Running the measurements

If you want the result and the intuition behind it, start with DETECTION.md. This page is for people who want to run the measurement code themselves.

The most important rule is simple. A comparison may use the known matching human when it is asking how close two populations are. A detector may not use that history when it is asking whether an unknown person's movement can be accused of being generated.

The code keeps those two jobs separate.

Select a Renderer by sampled full-session texture

Renderer checkpoint selection is separate from detection. It uses uninterrupted dense sessions because teacher-forced loss can improve while a free-running sampler gets worse. Score a fresh float checkpoint with:

python -m evaluation renderer-selection full_sessions/sessions.json \
  --backend float --model runs/renderer_p118345.pt \
  --specs w3 w5 --seed 7001 \
  --output runs/renderer_p118345_selection.json

Use --backend native without --model to score the authenticated packed release. The evaluator verifies source hashes, observes the first 256 physical reports once, carries one rollout through the complete remaining session, and reports the carried-session T/R/Z/D/S user-macro score. Its JSON replaces private IDs with ordinal aliases and records source-content hashes. The supplied panel determines the meaning of the result; the private eight-session panel is not bundled. One invocation scores one artifact and one draw seed. The frozen multi-model, multi-view, multi-draw aggregation must be assembled from separately retained cells.

Try the public example first

This builds one human/generated descriptor bundle from the included test movements:

python -m pip install -e ".[evaluation]"

python tools/build_descriptor_bundle.py examples/aim_test.npz \
  results/local/descriptors.npz --rows 256 --assume-quiet-preroll

That flag is explicit because the compact example begins at A and lacks 96 earlier reports. It left-pads the missing history with quiet reports for a smoke run only. For a real result, omit the flag and supply renderer_context_raw_dxdy [N,256,2].

That exact per-row context is part of this evaluation protocol. It is intentionally different from the recommended deployment schedule, where an application prepares one representative RendererProfile before B and reuses it across events. Do not silently substitute the deployment profile when claiming to reproduce the frozen tables, and do not treat these exact evaluation windows as a real-time requirement. The separate one-draw engineering probe is explicitly bounded in renderer_profile_sensitivity.json.

You can then measure human variation and run the labeled judge:

python -m evaluation floors results/local/descriptors.npz --panel full
python -m evaluation judge results/local/descriptors.npz --panel full

floors knows which sessions and identities match. That is intentional because it is measuring distance, not trying to identify an unknown source.

judge is given human/generated labels and learns the easiest population-level difference it can find. Its held-out AUC tells us whether some difference remains, but it does not tell us whether a real-world detector can use that difference without false accusations.

The small public bundle can also exercise a reduced version of the unknown-person code:

python -m evaluation cold-smoke results/local/descriptors.npz \
  --panels trajectory texture full

Its output says not_release_protocol: true. It is a useful code check, not a reproduction of the final study.

Build a complete warm/cold audit bundle

For a real audit, first generate at least two independent full-pipeline cells over the same source roster. --rows 0 keeps one B80-nearest row from every physical source:

python tools/build_descriptor_bundle.py events_with_renderer_context.npz \
  results/local/draw_a.npz --rows 0 --event-seed-domain draw-a

python tools/build_descriptor_bundle.py events_with_renderer_context.npz \
  results/local/draw_b.npz --rows 0 --event-seed-domain draw-b

python tools/build_audit_bundle.py \
  --cell draw-a=results/local/draw_a.npz \
  --cell draw-b=results/local/draw_b.npz \
  --output results/local/enriched_audit.npz

The first tool records the generation-known target geometry as causal context and keeps each generated row bound to its human source. The second verifies that human rows are identical across cells, freezes whole installation keys into reference and held populations using a declared SHA-256 seed, and selects exactly 32 audit rows per held session without reading generated outcomes.

The exact warm protocol also needs a disjoint 32-row validation group and at least 48 trusted reference rows in every held session. The builder therefore requires at least 112 human sources per held session, at least two held sessions, and at least two reference plus two held installation keys. It fails with a direct explanation rather than creating a smaller study under the same name.

To evaluate a retrained Renderer instead of the native artifact, pass the same --float-renderer-checkpoint runs/renderer_p118345.pt to both descriptor commands. Use distinct --event-seed-domain values to redraw both the Planner head and Renderer sampling, keeping the complete pipeline cells independent.

Running the real unknown-person test

The release command is cold:

python -m evaluation cold results/local/enriched_audit.npz \
  --panels trajectory texture full

For every identity being judged, the exact runner removes:

• the complete installation key;
• every human session stored under that key; and
• the matching generated rows and generator cell being evaluated.

It learns its directions from other identities. It sets its boundary using human movement only. Generated outcomes are inspected only after that boundary is fixed.

This is what makes the result a genuine unknown-source false-positive test instead of a same-person comparison in disguise.

What is inside a descriptor bundle?

Every command reads a pickle-free .npz with schema abcurves.descriptor_bundle.v1.

Array	Shape	What it contains
features	[N,D]	Finite trajectory, texture and combined descriptors
origin	[N]	human or generated
installation_key	[N]	Participant ID held out with all their sessions by the cold test
session_id	[N]	Recorded session boundary
source_id	[N]	Physical source-trial binding
order	[N]	Event order for chronological groups
task	[N]	Task or challenge stratum
feature_names	[D]	Names of the descriptor columns
panel_slices	[P]	name:start:stop declarations for each panel

The exact release studies need a little more custody information:

Array	Shape	Why it is needed
population_role	[N]	Separates reference humans from wholly held evaluation rows
generator_cell	[N]	Keeps model seed and draw cells separate during fitting
target_role	[N]	Preserves the task/role comparison stratum
causal_context	[N,C]	Stores information available at generation time, never future outcomes
block_order	[N]	Preserves stable ordering inside chronological groups
audit_panel	[N]	Marks the frozen 32-row evaluation panels
audit_order	[N]	Keeps frozen within-session order; non-panel rows may use -1

Generated rows carry the installation key, session, and source binding of the human context they continue. Exact cold mode therefore removes the target key's human and generated rows together. Exact warm mode aligns each candidate cell to the same frozen physical-source roster.

The exact cold and warm commands require these fields. They never silently fall back to the smaller smoke calculation.

The known-history experiment

There is also a warm command. It asks a different and easier technical question: what can a detector notice when it already owns a large, trusted set of clean human movement from the same recorded session?

That is deliberately a best-case laboratory condition. In ordinary use, sensitivity, mousepad, grip, posture, fatigue, and habits can change enough that an older reference no longer describes the person's current movement. Warm results therefore measure what perfectly matched history can reveal; they are not a promise that a real-world baseline stays valid.

Its primary gate is the maximum over cross-fitted Trajectory14, Texture19 and Full49 directional dense, sparse-tail/Berk-Jones, and subgroup statistics. Here alpha is the accusation cutoff: a smaller value requires stronger evidence. A disjoint human validation group selects alpha=0.0025 before the frozen panel or any generated outcome is evaluated.

python -m evaluation warm-smoke results/local/descriptors.npz \
  --installation-key KEY --session-id SESSION

python -m evaluation warm results/local/enriched_audit.npz \
  --panels trajectory texture full

The smoke command is only a small implementation check and marks its result not_release_protocol: true.

That experiment is useful when studying the judges, but it is not used for the unknown-person claim. It lives here for completeness and for people deliberately working on detection.

Exact release defaults

The unknown-person route uses 32-row groups and 16 nested mixture ledgers. The known-history route uses 48 context neighbours, 512 null-fit draws, 2,048 disjoint null-calibration draws, 32 nested ledgers, and the predeclared contamination schedule. These are the defaults behind the compact release results, not settings inferred from the bundled smoke example.

What can be reproduced from this repository?

The public repository can:

• verify every compact result file by SHA-256;
• build the same trajectory, texture and Full49 descriptors;
• run and test every split and statistic;
• run the exact algorithms on another correctly enriched corpus; and
• check the inference path and the descriptor/evaluation smoke paths on the bundled examples.

python -m evaluation verify-results
python -m evaluation verify-results results/inference/manifest.json

The final tables use the frozen 320-row panel and contributed hardware recordings. Both are publicly reproducible through the raw Release attachments and compact recipe in DETECTION_REPRODUCTION.md. That guide preserves the exact B80 selection, historical generation runtime and separate Renderer-headline aggregation; generic fresh bundle selection is not equivalent. The two compact receipts in results/detection/ retain the original numerical outcomes and sealed source digests.