Determinism

The claim, precisely#

Identical input bytes + identical artifact ⇒ identical integer outputs — latent_int, band, pd_ppm, every impact_int, and every reason code — on any conforming runtime, any hardware, any language.

Why it holds#

Floating-point nondeterminism comes from rounding under accumulation: (a + b) + c ≠ a + (b + c) in floats, so summation order, vectorization, and fused-multiply-add all matter. CompileML removes the accumulation entirely:

1. Compile-time quantization. Every leaf value becomes an integer once: value_micro = rha(leaf × learning_rate × 10⁶). The float model is then discarded — the integer model is the model.
2. Integer everything. Scoring sums value_micro in int64 (associative — order cannot matter). Band edges are integers. The calibration table is integers, interpolated with an integer division formula (spec §2.2). Attribution is integer differences.
3. The one float op left is comparison. Tree routing evaluates x <= threshold. IEEE 754 comparison is exact — it involves no rounding and no arithmetic. Given the same input bytes, every machine routes every row identically.

For models whose source framework compares in float32 internally (XGBoost hist trees), the artifact records input_precision: "float32" and every runtime quantizes inputs to binary32 before comparing — so even that subtlety is reproduced identically everywhere.

What is not claimed#

• Upstream float pipelines. If your feature pipeline produces different bytes on different systems, decisions can differ. Producing identical input bytes is the caller's contract; CompileML's contract starts at the feature vector.
• Cross-artifact equivalence. Two artifacts compiled from the same model are two artifacts. The unit of governance is the hash.
• Population-level stability. Determinism says nothing about drift in who applies. It says your measurement of drift is exactly reproducible — any change in a monitored metric is attributable to the portfolio, never the tooling.

How it's tested#

• The SQL export executes in a real engine and must match the Python runtime integer-for-integer on every row.
• The COBOL export compiles under GnuCOBOL in CI and is diffed the same way.
• Rebuilding an artifact from identical inputs must reproduce the identical hash (no timestamps or randomness live in hashed content).
• CI runs the full suite across operating systems and Python versions and compares artifact outputs byte-for-byte.