Principle. Roll out read-only first, prove value with logs and stamps, then enable advisory hooks, and only after a safety case, consider scoped closed loop. Collapse parity must hold at every boundary: phi((m,a)) = m. Clamp everywhere: a := clamp(a, -1+eps_a, +1-eps_a).
5.3 — KPIs to measure (tie to money/time/risk)
• Stability. nan_inf_count, oscillation_area, limit_cycle_count, false_alarm_rate.
– nan_inf_count := count(m is NaN or Inf)
– oscillation_area := sum_t |m[t] - m[t-1]| over the window
– limit_cycle_count := sustained sign-flips of m[t]-m[t-1] beyond a small deadband
• Efficiency. energy_per_task, fuel_per_km, perf_per_watt, scrap_rate.
• Predictability. eta_variance, range_variance, slo_jitter.
• Ops effort. tickets_avoided, hours_saved, rollbacks.
• Quality. rework_count, ooc_reversals, customer_incidents.
SQL sketch (daily roll-up; paste as plain text)
SELECT kpi_name,
COUNT(*) FILTER (WHERE band IN (‘A-‘,’A–‘)) AS low_conf_samples,
SUM(CASE WHEN is_nan_or_inf THEN 1 ELSE 0 END) AS nan_inf_count,
AVG(value_m) AS m_avg
FROM kpi_log
WHERE ts::date = CURRENT_DATE – INTERVAL ‘1 day’
GROUP BY kpi_name;
5.4 — Band design (simple, monotone, auditable)
• Five bands. A++, A+, A0, A-, A--.
• Conservative defaults. A++ >= 0.90, A+ >= 0.60, A0 in (-0.60,+0.60), A- <= -0.60, A-- <= -0.90.
• Alternative (liberal) profile (example). A++ >= 0.75, A+ >= 0.50, A0 >= 0.25, A- >= 0.10, else A--. Declare the chosen profile in the manifest.
• Hysteresis. Require delta_up ≈ 0.05 to move up; allow instant downgrade.
Hysteresis helper (one-liner; paste as plain text)
def hysteresis(prev_band, a, up=0.05):
# upgrade requires crossing threshold + up; downgrade is immediate
if prev_band == “A+” and a >= 0.90+up: return “A++”
if prev_band == “A0” and a >= 0.60+up: return “A+”
if prev_band == “A-” and a > -0.60+up: return “A0”
if prev_band == “A–” and a > -0.90+up: return “A-“
# plain band without hysteresis for other cases
return “A++” if a>=0.90 else “A+” if a>=0.60 else “A0” if a>-0.60 else “A-” if a>-0.90 else “A–“
5.5 — Telemetry and stamping (trust by default)
• Log tuple. time,kpi_name,m,a,band,knobs_hash,build_id.
• Daily roll-up. Append a single stamp line with the day’s manifest and checksum.
• Replays. Any future run must emit the same (m,a) given the same knobs and inputs.
CSV row + daily stamp (plain text)
YYYY-MM-DDThh:mm:ssZ,temp_loop,412.500000,0.812345,A++,3c90..a2,build_abc123,baseline
SSMH-STAMP day=YYYY-MM-DD manifest_sha256=3c90..a2 knobs={“eps_a”:1e-6,”eps_w”:1e-12,”bands”:[-1,-0.9,-0.6,0.6,0.9,1]} build_id=build_abc123 conformance=9b2e..44
Navigation
Back: Shunyaya Symbolic Mathematical Hardware – Deployment Modes & Week-1 Pilot (5.1–5.2)
Next: Shunyaya Symbolic Mathematical Hardware – Readiness & Safety Gates (5.6–5.7)
Directory of Pages
SSMH – Table of Contents
SHUNYAYA 