Walk the whole embedding runtime in one command

The last chapter of the book assembles a runtime rather than a claim. This command walks the whole path: it registers two spaces, compares them natively, fits a bridge, derives a third space, judges the derivation, compresses it, applies an operator and explains the lineage.

Run it

git clone https://github.com/ernanhughes/relate.git
cd relate
git checkout 1de8c95b3995791042d6eb1c150c64c7cdb3492f
python -m venv venv
venv/Scripts/pip install -e ".[dev]"
venv/Scripts/python -m relate.cli demo

Python 3.10 or later, NumPy only. No model download, no network, no service.

Expected outcome

SOURCE   demo-coarse [a1a5354ed043547f]
TARGET   demo-fine [70a68f6a3dae3c1c]
NATIVE   CKA 0.86  NN overlap 0.51  counterpart top-1 0.94
BRIDGE   ridge a1a5354ed043547f->70a68f6a3dae3c1c  anchors 320
DERIVED  demo-coarse [11588b564503b381]  (native target hash never reused)
VERDICT  retrieval         PASS
VERDICT  threshold_transfer FAIL
WHY      FAIL — threshold_transfer / - failed: calibration_transfer/far_increase = 0.075 (requires far_increase max_delta 0.05; frame target_native, reference 0.1)
COMPRESSION
VERDICT  compression/retrieval         PASS
VERDICT  compression/threshold_transfer FAIL
OPERATOR
VERDICT  operator/constant_delta    PASS (1.000)
VERDICT  operator/identity_map      FAIL (0.579)
DOCTRINE transform -> derive identity -> measure -> scoped verdict; lineage composes, permission does not.

Read it in order. The first three lines are the native comparison: two seeded spaces, and three different measures of how alike they are, none of them agreeing. The bridge is fitted from 320 anchors. The derived space has a different hash from the native target, and the run says so in brackets.

Then the verdicts, and this is the part that matters: the runtime does not return one answer. It returns a scoped verdict per capability, and two of them FAIL. The failure names the threshold, the measured value, the limit it exceeded, and the frame it was compared in.

Lineage composes, permission does not

The last line is the chapter’s summary and it is also the reason the verdicts are scoped. The steps compose: a transformation, then a derived identity, then a measurement, then a verdict. Permission does not compose the same way. A PASS on retrieval says retrieval passed; it does not carry over to threshold transfer, and the run reports the second as a separate FAIL rather than inheriting the first.

operator/identity_map FAIL (0.579) against operator/constant_delta PASS (1.000) is the same point about operators: the right operator for one transformation is not the right operator for another.

Mechanism and limitations

Documented: the demo’s steps, the demo space fixtures and the CLI contract are in src/relate/cli.py; the reconciliation of every chapter-26 component to its implementation object and benchmark is in evidence/CHAPTER-26-RECONCILIATION.md in the same repository.

Observed: the run above. The native comparison reported CKA 0.86, nearest neighbour overlap 0.51 and counterpart top-1 0.94; the bridge was fitted from 320 anchors; the derived space carried a different hash from the native target; and four of the six scoped verdicts failed, each with a stated reason.

Not shown. The demo spaces are synthetic and seeded. The two FAIL verdicts are the demo’s own fixtures failing their own criteria, which demonstrates that the gate can fail, not that these particular capabilities are hard. The 1.000 on the constant-delta operator is a fidelity on the demo’s pairs. No claim is made here about real encoders, real corpora or real bridges, and the lineage the run prints is a lineage over synthetic objects.

Canonical source and the chapter-26 reconciliation .