Array source revision dated 5 October 2026. The dated manifest and current revision record, rather than the unchanged version label alone, identify this source snapshot.
The current reproduction materials are in Reproduce_Analyses in this
repository. In the desktop user package that folder sits beside Editable_Source;
in Supplementary Code it is called Analyses beside Source. Run
python Reproduce_Analyses/repeat_analyses.py --output fresh_results from this
repository, with NumPy and SciPy installed. The user package also includes a
plain-language guide and a bundled Windows runtime.
A separately implemented comparator using established tree calculations matches all 57,488 tested profile values and has comparable computation time. Neither implementation is consistently faster across the complete case set. Earlier large speed ratios concern reference implementations that enumerate missing-record sets; they do not establish a general computational advantage. The UK threshold comparison also shows that high marginal coverage need not imply useful resolution. A calibrated box yields narrower intervals with lower marginal coverage. These results verify the calculations and their practical limits; they do not establish a new general mathematical theorem.
The Brazzaville comparison uses 8,208 public retail survey records, 16 vegetables and 14 market labels. April quantities fix two arrangements of seven numerical reports; August quantities evaluate them with an exact retained sample total. After one simulated report loss, the median worst-loss mean market width is 33.99% of the total for individual-market reports and 14.29% for paired-market reports. Without loss, individual reports have the smaller median width, 11.52% versus 14.29%. At a 10% share threshold, paired reports retain 108 of 224 below-threshold conclusions through every single loss; individual reports retain none. The paired width is the accounting identity 1/7. Both arrangements have full-total prospective widths under one loss, so that structural criterion alone does not distinguish them. Neither preserves an above-threshold conclusion through every loss.
This is a retrospective comparison using survey-derived quantities and simulated losses after aggregation. Reports share the original survey errors; the test does not validate independent backup acquisition or real field failures. All 7,168 endpoints agree with an independent linear program and the public engine. Inputs, protocol, results and 32 executable payload examples accompany the reproduction materials. Source and derived Brazzaville data retain CC BY-NC 4.0, separately from the software licence.
The single exhaustive compartment is fixed by conservation: its amount equals the remainder. The structural kernel is unchanged. Public origin-destination and CPC data are retained; private contaminant values are excluded. The source manifest identifies the supplied files, and the repository commit identifies the corresponding published snapshot when available.
TEM-FLOW reconstructs source–transient–destination mass-flow intervals when route evidence is incomplete, dated differently, or collected by separate institutions. It preserves mass balance and the identity of the material while distinguishing observed, source-reported, transparently derived, and model-derived quantities.
TEM-FLOW does not force a single allocation when several networks satisfy the evidence. Its main outputs are feasible intervals, supporting solutions, evidence links, sufficiency measures, and explicit refusal states. A point allocation is optional and is always labelled as a prior-dependent working value.
- Evidence-Resolved Reconstruction (ERR) for exact bounds over a physical feasible set and its least-unresolved subset.
- A joint destination-share set on the simplex, calibrated with total-variation residuals, for evidence states that contain earlier route shares and a compatible current downstream anchor.
- Typed identities, transformation certificates, dated-evidence matching, and an unresolved-destination compartment.
- The inherited ED-FLOW-style reviewer interface with the TEM-FLOW computation and reporting rules underneath it.
- Optional downstream CPC and chemistry consumers. These are applications of a compatible mass interval, not part of the method's novelty claim.
- Explicit include/exclude switches for ERR, CPC compatibility, contaminant tracing, exposure/ED/THQ, market access, flow irregularity, postharvest loss, supply continuity, and price/affordability monitoring.
- An explicit exposure-scenario consumer can combine a declared last-known or projected consumer population, last-observed or modelled retained mass, and last-reported food concentration to calculate CPC, estimated daily intake, and THQ. The result records every input basis and its analysis date. A red node ripple and route pulse is available only after a declared THQ exceedance and only when the reviewer clicks Display exposure animation. An interval that only crosses the threshold uses amber. Running the model never displays the animation automatically, and route pulses do not assert that contaminated product travelled along the displayed route.
The public package contains the computational slots and a local CSV or JSON file loader, but no contaminant-chemistry application ledger. A dated CSV uses one observation per row and is validated and converted internally to the engine's JSON contract. Sensitive application records can be supplied on demand and remain in browser memory for the active session. See the plain-language input guide and the blank CSV template. A filled synthetic example shows four record types without disclosing application data. Private application ledgers must be supplied separately; none is included here.
Map coordinates are drawn from the audited predecessor coordinate registers where an exact or documented same-feature match exists. Commodity-state child nodes inherit their physical parent's audited coordinate. Unresolved registry nodes remain available to the model and lists, but their display-only layout proxies are hidden by default and can be shown explicitly. Double-click selects a map feature; left-drag pans; right-drag up/down zooms; Escape clears floating messages, map selection, and exposure animation.
Source mass alone is not treated as enough evidence for useful destination allocation. The Karg validation returned a median interval width equal to 98.32% of source mass. Likewise, a missing downstream record does not imply 100% source retention. Both conditions return broad intervals or refusal until compatible evidence is supplied.
python -m pip install .
python -m temflow validate
python -m unittest discover -s tests -v
python analysis/python/reproduce_all.py --output analysis/results/REPRODUCTION_REPORT_V1_0_0.jsonThe last command distinguishes raw-to-result reproduction from independent
rescoring when raw evidence is too large, externally hosted, or restricted.
See analysis/README.md for the case-by-case data boundary.
Run the local reviewer interface:
python -m temflow patterns-engine --host 127.0.0.1 --port 8765 --data-dir .\user_data_temflowCore HTTP operations are POST /api/err/run and
POST /api/compositional/run. The integrated POST /api/model/run retains the
same map and evidence-selection architecture.
src/temflow/: installable software and reviewer interface.tests/: source tests.FORMULATION/: method definitions and proof obligations.ledger/: append-only evidence and transformation ledgers.analysis/data/: isolated CSV inputs used by the reported analyses.analysis/python/: Python analysis and independent-check scripts.analysis/R/: an independent base-R check of the paper's result table.analysis/results/: frozen scores and checksums.analysis/protocols/: pre-analysis protocols, seals, and amendments.
Downloaded working archives, partial range downloads, build directories, rendered pages, caches, and superseded outputs are deliberately excluded.
The strongest branching test used independently collected UK cattle-movement records. Under a sealed split, 99.53% of positive withheld compartments were inside the intervals, the median width was 48.80% of the unmeasured remainder, and 99.12% of positive same-county retention cases were covered. These are administrative movement counts at county scale, not instrument-weighed mass or farm-scale retention.
An analyst-blind Dryad test covered 88.64% of 440 hidden transient mass states. The source-only Karg test failed the informativeness and unresolved-coverage gates. NASS and Zambia evidence reject a general no-route-means-100%-retention rule. These negative results are part of the release.
TEM-FLOW software is released under the BSD 3-Clause License. Third-party data retain their source licences; in particular, Brazzaville source and derived data are CC BY-NC 4.0. The maintained source and
versioned releases are available at
github.com/TheDOMEMANN/TEM-FLOW.
Add the archival DOI to CITATION.cff after an archival record is deposited.
Version remains 1.0.0. The new aggregate-record calculation reports both ranges for entered readings and worst uncertainty under missing records. Read the input guide. The original published v1.0.0 tag predates this revision and is retained unchanged. Use the revision manifest to identify the code used for the revised manuscript.