Own Gooo semantic composition tiny v1
Experimental 12,728-parameter models initialized from our own random state. No Laya/pretrained/earlier own-model weights. They rank legal compiler-owned Gooo paths from typed source plus complete Korean/English intent; they are not unrestricted text code generators. Optional disconnected generation is deterministic.
Six v2/v3 FP32/PTQ/QAT exports retain both improvements and regressions. Calibration selected v3 QAT; development v3 FP32 needs fewer extra attempts. All 384 development finite contracts finish within four candidates in each policy. Warm selected inference: about 9.7 microseconds, zero per-call heap allocations. Ternary weights: 2,759 disk bytes, decoded int8 tensors 12,896 bytes plus scales; 1,248-byte workspace. Whole-process RAM and packed arithmetic are separate.
Actual Gooo main dogfood: 144 calls, 370 predictions, 144 independently compiled Go executions and 2,304 function invocations. Every full 16-case contract passes. Read results.md for language disagreement, retained capture/counter errors, complete-function curves and timing/CPU/RSS measurement scope.
Use released Gooo main with SDK v0.2.11 and an explicit path model:
gooo body-codegen --json --path-plan plan.json --path-model v3/models/qat_ternary/model.json --path-step-attempts 1 --path-feedback-rounds 3 --path-feedback-unfixed --activity ChoosePath source.gooo
The original source and typed plan must bind; the compiler creates the source context. Omit the path model for deterministic continuation. Initial inputs exclude test outcomes; later actual failures may rank remaining paths only.
protocol.md was frozen before program/oracle implementation and training. raw-evidence.zip preserves SDK/native raw captures, actual-Go values and source. Only deliberately public synthetic evidence is included; publication-manifest.json lists exact payload/archive hashes. Native bilingual counter correction is in independent-audit.json; old raw reports are preserved. provenance/*.ttl links each model's weights and metadata to its training/export activity, split roles, own random initialization and compiler source using PROV-O.
Source/trainer/runtime: https://github.com/kimjooyoon/gooo-neural-decision-experiments SDK: https://github.com/kimjooyoon/gooo-decision-runtime/releases/tag/v0.2.11-experimental Compiler main: https://github.com/kimjooyoon/meta-ontology-go/commit/f4813dc6251037767c8cff7295ccfdab2b044ff2 Public source-bound curriculum, including rejected first capture: https://huggingface.co/asketeddy/gooo-compiler-context-tiny-v1/tree/3537f6d3f77ec163fe478d4aef440baee17b9121/research/fresh-composition-curriculum-20261002
These are six compositions with goal, parameter and language variants, one matched training seed and four enumerable candidates; not broad Gooo language understanding or universal semantic correctness. No default model promotion.