ASA × X FOR YOU ALGORITHM / VECTOR-NATIVE TRAJECTORY HOLDOUT
A recommendation system can change its internal trajectory while a familiar headline metric remains stable. ASA-XFA-002 tested whether preserving the full vector structure makes that change observable before it fully forms. The experiment used the pinned, publicly released Phoenix reference component as an external algorithmic environment; ASA did not receive the perturbation label.
ASA-XFA-001 passed every adverse path but also alarmed on every matched control. The semantic Core had been given numeric recommender telemetry formatted as text. That interface was invalid for this domain. The experiment remains frozen as FAIL; its threshold was not changed after reveal.
ASA-XFA-002 introduced a separate vector-native channel. The first ten turns establish a local reference distribution. The channel then measures Jensen–Shannon displacement of the complete 16-dimensional candidate-topic vector and requires persistent pressure across three turns. It outputs an uncalibrated trajectory signal, not a probability.
Six development seeds were used to establish the instrument. Code, thresholds, perturbations, holdout seeds and five PASS/FAIL gates were then hashed. Ten untouched seeds produced three matched cases each: an unchanged control, progressive restriction and structural rotation. ASA observed only the evolving numeric distributions; reveal labels remained outside the scoring input.
No perturbation; ASA remained silent.
Progressive concentration toward one topic.
Composition moved while concentration remained largely unchanged.
HHI compresses the candidate distribution to one concentration number. When topic populations exchanged position, it discarded which component moved and where.
ASA retained the full vector geometry. It registered persistent structural displacement while remaining silent on all ten controls.
Local health can hide global movement. Individual outputs, averages and thresholds may remain acceptable while the joint state is moving toward a different regime.
Long-running systems create trajectories. Agents, recommenders and autonomous workflows change memory, tools, distributions and recovery options across time. A snapshot cannot represent that history.
Observability must preserve structure. ASA is designed to retain relationships, persistence and temporal formation so operators can inspect what is changing before only the terminal consequence remains.
Seed 5105 was not detected by ASA.
Seed 5104 was not detected by ASA.
The complete holdout reproduced bit-for-bit in a fresh process.
RESULT SHA-256 9ac42463f92fbde4d360ba1ca0fea4cfa0229f29483b964c197f290060dad57cCANONICAL SHA-256 603e467dd860e969ba7bb3c748525538f6778c0b5b5b1956dfb4437884ca1981This is an internal frozen holdout using synthetic data generated by a pinned public Phoenix reference component from the open-source X For You repository. It is not a test of the production X service, uses no real-user data, is not independent validation and does not establish calibrated probabilities or production readiness.
Independent ASA Research using publicly released code. Not affiliated with, endorsed by or executed on the production systems of X or xAI.