environment/models-agent/supplied/README.md

SDL-2 v1.3 agent copy. In every package here, posterior_tool.npz holds a random subset of the stored posterior rows (about half of each density and magnetic mechanism); the rows that are not shipped are reserved for hidden worlds. The hidden lot is drawn from the same posterior, never from a shipped row, and its anneal parameters come from an independent run of the same anneal fit. Sentences below that call a file byte-identical to another release describe the original release, not this copy.

SDL-2 v1.2 episode. These are the live campaign's supplied packages, byte for byte except that each package's computed decision rankings are omitted (see its README). In this episode request_simulation(simulation="magnetic") is the virtual lot described in START_HERE.md, which returns numeric hypothetical loops; it is not the C7.3 twin's model distribution.

Supplied models — current entry point

Current: C7.3/ (October 23, 2026, corrective). The shared physics tool is ndfeb-sps-physics-tool-20261023.8.3, and the campaign twin is ndfeb_c7_campaign_v1 (lot CAMPAIGN-C7.3-LOT), which answers request_simulation(simulation="magnetic").

• What changed, and what C7.3 and C7.2 correct: C7.3/README.md. Changes since C6: the SPS density update on #7/#8, the anneal-densification layer, stage-specific densities, one alignment input, rejection of unknown route keys, and joint-target probabilities on both magnetic paths. Inherited C7.1 corrections: segmentation-consistent annealing, one hold dependence on both magnetic paths, and anneal_eig() / condition_anneal() for anneal experiments.
• What each model consumes and what it is calibrated on: C7.3/INVENTORY.md.

Superseded the same day, kept byte for byte: C7.2/ (ndfeb-sps-physics-tool-20261023.8.2). Its condition_anneal() keyed reading identity on the source string, so a reading entered under another source, evidence id or specimen label could count twice. Use it to reproduce C7.2 results.

Superseded the same day, kept byte for byte: C7/ (ndfeb-sps-physics-tool-20261023.8.0). Its CAPACITY and CLOSED anneal mechanisms depend on how a program is segmented. Use it only to reproduce C7 results.

Earlier, kept byte for byte: C6-development/ (C6 v2, ndfeb-sps-physics-tool-20261018.7.0-dev, density cutoff P3). Use it to reproduce earlier work and forecasts. Its twin family ndfeb_c6_campaign_v1 and its virtual specimens remain records. C0–C5 sit under library/tools/ and the source archive.

All packages load in one Python process (import each ndfeb_physics_tool.py under its own module name; each calls load()), so you can compare them directly. The packages are model hypotheses: neither is a measurement, and neither is a replacement for a model you adopted yourself.

C7.2 added observation identity, duplicate rejection and consistent initial-piece evolution; C7.3 resolves that identity against the released ledger. C7.1 remains unchanged for replay. P7 adds images, not a numeric calibration. See C7.3/README.md for qualified density-route conclusions.