Files
ARCADIA-Studio/harness/experiments/ordinary-source8-20260925/CHECKER-NOTE.md
T

2.5 KiB

Checker encoding correction

The first actual data-only checker run refused at inspect_runtime.py:354, assert (fixture / name).read_bytes() == text.encode(). Exit1; no successful readback JSON was written. The unchanged CLI project fixture was149bytes, whereas the executed checker incorrectly expected153bytes. Participants106bytes and children85bytes matched.

checker-first-refusal.log preserves the actual traceback text returned by exec for that run (transcribed from the tool result, not a separately redirected raw stderr file and not a replay). The preserved original checker was actually executed and refused; it is not an unrun predecessor.

This was an evaluator authoring defect: a Windows default-encoding read_text() source rewrite turned the intended literal 雨 into mojibake. Actual fixture bytes were UTF-8 e9 9b a8; the executed checker literal contained c3 a9 e2 80 ba c2 a8 (雨). Python reads its source as UTF-8, and parsing the preserved raw checker bytes confirms153bytes. The actual CLI fixture and captured original/after inventories remained correct. No product/source/fixture or runtime receipt changed.

Independent packet audit corrected an error in the first version of this note: it had claimed161bytes and a second encoding expansion. That result came from the author's secondary diagnostic probe, which called ast.parse(p.read_text()) using Windows default encoding on the already corrupt UTF-8 checker file. Repeating that exact data-only probe yields161bytes and prefix c3 83 c2 a9 c3 a2 e2 82 ac c2 ba c3 82 c2 a8; it is not what Python executed. Parsing raw.decode('utf-8') instead yields the correct153-byte account above. The first packet candidate is preserved locally; its note is superseded by this explicit correction. Original failed checker, refusal transcript, corrected checker, successful readback and application evidence are unchanged.

Preserved failed checker bytes as inspect_runtime_encoding_original.py, SHA256 b544cdb20dc8f5cf1b5d29c5ab78791dd3ac6701a4c1f7f7229f0d73dac7b81e. The only correction replaced that project-comment literal with Python's explicit \u96e8 escape using an explicit UTF-8 read/write. Corrected checker SHA256 db853cc7677421e1abc13e59ca7d181fdcf7bcb45033164e51e09f13ca51c172 then exited0 and produced ORDINARY-RUNTIME-READBACK.json, SHA256 3ee36b331431736062db7ea9fa94876795f61711d44820f5b5093e97cc515db5.

No comparisons were removed or normalized; complete expected UTF-8 bytes still must match. Only Git and data reads ran, never Cargo or product executables.