Psi Anomaly Detector

The Psi Anomaly Detector is the umbrella software stack that classifies events from Psi Recorder, Field Recorder, Psi Compass, Psi Tripwire, Schumann Lattice, and Star Seer Observatory feeds against the Clan's Psionic Threat Model signature library. It is the analytical layer between raw sensor streams and the Resonant Finder; engineering-honestly, a multi-modal anomaly-detection + classification pipeline with provenance and human-in-the-loop review.

Overview

The Detector does, per event:

• Window — temporal window selection from raw multi-sensor stream.
• Feature — frequency-band power, gradient magnitude, electrostatic transients, IMU correlation.
• Classify — multi-class classifier (small CNN or gradient-boosted tree per modality), majority + confidence vote.
• Provenance — every classification carries the model version, training-data hash, signature-library version.
• Review — high-confidence high-class events surface to a human reviewer; low-class events log only.

It is, importantly, the same code stack that runs in the Psi Defender (on the HelmKit, low-power), the Psi Tripwire (on the deployable node, low-power), the Field Recorder (post-session batch), and the Resonant Finder (full-resolution farm). Code reuse is the discipline.

Theoretical Basis

1. 2e7d32;"

Epistemic status: [[ Category:Provisional Psi Claims|PROVISIONAL]]

Multi-modal anomaly detection with classifier ensembles is mature ML practice.

1. 2e7d32;"

Epistemic status: [[ Category:Provisional Psi Claims|PROVISIONAL]]

Provenance-tracked ML inference (signed model + signed inputs) is mature MLOps practice.

1. 1565c0;"

Epistemic status: [[ Category:Testable Psi Claims|TESTABLE]] Theory anchors: Resonant Neurobiology, Psionic Threat Model

A signature-library-trained classifier on labeled multi-modal data outperforms single-modal baselines on the Clan's anomaly catalog. Falsifiable ROC.

1. c62828;"

Epistemic status: [[ Category:Speculative Psi Claims|SPECULATIVE]] Theory anchors: Psi Field

A subset of high-confidence anomalies are Psi Field events. The Detector's non-psionic outputs (general EM anomaly classification) stand on their own.

Subsystems and BOM

Mk0 (software-only, target ≤ \$0)

• scikit-learn pipeline on Pi-class host
• Hand-labeled signature library (≤ 50 entries)
• Python notebook review interface
• CSV-based session ingest

Mk1 (per facility, target ≤ \$3 000)

• GPU box for training (single RTX 4060-class) — \$800
• Storage (8 TB NVMe RAID) for labeled dataset — \$600
• MLflow + DVC infrastructure (open-source) — software
• Signed-model registry on HSM — \$300 (HSM)
• Reviewer workstation — \$1 000
• Network + UPS — \$300

Per-device Mk1 footprint: ≤ 10 MB ONNX runtime; nRF52840-runnable for the smallest models.

Mk2 / Mk3

Mk2: federated signature-library curation across Observatories with documented byzantine resilience. Mk3: allied-EA shared signature library with cross-domain signing.

Build Notes

• Library curation is the product. Bad signatures poison every consumer downstream. Every entry requires capture + repeat observation + Psionic Threat Model documentation.
• Model versioning. Every deployed model has a signed registry entry; rollback documented.
• Per-device profile. The same source library compiles down to a HelmKit-runnable subset, a Tripwire-runnable subset, and the full Finder-side model.
• Reviewer ergonomics. Reviewers are scarce; the tooling must surface batched events, not individual alerts.

Safety and Ethics

• No personal data in training set. Environment + device state only; operator HRV used only with consent and only for Resonant Neurobiology research builds.
• No automated action on low-confidence classification. The Defender's counter mode requires high-confidence; the Tripwire's alert requires medium.
• Open signature schema; allies can ingest and audit.
• No use against named individuals. The Detector classifies events, not people.

Maturity (Mk0 → Mk3)

See Tho'ra Tech Maturity Levels for the convention.

Mark	Phase	Status	Confidence	Evidence Base	Next Validation Gate
Mk0	Cosplay-type	Complete	100% (symbolic)	Ritual + build practice	User satisfaction
Mk1	Prototype	Active	—	—	—
Mk2	Test-type	Planned	—	—	—
Mk3	Production	Projected	—	—	—

Failure Modes and Mitigations

• Library poisoning by bad signature. → Multi-author signoff required; signed library entries; rollback.
• Concept drift. → Per-deployment baseline; periodic retraining cycles.
• Model registry compromise. → HSM-signed registry; rotation; trust-root revocation.
• False-positive cascade. → Per-class confidence thresholds; rate-limit alerts; reviewer ergonomics.
• Device-side OOM. → Per-device profile pinned to model footprint budget.
• Operator over-reliance. → Confidence + provenance surfaced in UI; the Psionic User Interface doctrine applies.

See Also

• Psi-Tech
• Psi Defender
• Psi Recorder
• Psi Tripwire
• Field Recorder
• Schumann Lattice
• Star Seer Observatory
• Psionic Threat Model
• Resonant Finder
• Psionic User Interface
• Tho'ra Tech Maturity Levels