A proposed field adds no separation over the existing fence: a separability census
On a 96-cell synthetic grid the candidate field partitioned objects exactly as the existing fence (symmetric difference 0); 12 of 12 thresholds and 6 of 8 predictions held; 12 of 96 cells were reachable.
Mixed resultCompared with simpler mechanismsWorking note
- Designed for
- Built to test whether a field proposed against a hostile model, the receipt e-mail, separates lawful from foreign objects better than the ownership fence the guard prototype already had.
What was tested
An instrument reproduced, through tool calls on the deterministic probe, hand arithmetic for candidate predicates (a receipt e-mail field, an amount equal to the order total) on a grid of 8 anchors and 12 objects. Predictions and thresholds were computed by hand beforehand; no language model ran.
What we observed
The candidate’s partition and the existing fence’s partition were identical: symmetric difference 0. Thresholds held 12 of 12; predictions held 6 of 8, two being refuted as written. An unrepelled foreign class grew 2, 4, 6 and 10 across four corpus sizes. Only 12 of 96 cells (12.50%) were reachable in the world and 10 of 96 through the guard, with one anchor where the two diverge.
What this means for Heimdall
This asks whether a field proposed against a hostile model separates objects better than the guard prototype’s ownership fence. It did not: both drew the same line, so on this fixture the existing fence already did that work. The low reachability also weakens the earlier shared-grid comparison.
Limits
One synthetic fixture and four ladder points, so growth claims are not supported; the document stating the premise is not in the tree; nothing here measures real hostile model behaviour.
What remains open
None planned on this fixture.