SSMT – ML Hygiene, Sensor Fault Detection, and Fleet Snapshot Testing (4.9–4.11)

Cleaner models, honest alarms, and replayable evidence — all from one symbolic channel.

Context.
This page shows how Shunyaya Symbolic Mathematical Temperature (SSMT) improves machine learning stability, exposes bad sensors instantly, and produces evidence you can take to audit, safety review, or insurance. This is where SSMT stops being “a better temperature reading” and becomes “a defensible operational layer.”

4.9 ML feature swap (hygiene, no re-tuning per unit)

Problem.
Most analytics and ML models today are fed raw temperatures in °C, °F, or K. That causes three recurring problems:

  • You have to rescale the feature for each site or device.
  • Outliers (spikes, stuck sensors, unit jumps) destabilize training.
  • Porting a model from one environment to another means weeks of re-tuning thresholds.

SSMT approach.
Replace raw temperature inputs with symbolic signals that are already centered, bounded, and comparable:

  • Use e_T (unitless contrast around T_ref) instead of raw °C/°F/K.
  • Optionally add a_phase if survival-side matters (freeze, warp, stress zone).
  • Optionally add Q_phase if you want “sustained risk memory” instead of frame-by-frame jitter.

Result:

  • e_T is zero at the declared baseline, positive when hotter, negative when colder.
  • a_phase and Q_phase are bounded ((-1,+1) or [0,1]) and flicker-resistant.
  • Thresholds carry over to new locations without renormalizing everything.

In practice:

  • The same model input spec works across multiple sites.
  • You do not keep rewriting “what counts as hot” every time you deploy.
  • False drift alarms drop because the feature space stops jumping just because someone expressed the same physics in a different unit.

This is ML hygiene: strip unit drama at ingest, not in post-processing.

4.10 Fault detection (unit flip shock)

Problem.
When a data stream silently flips from °C to °F (or any other scaling shift), downstream systems often don’t notice until after a bad decision. Humans only catch it later during forensics.

SSMT approach.
Because SSMT does a one-time “to Kelvin” and then produces e_T, any sudden unit flip shows up as a hard step change in e_T. That is both detectable and auditable.

If a stream flips °C↔°F by mistake,
the one-time to-K step isolates it as a sudden level shift in e_T;
drift monitors catch it.
Action:
  set health.sensor_ok = false
  flag drift = true
  trigger investigation
  keep thresholds in symbol space (do not silently adapt)

Key point:

  • We do not “auto-heal” the stream.
  • We explicitly mark it unhealthy (sensor_ok = false) so nobody can pretend it was valid data.
  • We keep alarm thresholds in symbolic space, so policies don’t get corrupted by garbage input.

This is compliance-grade behavior. It creates an objective trail: “the feed was marked unsafe at time T, decisions after that were quarantined.”

4.11 Empirical snapshot (symbol space across multiple stations)

Goal.
Show that SSMT’s dials (e_T, a_phase, Q_phase) reduce chaos in practice: fewer flicker alerts, more stable thresholds, and clearer portability across a group of stations — without rewriting physics.

Setup.

  • log lens
  • T_ref = 298.15 K
  • Phase dial near a critical transition with
    T_m = 273.15 K, DeltaT_m = 2.0 K, c_m = 1.2, eps_a = 1e-6
  • Soft hysteresis with
    rho = 0.90, k_side = 2.0
  • Symbolic alert rules:
    • Freeze-style rule: a_phase <= -Phi_freeze with Phi_freeze = 0.10
    • Hot-style rule: e_T >= E_hot with E_hot = 0.8
Rules.
Freeze (symbolic): a_phase <= -Phi_freeze
Hot (symbolic):    e_T >= E_hot

Observed slice.
Across 5 stations:
• Phase-aware rules reduced alert flicker by 10.45% (134 -> 120 toggles).
• 4 of 5 stations showed equal or lower flicker after using Q_phase-based stability.
• A synthetic unit flip (as if someone changed °C↔°F mid-stream) produced a visible step in e_T,
  automatically set sensor_ok = false and drift = true,
  and symbolic thresholds did not move.

Hot threshold portability:
• Using the same E_hot for all stations did not create spurious "hot" flags,
  meaning no site demanded its own private threshold in this slice.

Meaning.

  • Flicker dropped when alerts were driven by Q_phase (sustained state) instead of raw temperature crossings.
  • Station-to-station fairness got better because everyone used the same symbolic thresholds (E_hot, Phi_freeze) instead of “whatever our meter says is hot.”
  • Sensor tampering / miscalibration / unit error became obvious immediately in the symbol channel.

Why this matters.
This is the bridge from math to governance:

  • You can quantify “before SSMT vs after SSMT” using flicker count, false toggles, time in sustained risk, and threshold portability.
  • You can show that the alarm policy was stable, explainable, and replayable.
  • You can defend operational choices without arguing about raw units.

This is how temperature stops being a loose number and becomes part of an audit trail.

Navigation
Previous: SSMT – Cross-Site Comparability, Rail Stress, Fuel Survival, and Spaceflight Cryo (4.5–4.8)
Next: SSMT – Tier S1 Validation: Proving the Math Holds Before Touching Real Data (5.0–5.3)

Directory of Pages
SSMT – Table of Contents