. They share the phase names of the assessment folder (TDB) they are written for, so they load together with `experimental/` and `invariants/` of that TDB. Use them by choice, never as measured data.

- `derived/invariant-brackets/<SYS>/<TDB>.<file>.json` (tags `derived`, `derived_bracket`): For each binary invariant reaction (eutectic, peritectic, monotectic, eutectoid, peritectoid, metatectic, syntectic, congruent melting or transformation, polymorphic) with printed temperature T and printed temperature uncertainty u (u = sigma_effective of the value's uncertainty budget: the printed sigma or the method floor, whichever is larger; else the printed sigma; else 5 K, documented default), two ZPF points are written at the invariant composition x0: at T + max(2u, 10 K) the phase set stable above the reaction and at T - max(2u, 10 K) the set stable below. The reaction is read as printed with the cooling convention (left side = high temperature; a reaction written on heating, i.e. a single liquid formed from solids, is turned round). x0 is the composition of the phase alone on its side (eutectic: the liquid; peritectic: the product; congruent: the common composition); the other side's phases are written with their printed compositions (null where not printed or for line compounds, E2). The target hyperplane is estimated at x0 - d and x0 + d (d = min(0.01, half the distance to the nearest printed composition or to a pure element)) so that a line compound at x0 cannot make the hyperplane undefined (ESPEI 0.9.1 returns zero driving force when the equilibrium at the hyperplane composition is a single phase without internal degrees of freedom). A point is not written when another invariant of the system that shares a non-liquid phase (or a liquid within 0.05 in composition of x0) lies between T and the bracket temperature. These points encode the semantics of the printed reaction (which phases are stable on which side); they are DERIVED, not measured. Each point records its source observation, T, u and offset (`orchid.points`).
- `derived/liquid-stability/<SYS>/<TDB>.liquid-stability.json` (tags `derived`, `derived_liquid_stability`): Single-phase LIQUID points at x = 0.1 ... 0.9 at T = T_liq_max + 50 K, where T_liq_max is the highest of the melting points of the two elements (SGTE unary data), every accepted printed temperature of an experimental literature value of the system that involves the liquid (invariants, liquidus and ZPF points) and every assessed (handbook) two-phase field point that contains the liquid. Written only where the evidence shows complete liquid miscibility: the system has an assessed (handbook) diagram with a single-phase liquid field and no second liquid (no L#2 field), no literature value of the system has two liquids (monotectic, syntectic or two liquid labels), both elements are condensed liquids at 1 atm (metals and metalloids) and T stays below both normal boiling points. Purpose: forbid spurious liquid miscibility gaps (and solids stable above the liquidus) in a fit. DERIVED, not measured. `orchid.basis` gives the temperatures used.
- `assessed/<SYS>/<TDB>.<DIAGRAM>.json` (tag `assessed_phase_diagram`): the ZPF datasets of assessed-phase-assemblages (one point inside each named field of an expert-drawn handbook diagram) with the phase names of that TDB (phase registry); points with a phase that the TDB does not model, or models twice, are left out; `orchid.confidence` per point. Evidence class assessed_phase_diagram: neither experimental nor CALPHAD.
- `assessed/<SYS>/<TDB>.<DIAGRAM>.dense.json` (tags `assessed_phase_diagram`, `dense_field_grid`; validated diagrams only, currently Cu-Zr): the dense interior grid of the same diagram (assessed-phase-assemblages `dense-field-grid/`), as separate datasets so that they can be weighted apart from the one-point-per-field datasets; `orchid.field_weight_share` per point (1 / points in the field: one observation per field).

Counts (brackets written and skipped by reason, systems with liquid-stability points) are in validation.json (`phase_constraints`). Suggested use: weight each group so that it contributes comparably to the measured data (e.g. per-type weights proportional to 1/N), and check the fitted model's invariant temperatures against the printed ones after the fit.

## Citation

R. Arróyave, J. Saal, D. Shin, ORCHID ESPEI datasets: literature and database-derived thermodynamic fitting inputs (version 2026.10; Version 1, machine-extracted), Zenodo (2026). https://doi.org/10.5281/zenodo.23735788

Part of: R. Arróyave, J. Saal, D. Shin, ORCHID data: umbrella record of the ORCHID datasets (version 2026.10; Version 1, machine-extracted), Zenodo (2026). https://doi.org/10.5281/zenodo.23736065

Also cite the original source of every value you use (each dataset's `reference` / `bibtex`, references.bib; index.csv lists the DOIs); A. Jain et al., The Materials Project: A materials genome approach to accelerating materials innovation, APL Materials 1 (2013) 011002, https://doi.org/10.1063/1.4812323 (CC BY 4.0); J. E. Saal et al., JOM 65 (2013) 1501, https://doi.org/10.1007/s11837-013-0755-4 and S. Kirklin et al., npj Comput. Mater. 1 (2015) 15010, https://doi.org/10.1038/npjcompumats.2015.10 (CC BY 4.0).

Licence: CC BY 4.0 (derived data; MP and OQMD data CC BY 4.0). The source papers are not redistributed; only their citations and DOIs.