From 50490b2c13cc35b8cc46029c04de58e433f10551 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Sat, 14 Nov 2026 15:37:55 +0800 Subject: [PATCH] docs(spec): mandate fp32 graph I/O in IMF v1 (owner-flagged follow-up) The runtime guard shipped in interscript-ts #134 (fp16 feeds rejected with a named artifact-contract error, locked by a 1KB dual-branch fixture); the WO17 close-out owner-flagged the SPEC line that makes the contract normative. Weight packing stays free; the mandate is on graph edges. --- SPEC.md | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/SPEC.md b/SPEC.md index 0b48976a..65e42d61 100644 --- a/SPEC.md +++ b/SPEC.md @@ -73,9 +73,18 @@ is the manifest's job. encoder.onnx: decoder.onnx: +Graph I/O dtype: every ONNX graph's inputs and outputs MUST be fp32. +Weight tensors MAY pack to any precision, but the runtime contract is +fp32 edges — fp16 graph I/O is nonconforming (interscript-ts #111: +fp16 edges surface as 0-byte feeds through the Node >= 24 N-API +binding; the repack tool is normalize_fp16_io.py). A conforming +runtime MUST reject fp16-IO artifacts with a named artifact-contract +error, not silently misfeed. + A conforming loader MUST reject: any other `format`, any `tokenizer` -other than `bytes`, missing required members, and uncovered or -mismatched member digests. +other than `bytes`, missing required members, uncovered or mismatched +member digests, and any artifact whose graph inputs or outputs are not +fp32. ## §4 Byte tokenizer