Skip to main content

​
The constants

A backend must satisfy all three simultaneously to earn the THRINDEX Certified [v0] badge:

​
Fixture set

The ratification fixture set is frozen at CRC32 e2ebd845 (120 SHD test samples, stratified 6 per class × 20 classes). It lives at conformance/fixtures/shd_ratify_v1/frozen/ in the monorepo.

​
Quantization model

The envelope was calibrated against per-channel symmetric int8 — the hardware-class quantization that real neuromorphic digital chips implement:
  • Scale per output channel: scale_n = max_abs(row_n) / 127
  • Rounding: round-half-to-even (banker’s rounding, matches thrindex-numerics)
  • Clamp: [-127, 127]; dequantize: q as f32 × scale_n
  • Biases: left in f32

​
Ratification evidence

Measured on the frozen 120-sample fixture set. Reference model: templates/keyword-spotting/model.thx (Dense 700→512→20, τ_mem=20ms, 64.66% SHD test accuracy). INT8 per-channel — should PASS (this is the hardware model the envelope targets): INT4 per-channel — should FAIL (confirms thresholds are not vacuously loose): The separation is decisive. There is no metric where int8 and int4 are ambiguous.

​
Threshold derivation

T_mean = 0.020 — the provisional value was confirmed by data. INT8 sits at 26% of the limit (3.8× headroom); INT4 at 419% (4.2× above). The 15.9× int8/int4 gap means the threshold is well-placed: tight enough to catch INT4 decisively, loose enough to accommodate INT8 with margin. T_max = 0.130 — the provisional value (0.100) equaled the observed INT8 agg_max exactly (zero margin). One extreme sample had a neuron differing by 10 spikes over 100 timesteps (rate error = 0.10). Adding 30% headroom gives 0.130. The auto-derived 40%-into-gap value (0.428) was rejected as too loose — certifying a backend with worst-neuron divergence of 43% has no meaningful inference interpretation. 0.130 passes INT8 with margin and fails INT4 (0.920) by 7.1×. P_min = 0.900 — the provisional value (0.990) was unreachable: INT8’s observed overall agreement was 0.983, with boundary-case agreement 0.9375 (15/16 near-tie samples). 0.900 seats below both with 8.3% and 4.2% headroom respectively. INT4 at 0.617 fails by 32%. Semantic interpretation: the backend must agree with the float reference on ≥90% of test inputs.

​
V0 scope

This envelope was ratified against one reference model (SHD keyword spotting, Dense 700→512→20, 64.66% accuracy, 120 frozen test samples). It brackets per-channel int8 divergence on that specific architecture and weight distribution. It is not yet a universal constant. A backend certified under V0 carries the badge THRINDEX Certified [v0].

​
Reopening triggers

Any one of the following mandates a V1 measurement:
  1. A structurally different model enters the conformance suite (recurrent layers, different T, Conv2d-heavy, or substantially different weight distribution)
  2. A real silicon backend is added whose quantization characteristics differ from per-channel int8
  3. A latency-code or time-to-first-spike model enters the suite

​
Re-certification

Tightening any threshold is a breaking change: V0-certified backends must be re-certified under V1. Loosening requires the same approval path. The harness supports --envelope v0 indefinitely for retesting. The V0 badge is retained in published history but marked superseded when V1 ships.