The instrument
The certificate obeys two hard rules: it never collapses into a single score, and the measured-versus-modeled boundary and exposure status are legible at a glance. The asset value field reads Not Assessed, always.
Live record
Everything else Voltry publishes is paper and rules. The certificate is the exception: a dark, foil-marked record with a single engraved waveform, depth under the pointer, and a tamper state that is visibly broken when the chain is. Switch the tier to see how evidence depth, not asset quality, changes what the record can claim. Then simulate tamper to watch the Merkle root scramble and the status flip.
Tier is evidence depth. Bronze, Silver, Gold never describe asset quality.
Certificate of Record · Kolibri Engine
VTR-2026-1320A
NVIDIA H100 SXM5 · 80GB · born on fleet
Born on 2026-01-09 · 11,418 operating hours · CONTINUOUS
Certificate of Record · Kolibri Engine
VTR-2026-0A41
NVIDIA H100 SXM5 · 80GB · reconstructed
Born on unknown · single read, no history · POINT-IN-TIME
Anatomy
The order is load-bearing. A reader meets the deterministic facts first, sees exactly where a model enters, and never confronts a single opaque scalar standing in for all of it.
Pass or fail, no model. Authenticity, firmware and VBIOS integrity, functional burn-in, the silent-corruption functional test, and data sanitization to IEEE 2883-2022. Identity, authenticity, and firmware are disqualifying: a unit that fails them cannot be certified as genuine.
Raw values against normal-versus-degraded thresholds. ECC and Xid error-history counts, retired and remapped pages, spare rows remaining, the critical Hopper end-of-life proxy, and throttle, clock, and power-stability reads. No weights are applied.
Each estimate carries its band. Memory wear, hardware health, and thermal and connector fatigue indices, the Modeled Wear Trajectory with a stated-coverage lower predictive bound, the transfer-error tag, and the methodology-version hash and calibration-snapshot ID for replay.
Certification tier, history as born-on or reconstructed, chain gaps yes or no, and Exposure Assessed rendered prominently. The two things a buyer must not miss, whether the modeled fields rest on a single read and whether exposure was assessed, are legible at a glance.
Worked examples
The difference between these two records is exactly the difference the methodology is built to price. The band is tighter in B not because Voltry asserts more confidence but because there is more measurement. Confidence is earned by evidence and shown on the certificate, never dialed up by hand.
A discrete H100 SXM bought on the secondary market, seen once at certification, from an un-instrumented facility. What a buyer should expect for most resale deals.
90% lower predictive bound low, wide band
The same model GPU, monitored from birth in a Voltry-instrumented facility, observed through the infant-mortality phase of the bathtub curve.
90% lower predictive bound higher, narrow band
Tiers
A tier never describes how good the silicon is. It describes how much was observed and how it was observed. A Bronze record is not a low-quality part; it is a part seen with less evidence. A high tier with poor measured facts is entirely possible, and the certificate shows it.
Tier multipliers feed the fee as work, never as a function of asset value. A Gold setup precedes a recurring monitoring subscription. The tier on the face of the record tells a reader how to weigh the modeled fields, not what the hardware is worth.