Changelog

Unreleased

ST — the stream table (ADR 0068, accepted 2026-11-08)

• The web shell's streams screen (#/job/{jid}/streams) shows a run's stream table from get_stream_table, below the n, T, P view: per stream its total and its phases, the phase source and the near-phase-boundary flag, phase fractions, flows and enthalpy flows, and the energy summary per unit, in SI units as the columns name them; a value the table does not give carries its reason in plain words (Frank's O3). The shell now calls 18 operations (L-WEB-1).

X05 — replay from the wheel (ADR 0069, accepted 2026-11-08)

• The wheel and the sdist carry requirements.lock (openflowsheet/_data/requirements.lock, a link to the root file in a checkout). environment() records its SHA-256 only after checking the nine distributions of the runtime closure (attrs, casadi, jsonschema, jsonschema-specifications, numpy, pyyaml, referencing, rpds-py, scipy) against it by exact version string, in a checkout and an installed package alike; otherwise it records "", and replay names the failing distributions in its reason. An installed package at the lock now replays its own bundles exact_replay / MATCH where ADR 0007 D4 allows (envelope L41 narrowed).
• A checkout whose environment is off the lock now records "" where it recorded the lock's hash before. Optional extras are not part of the identity.
• The closure's four transitive members (attrs, jsonschema-specifications, referencing, rpds-py) are not pinned in [project].dependencies (ADR 0069 Amendment 1, Frank's Q-F1 = no), so users keep upgrade freedom over them. An install constrained by the shipped lock replays exactly: pip install -c requirements.lock openflowsheet, with the requirements.lock of the release's tag. A plain pip install openflowsheet whose resolver picks other versions of those four records "" and names them in every replay reason.
• No schema change: the environment document keeps its eight keys.

RRC — revision-region continuation (ADR 0072, Accepted 2026-11-08; R-990…R-995)

• New solve policy RRC-revision-v1, offered by explicit id and never the default: T06-revision-v2 with one more edge-3 action on revision_eo. A failed revision solve continues the changed operating values (pinned state, outlet and duty specifications; registered operating parameters such as split_fraction or a column's reflux and boil-up) from the nearest ancestor revision with a VERIFIED solve, starting at that ancestor's certified root, before the sequential restart. One recovery solve per step; any refusal is recorded and falls through to the restart unchanged.
• Not a superset of T06-revision-v2: on every measured case the restart also converged, in fewer iterations, and a continuation that stops part-way ends the step without trying the restart.
• Records: the certificate's recovery item (continuation.type: revision_continuation), the solve-path.json member revision_continuation (replayed alone: a rerun does no store lookup), and one initializer_rejected event on a refusal. Recovery-edge inventory row E3c; envelope L21 and L-COL-9 narrowed.
• Schemas: three frozen schemas each accept one more value (globalization.eo_recovery; continuation.type in the attempt context and the certificate); no member is added and every existing document validates unchanged. No registered policy, record or replay moves.

v0.3.0.post1 — the PyPI page: a current README and the project links; no code change

A post-release of 0.3.0 (R-922): the same release, republished so that its PyPI page is current. It is gated as 0.3.0 at 0.3.0's candidate C = a3e508e35fc5451fe56fa7b70941621992d35b50, whose verdicts carry unchanged; the envelope stays v0.3-envelope-1 (release: 0.3.0).

• README.md, the PyPI long description, is rewritten for a PyPI reader: what OpenFlowsheet is, what 0.3.0 contains and what it has and has not established, how to install it and its extras, a quickstart, and a links section; every link is absolute.
• pyproject.toml gains [project.urls]: the documentation site (Homepage, Documentation), the repository, the issue tracker, the changelog and the GitHub releases.
• No code, schema, data, dependency or lock change: the wheel and the sdist differ from 0.3.0's only in the version, the README and the project links. Everything the ## v0.3.0 section below states holds for 0.3.0.post1 unchanged.

v0.3.0 — unreleased

Draft (V03 WO-4); nothing below is released. The release candidate C, its gate table and the verdicts are added to this section at C (V03 spec §4, ADR 0053). 0.3.0 claims V11–V20, W21–W26 and X31; W27 is not claimed.

S01 — the countercurrent equilibrium-stage subsystem

• Equilibrium stages with explicit degrees of freedom and energy and phase coupling, an initializer and Newton continuation, on SYN-001. Manifest evidence/S01/465240a4…/manifest.json.

S02 — the MESH column

• Stage indexing, feed, condenser, reboiler, pressure and specification conventions; four matched cases and the invalid-specification and invalid-phase refusals. Manifest evidence/S02/1810dde8…/manifest.json.

S03 — independent column references

• Column answers compared with an independently written implementation across the available suite, with perturbation studies and the supported nonideal scope. Manifest evidence/S03/930aad1b…/manifest.json.

S04 — the Peng–Robinson component library pr-lib-v1 (library-only)

• A vapour-phase Peng–Robinson library provider; a revision cannot select it, and its records do not ship. Manifest evidence/S04/5c29fbf6…/manifest.json.

S05 — IAPWS-95 water iapws95-v1

• The IAPWS-95 formulation as a provider with its own reference convention IAPWS95-ref-v1; the coefficients ship with IAPWS's attribution in NOTICE. Manifest evidence/S05/01cdeed4…/manifest.json.

S06 — column specifications by purity, recovery and stage temperature

• Purity, recovery and stage-temperature specifications for the column, with their rows, degrees of freedom and typed refusals. Manifest evidence/S06/ad264a23…/manifest.json.

S07 — CoolProp pure fluids coolprop-heos-v1 (optional extra coolprop)

• CoolProp 8.0.0's Helmholtz equations of state for 123 fluids in two tiers, reference convention COOLPROP-DEF-ref-v1, compared across architectures within registered bounds (ADR 0046). Manifest evidence/S07/7bfab70d…/manifest.json.

S08 and S08b — one-phase column sections and the phase contract

• Block openings for column sections in one phase (ADR 0048), and signature items by declaration with saturation consistency for S01 cascades. Manifests evidence/S08/6d7d1a6d…/manifest.json and evidence/S08b/acbd1266…/manifest.json.

S09 — BioSTEAM as a third column reference

• A blind third reference for the column cases: 19 of 19 qualified, X31 met where available for 3 of 4 cases. Manifest evidence/S09/4d4df087…/manifest.json.

S10 and S10-C — the Wilson liquid-mixture method on real records

• wilson-ig-v1: Wilson activity coefficients with an ideal-gas vapour for methanol, ethanol and water, Wilson parameters fitted to published equilibrium data (Kurihara et al. 1993), with the VAL-1–3 validation bands (ADR 0047); S10-C adds the γ–φ PH flash and the column on Wilson (ADR 0049). Wilson's records and the Kurihara 1993 points ship as package data (V03 spec §9.1). Manifests evidence/S10/f6f27c3f…/manifest.json and evidence/S10c/2aedd250…/manifest.json.

S11 — the binding package

• Water (IAPWS-95), CoolProp's pure fluids and Wilson selectable in flowsheets; the seven thermal.* units; the column as a flowsheet unit (ADRs 0050, 0051). semantic_role gains compressor, turbine and other (ADR 0054). Manifest evidence/S11/01e72ca4…/manifest.json.

S12 (E01) — cost estimation and LCA

• Cost and life-cycle impact as differentiable post-solve functionals of a verified state, with an accuracy class (ADR 0052); the LCA table builder needs the optional extra lca. Manifest evidence/S12/02fc8d2f…/manifest.json.

What it is not

• numerical verification, not empirical validation
• no global optimality
• Linux only
• not a sandbox
• the stand-in reactor is synthetic
• no reactor result has been compared with experimental data by this project: the group's published comparison is the group's evidence, and every real-reactor number is the group's model at the pin, a regression, whose discretization bias is 1.0–1.4 % in NH₃ (M02 design note §13)
• the optimization routes need the optional nlp extra
• the CoolProp provider needs the optional coolprop extra
• cost and LCA results are assessment estimates with an accuracy class
• cost and LCA are Python-only
• no ecoinvent data
• Wilson is vapour-liquid only; no liquid-liquid split
• no automatic liquid/vapour regime switching for pure fluids
• X31: one independent reference for the refluxed absorber
• no rate-based column models
• no W27 claim in 0.3.0
• no CFD or Peclet refinement
• no external benchmark score

v0.2.0a1 — release candidate C_α = cc73219a77f6a54579c8acb151ed6fae228c8a4c (0.2.0a1); a pre-release; tag and publication by Frank

ADR 0021 D4, reused by ADR 0028. The gate table below is what scripts/v0_2_gate.py --rc C_α --markdown prints at the tag candidate, after the design lane's verdicts on C_α (docs/reviews/V02-alpha-verdicts.md: judged 2026-10-16, V13, V17 and W21 re-judged 2026-10-17 on Frank's answers in ADR 0028). Nothing in this section is released until the tag; the tag and every publication are Frank's (ADR 0021 D5).

What it does. Everything v0.1.0 does (the application contract, the thirteen unit models, the planner, the verifier, the certificate or the typed failure, the replay bundle) and, added since v0.1.1: an adapter that runs the pinned PyMRM reactor as a unit model, with its coupled route and its experiment artifacts (M02); the Peng–Robinson property route of the ammonia loop (M01); and a diagnostic web shell served by serve-http --ui, which edits nothing (Solve and Cancel are its only effects; M06). The supported envelope is docs/support-matrix.md: unlisted is unsupported.

Gates (V11–V20, W21, W22), as scripts/v0_2_gate.py prints them. A FAIL is printed FAIL.

Gate	Verdict at C	Required evidence	Clauses and travelling limitations
V11	PASS	Semantic diff, drafts, task validation, units/references, real migration when applicable	limitations: L38
V12	PASS	Required unit library with per-model definition of done	limitations: L02, L26, L27, L28, L29, L30, L-CSTR-1, L-CSTR-2, L-CSTR-3
V13	PASS	DM/SCC/BTF, tear and EO, scaling and initialization	limitations: L10, L22, L23, L24, L36, L38, L43, L-WS-1, L-WS-2
V14	FAIL	Homotopy, qualified PTC/SER and three bounded recovery edges	failing: V14 (b); V14 (b) FAIL, accepted by Frank 2026-10-17 (ADR 0028 D3); limitations: L01, L08, L20, L21, L43
V15	PASS	Frozen phases, EO cross-unit spec, multiple-root evidence	limitations: L02, L14, L19, L20, L33
V16	PASS	30+ distinct cases, eight shared comparisons, two both-reference cases where feasible, registered 20x20 ensemble	limitations: L16, L17, L18, L32
V17	PASS	Local/HTTP/MCP, ten agent tasks, >=80% success, zero unauthorized actions/false verification	limitations: L11, L12, L13
V18	PASS	Distribution/data gates, local rank evidence, qualified energy checks, all recovery edges tested	limitations: L03, L05, L06, L15, L19, L25, L34, L37, L39, L40, L41
V19	PASS	Real-chemistry dossier with usable data/kinetics/property/reference route	—
V20	PASS	>=95% nominal robustness, correct invalid-structure handling, no adversarial false verification	limitations: L04, L07, L09, L31, L32, L38
W21	PASS	PyMRM boundary/accuracy/execution and provenance	limitations: L05, L06, L07, L15, L19, L20, L44, L45, L46, L48
W22	PASS	One validated needed nonideal property route	limitations: L42, L47
W23	not claimed	Surrogate conservation/domain/derivative/default UQ and promotion	—
W24	not claimed	Fixed-topology optimization, eligible trust-region example and parent-model checks	—
W25	not claimed	End-to-end real reactor refinement with honest evidence limits	—
W26	not claimed	Diagnostic web shell through shared contract	—
W27	not claimed	External agent benchmark adaptation attempted; inaccessible assets disclosed	—

What it is not.

• Not 0.2.0. A pre-release: W23–W27 are not claimed, and the web shell is included and not claimed.
• Not empirically validated. SYN-001 is synthetic, and every result is numerical verification against the same equations. Also, no reactor result has been compared with experimental data by this project: the group's published comparison is the group's evidence, and every real-reactor number is the group's model at the pin, a regression, whose discretization bias is 1.0–1.4 % in NH₃ (M02 design note §13).
• Not a guarantee. A human sign-off records that someone with the relevant expertise examined the evidence and raised no objection, not that the content is correct (R-017).
• Not general beyond the support matrix. What docs/support-matrix.md does not list is unsupported, and each unsupported capability ends typed.
• Not a certificate of solution accuracy. VERIFIED certifies residual accuracy at the registered tolerances; the solution-error bound is recorded and disclosed, never promised (ADR 0007 F3).
• PTC is experimental. No PTC family is qualified: V14 (b) is judged FAIL at C_α.
• The agent evidence is narrow. One pinned model and configuration, MCP only. V17's agent evidence is carried from v17-c2's surface, not re-measured on v0.2's: Frank accepted the carry (ADR 0028, 2026-10-17), and the release candidate's job checks the recorded decomposition chain from that surface to this one (M07a.A37).
• The reactor is opt-in and measured on one machine. Real-reactor runs are opt-in (they need the separately built, pinned reactor environment) and of reproducibility class R3 (ADR 0034 D6). The reactor has been run on Linux x86-64 on one machine (ref-x86-64): it is not measured on aarch64, and no CI job runs the reactor. What CI runs in its place is a stand-in, and the reactor stand-in is synthetic; it certifies nothing about the reactor. There is no warm start of the reactor (L44).
• The reactor execution profile is not a sandbox. The group's pinned code runs with the user's privileges, unconfined (L45). The profile is registered on Linux; macOS is best effort, and Windows is refused before spawn (ADR 0033 D2).
• The property route is narrow. Validation covers pure-component behaviour within R-160's bands against the reference equations of state; k_ij = 0; there is no mixture VLE validation. Of pr-c1-v1's liquid side: the LIQUID regime and pure-NH₃ flashes are deferred and refused typed (U15), and against the reference equations of state its saturated-liquid volume is 11–30 % high and its enthalpy of vaporization 2–5 % high (L47); no v0.2 loop decision uses either.
• Rights. The group's 4TU dataset may be used and referenced (Frank, 2026-10-02; the real chemistry dossier §8, row 8); no result of this release is compared with it.

Known limitations, by support-matrix id (the text of each is in docs/support-matrix.md): L01, L02, L03, L04, L05, L06, L07, L08, L09, L10, L11, L12, L13, L14, L15, L16, L17, L-CSTR-1, L-CSTR-2, L-CSTR-3, L-WS-1, L-WS-2, L18, L19, L20, L21, L22, L36, L37, L38, L39, L23, L24, L25, L26, L27, L28, L29, L30, L31, L32, L33, L34, L35, L40, L41, L42, L43, L-WEB-1, L-WEB-2, L-WEB-3, L-WEB-4, L44, L45, L46, L47, L48.

Reproducibility. In the words of ADR 0007 F1: "the project reports bitwise agreement and never promises it, on any platform including a single machine; R0 bit-identity of structure plus D2 agreement for floats is the promise."

Review status of every package (each manifest's status at C_α, the manifest at the latest committer time where there are several; human review fields as recorded; M07a, whose manifest is the record of this gate and is written after C_α, has no row, R-320):

Package	Status	Human review (numerical / process model)
K01	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K02	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K03	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K04	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K05	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K06	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
M01	tested	pending / pending
M02	tested	pending / pending
M06	tested	pending / pending
P00	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
P01	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
P02	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
P03	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
T01	tested	pending / pending
T02	tested	pending / pending
T03	tested	pending / pending
T04	tested	pending / pending
T05	tested	pending / pending
T05b	tested	pending / pending
T06	tested	pending / pending
T07	tested	pending / pending
T08	tested	pending / pending

v0.2.0 — unreleased

In development on main; nothing below is released. The gate table is added to this section at the release candidate (M07).

M01 — the Peng–Robinson route of the ammonia loop (W22)

• C1 property route. Peng–Robinson with the light gases vapour-only, its root and phase rules, an equilibrium-vapour flash, and the formation datum PR-C1-ref-v1 (ADR 0026); the parameters are cited and registered. The reactor boundary of the loop is ADR 0027.

M02 — the PyMRM reactor adapter (W21)

• External reactor as a unit model. The pinned PyMRM reactor runs in an out-of-process child, one per attempt, under a three-layer kill chain, with process-level experiment identity and an exact cache of deterministic outcomes (ADR 0033); it runs inside a flowsheet solve (ADR 0034), with model replacement and promotion as a checked commit (ADR 0035). Experiment records and the coupled route are replay artifacts. A synthetic stand-in reactor ships for the default gate and is listed as synthetic.

M03 — sensitivities, studies and the NLP bridge (W24, part)

• Parametric sensitivities and studies. Parameters are pinned inputs; their derivative comes from a bit-identical parametric twin and is issued only at a qualified regular root; sweeps and the estimation example are studies over those parameters (ADR 0031).
• General NLP bridge. A PyNumero gray box over the parametric twin solved by an audited cyipopt, judged only on the re-solved, certified simulation (ADR 0032). It needs the optional nlp extra and its licences (accepted by Frank 2026-10-16).

M04 — surrogate evidence (W23)

• Surrogates as evidence. A registered uniform reference box, a plan registered before the parent runs, joint split-conformal intervals, and promotion decided by an integer coverage test (ADR 0036). The surrogate is a native unit pinned by its manifest's hash, promoted and rolled back through the checked commit (ADR 0037); the surrogate_study job kind joins the contract.

M05 — the trust-region study of the C1 loop (W24)

• Trust-region adapter. Pyomo's trust-region filter over a projection of the canonical problem, with the audited Ipopt executable for its subproblems (ADR 0038); a conditional scipy-only fallback is ADR 0040. The C1 study optimizes the reactor inlet temperature for liquid NH₃ product; candidates are judged by parent checks and every true-model call is accounted (ADR 0039).

M06 — the diagnostic web shell (W26)

• serve-http --ui. A browser view of one project at /ui/, on the same origin as the unchanged HTTP API: revisions with a flowsheet drawing, validation and degrees of freedom, the equation view (structure), runs with attempts, trace and plan (live while a job runs), certificates, failure bundles, streams and units, bundle files with download, revision and run comparison, and the agent history. Solve and Cancel are its only actions; it edits nothing. docs/web-shell.md says what each screen shows and does not claim; ADR 0030 (proposed) and docs/design/M06-web-shell.md are the design.
• Contract only. Hand-written ES modules and CSS shipped as package data: no build step, no framework, nothing third-party. One fetch call site, URLs from a route table generated from OPERATIONS, exactly 17 operations called; records read whole through the raw export and checked against their registered SHA-256. The bearer token stays in session storage and the Authorization header; every /ui response carries a strict Content-Security-Policy.
• Contract amendments (ADR 0019 Amendment 3, proposed). inspect_structure gains a row and column index and, for an unroutable revision, the validation's own analysis; diff_revisions gains element-level detail; a new operation, list_audit, reads the project's audit history (own rows with read, other principals' with policy).
• Tests. Static scans of the shell's sources; Node's built-in runner on the view models and screens against fixtures captured from real responses (check.sh needs Node ≥ 22 unless OPENFLOWSHEET_SKIP_WEB_TESTS=1); every used operation over HTTP; a headless-Chromium test of every screen with a request recorder, Solve, a live and a cancelled job, and the downloads (skipped without a browser, required on CI's x86-64 leg).

M07 — the reference journey and the release gate (W25)

• Reference journey. One decision (the reactor inlet temperature) taken through four rungs of fidelity, with decisions compared by the parent's objective against its noise floor; W25 is judged on reporting what happened, not on whether the decision changed (ADR 0041).
• The 0.2.0 gate. scripts/v0_2_gate.py gains v0.2-gate-v1, which claims all seventeen gates V11–V20 and W21–W27, each judged at its own candidate, with per-release acceptance of failures (ADR 0042, proposed).

What it is not

• not a sandbox
• Linux only
• reactor results reproduce from records (R3)
• T_in 653.15-693.15 K, P 9-11 MPa, H2/N2 2.5-3.5, inerts 0.035-0.2, per-tube flow 0.5-2 x nominal
• no empirical validation of the C1 loop
• no reactor result has been compared with experimental data by this project: the group's published comparison is the group's evidence, and every real-reactor number is the group's model at the pin, a regression, whose discretization bias is 1.0–1.4 % in NH₃ (M02 design note §13)
• no global optimality
• the optimization routes need the optional nlp extra
• the stand-in reactor is synthetic
• no external benchmark score
• no CFD or Peclet refinement
• no equilibrium-stage subsystem

v0.1.1 — the single public repository; release by workflow; no change to the simulator

The code and data of the simulator are v0.1.0's: under the ADR 0021 D2.4 paths (src/, schemas/, benchmarks/, requirements.lock, pyproject.toml, MANIFEST.in, README.md, LICENSE, NOTICE) this release differs from the release candidate C = 67c66d9 only in the version, as v0.1.0 did, and scripts/v0_1_gate.py --rc 67c66d9… proposes it under the v0.1 gates. 0.1.1 is the first release published through the release workflow.

• One repository. computational-chemical-engineering/openflowsheet is now the working repository — branches, pull requests, CI, tags and releases (R-150). Its history starts at v0.1.0; the development history before it is archived, read-only, in openflowsheet-dev, and the commit ids cited under evidence/ refer to that archive (docs/HISTORY.md).
• Release workflow. .github/workflows/release.yml, dispatched by hand and a dry run by default: the release gate, the tag, the sdist and the wheel (built twice and inspected), the PyPI upload by trusted publishing behind an approval, and the GitHub release with this section as its notes (docs/RELEASING.md). The two-repository snapshot release that preceded it is removed.
• The gate without C. Where C is not in the repository, scripts/v0_1_gate.py compares the candidate with C's recorded file hashes, release/rc-trees/<C>.json, by the same rule (R-151); it prints the version it judges (vX.Y.Z tag may be proposed) and covers the 0.1 line.
• Pre-push guard. scripts/install-hooks.sh installs a hook that refuses to push to the public repository any commit that does not descend from v0.1.0.
• T08 manifest generator. §A4.6's ancestry rule is judged on the development history only; a release commit without that history (the public v0.1.0) is listed as unrelated, not counted against it.
• Tests without the archived history. Tests that read pre-0.1.0 commits skip explicitly where those commits are absent, with the reason "pre-0.1.0 development history is archived in openflowsheet-dev (R-150)", and run unchanged where they are present.

v0.1.0 — release candidate C = 67c66d9 (0.1.0rc1), version-only change to 0.1.0; tag and publication by Frank

ADR 0021 D4. The gate table below is what scripts/v0_1_gate.py --markdown prints for the release candidate C = 67c66d98587f23bd7dfe8da28a8facccc92da21e after the design lane's verdicts (docs/reviews/T08-verdicts.md); T08's evidence is evidence/T08/67c66d9…/manifest.json (tested). The tag commit differs from C only in the version (ADR 0021 D2.4); the tag and every publication are Frank's (ADR 0021 D5).

What it does. A process simulator for ideal three-component processes that an agent or a person drives through one application contract — Python, CLI, HTTP and MCP over stdio, twenty operations — in which every solve ends in a certificate or a typed failure and writes a replay bundle. A flowsheet is a content-addressed revision of thirteen unit models (mixer, splitter, TP and PH flashes, heater, exchanger, pump, valve, component separator, conversion reactor, kinetic CSTR, feed and sink). Validation reports structure (DM/BTF), units and components before a solve. The planner tears recycles with safeguarded Anderson acceleration or merges them into equation-oriented regions, promotes cross-unit specifications into EO, and runs bounded phase attempts and the bounded recovery edges of docs/recovery-edges.yaml. A verifier that shares no factorization with the solver issues a SolutionCertificate (residuals, balances, regularity, alias certificates) or a failure bundle. Compatible warm starts and pseudo-transient continuation are offered by explicit policy only (T08-warm-v1, T08-ptc-v1). The supported envelope is docs/support-matrix.md: unlisted is unsupported.

Name. The project is OpenFlowsheet: distribution, import package and console script are openflowsheet (they were the provisional process-runtime and process_runtime; there is no alias), and the environment variables are OPENFLOWSHEET_* (R-149). Install the server extra with pip install 'openflowsheet[server]'; solve with openflowsheet solve SYN-001-nominal --out ./bundle. The schema $ids keep their v0.1 URLs. The rename moved the SYN-001 provider's self-hash and so the registered identity values that cover it, by substitution only (R-149).

Gates (V11–V20), as scripts/v0_1_gate.py prints them. A FAIL is printed FAIL; a released FAIL is only a clause ADR 0021 D3 lists with Frank's acceptance.

Gate	Verdict at C	Required evidence	Clauses and travelling limitations
V11	PASS	Semantic diff, drafts, task validation, units/references, real migration when applicable	limitations: L38
V12	PASS	Required unit library with per-model definition of done	limitations: L02, L26, L27, L28, L29, L30, L-CSTR-1, L-CSTR-2, L-CSTR-3
V13	PASS	DM/SCC/BTF, tear and EO, scaling and initialization	limitations: L10, L22, L23, L24, L36, L38, L-WS-1, L-WS-2
V14	FAIL	Homotopy, qualified PTC/SER and three bounded recovery edges	failing: V14 (b); V14 (b) FAIL, accepted by Frank 2026-09-29 (ADR 0021 D3); limitations: L01, L08, L20, L21
V15	PASS	Frozen phases, EO cross-unit spec, multiple-root evidence	limitations: L02, L14, L19, L20, L33
V16	PASS	30+ distinct cases, eight shared comparisons, two both-reference cases where feasible, registered 20x20 ensemble	limitations: L16, L17, L18, L32
V17	PASS	Local/HTTP/MCP, ten agent tasks, >=80% success, zero unauthorized actions/false verification	limitations: L11, L12, L13
V18	PASS	Distribution/data gates, local rank evidence, qualified energy checks, all recovery edges tested	limitations: L03, L05, L06, L15, L19, L25, L34, L37, L39, L40, L41
V19	PASS	Real-chemistry dossier with usable data/kinetics/property/reference route	—
V20	PASS	>=95% nominal robustness, correct invalid-structure handling, no adversarial false verification	limitations: L04, L07, L09, L31, L32, L38

What it is not.

• Not empirically validated. SYN-001 is synthetic — three pseudo-components with invented constants and ideal VLE — and every result is numerical verification against the same equations (L18).
• Not a guarantee. A human sign-off records that someone with the relevant expertise examined the evidence and raised no objection, not that the content is correct (R-017).
• Not general beyond the support matrix. What docs/support-matrix.md does not list is unsupported, and each unsupported capability ends typed (U01–U14).
• Not a certificate of solution accuracy. VERIFIED certifies residual accuracy at the registered tolerances; the solution-error bound is recorded and disclosed, never promised (ADR 0007 F3; L05).
• PTC is experimental. No PTC family is qualified: V14 (b) is FAIL (on PTC-R1, PTC ends at the saddle from 149 of 441 starts) and carried by Frank's acceptance (ADR 0021 D3, L01).
• The agent evidence is narrow. One pinned model and configuration, MCP only, not out of sample (L12).

Known limitations, by support-matrix id (the text of each is in docs/support-matrix.md): L01, L02, L03, L04, L05, L06, L07, L08, L09, L10, L11, L12, L13, L14, L15, L16, L17, L-CSTR-1, L-CSTR-2, L-CSTR-3, L-WS-1, L-WS-2, L18, L19, L20, L21, L22, L36, L37, L38, L39, L23, L24, L25, L26, L27, L28, L29, L30, L31, L32, L33, L34, L35, L40, L41.

Reproducibility. In the words of ADR 0007 F1: "the project reports bitwise agreement and never promises it, on any platform including a single machine; R0 bit-identity of structure plus D2 agreement for floats is the promise." The registered platforms are Linux x86-64 and aarch64 with Python 3.13 and the one lock (requirements.lock, ead4edf1…); floats are compared under K04-numerical-policy-v1.

Review status of every package (each manifest's status; human review fields as recorded):

Package	Status	Human review (numerical / process model)
P00	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
P01	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
P02	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
P03	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K01	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K02	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K03	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K04	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K05	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
K06	reviewed	signed 2026-09-23 (R-017) / signed 2026-09-23 (R-017)
T01	tested	pending / pending
T02	tested	pending / pending
T03	tested	pending / pending
T04	tested	pending / pending
T05	tested	pending / pending
T05b	tested	pending / pending
T06	tested	pending / pending
T07	tested	pending / pending
T08	tested	pending / pending

reviewed on P00–K06 is Frank's sign-off of 2026-09-23 in R-017's qualified form: it records examination without objection, not a guarantee of correctness. T01–T08 are tested: their design-lane reviews are recorded under docs/reviews/, and their human review fields are pending. T08's manifest is at the release candidate C = 67c66d9 (evidence/T08/67c66d98587f23bd7dfe8da28a8facccc92da21e/manifest.json): 85 checks, 83 pass and 2 FAIL (T08.A70 and B23, V14 (b), accepted by Frank in ADR 0021 D3).

Licensing. Apache-2.0 (LICENSE, NOTICE). The sdist and the wheel carry the project's own code, its package data and its licence files only (ADR 0006 mode A); CasADi (LGPL-3.0-or-later), numpy and scipy are fetched by the installer from PyPI.

v0.0.0 — 2026-09-23

The first vertical slice: an ideal three-component flash–recycle process, solved, independently verified, replayable, and inspectable from the command line.

What it does. Compiles a declared flowsheet to a sparse residual and exact Jacobian (CasADi 3.8.0); solves a three-variable tear with a damped Newton and bounded phase attempts; verifies the answer with a verifier that shares no factorization, cache or intermediate with the solver; issues a SolutionCertificate or a FailureBundle; writes a replay bundle that a third party can re-run or, failing that, inspect.

python -m process_runtime.application.cli solve SYN-001-nominal --out ./bundle
python -m process_runtime.application.cli inspect ./bundle
python -m process_runtime.application.cli replay ./bundle

Acceptance. All seven gates of plan §5.1 are met — PYTHONPATH=. python scripts/v0_0_gate.py reads them from the committed evidence rather than re-deriving them. Six of the seven are met with limitations, and the limitations are on the report rather than in a footnote. 1320 tests, green on x86-64 and aarch64.

What it is not.

• Not empirically validated. SYN-001 is synthetic: three pseudo-components with invented constants. Nothing here has been compared against experiment or an independent simulator.
• Not a guarantee. The human sign-off on every evidence manifest records that someone with the relevant expertise examined it and raised no objection — register R-017, and the field says so itself. Defects are expected to surface in verification and validation studies.
• Not general. The check set, the balances and the tear are written for the SYN-001 topology. A general structural decomposition is T01 and does not exist, which is what most of the recorded limitations are waiting on.
• Not a certificate of solution accuracy. VERIFIED certifies the residuals at the registered tolerances; the first-order solution-error bound is recorded and disclosed, never promised.

Known limitations, by gate. K01's second-order derivatives through an opaque callback (unsupported, reported rather than fabricated); K02's rank report; K03's initializer chain, of which only the registered local initializer is wired; K04 and K06's structural over-specification, which needs T01's matching; K05's clean-container replay.

Reproducibility. Structural artifacts are bit-identical across x86-64 and aarch64 — measured, in CI, every push. Floating-point agreement is reported and never promised: two ubuntu-latest runners were measured disagreeing on a converged state's last bits, and the project's promise is R0 structural identity plus agreement within the declared numerical policy. A changed dependency cannot produce an exact replay, because the mode is decided from the environment before anything runs.

Licensing. Apache-2.0. No third-party binaries are redistributed; CasADi is fetched from PyPI by the installer.