asketeddy/gooo-record-field-tiny-v1

Gooo record field tiny v1

Gooo가 작성한 조건문과 변수, 레코드 필드 안에서 작은 모델이 조립 순서를 제안하는 실험입니다. 설계도인 Gooo 소스에 허용된 부품을 적고, 모델이 그 부품의 조합을 먼저 골라보는 방식입니다. 실제 실행에서 맞는 필드와 남은 부분을 각각 셉니다.

이번 모델은 제목 복사·상태 설정·사유 뒤에 문자열 붙이기라는 세 역할을 대상으로 직접 초기화하고 GPU에서 480회 학습했습니다. Go가 추론과 실행을 담당합니다. 학습에 쓰지 않은 192개 소스에서 첫 후보 일치는 FP32 192개, QAT 134개였습니다. 같은 선택지에서 의도를 반대로 바꾸면 첫 후보 필드 충족률이 50%로 떨어졌으며, 이 실패도 함께 공개합니다. 서로 반대되는 의도를 짝지은 작은 학습 자료를 늘리는 것이 다음 단계입니다.

Public preparation/training source and Go evaluation/publication source. Use the SDK's LoadRecordThree and PredictRecordInto entry points or the compiler CLI example below. This custom Go model format has eight mask scores. models/fp32, models/ptq_ternary and models/qat_ternary contain independent metadata and weight files. The complete executed pilot is in evidence.zip.

Learning from actual Gooo field choices

Compiler dev PR1229 and main PR1230 each passed all 12 CI jobs with independently verified source-bound evidence. Main aeff3641254ac5f795fa1fb8c94bed702f10cbe4 is installed with Go1.27.1 and SDK v0.2.22. Research PR15 passed all 24 jobs and merged as 2a8f21794bfe42a9bdee22f5e6218ea9eb7499ce.

The installed compiler completed a separate 100 constructions/200 native runs. All finite counts and generated Gooo/Go/driver identities match the original candidate, including counter-intent failures. Native execution adds zero model calls. Two worker goroutines also completed eight record constructions with 120/120 selection fields; record graph execution uses body-compose. installed-native.zip retains the complete separate installed observations. release-validation.json pins these releases and their CI/source identities.

In the counter-intent budget-one result, the 12/24 matching record fields belong to cases that return their unchanged input. On the fields requiring an actual transformation, the fresh models matched 0/12, deterministic order 8/12 and the frozen ordinal control 4/12. This makes the need for paired opposing intents explicit. All weights, calibration settings and initial pilot results remain the recorded v1 exports.

Gooo describes a small construction space: conditions and local assignments, three record fields, two permitted expressions per field, intent and examples. Our own small model orders the eight possible combinations. The compiler assembles those pieces and measures the resulting behavior. A useful analogy is a drawing with numbered parts: the model proposes an assembly order, and execution shows which parts fit the requested drawing.

This pilot adds actual field expressions to the model context and makes ordinary conditional bodies easier to write. if can fall through without else, including early-return guards and nested local updates. Every reachable return, variable scope and assignment type is still checked by the compiler.

What was executed

• Clean compiler source: af811fe3b5dbb709edd61fe9de216f6af0850911.
• Go SDK: v0.2.22-experimental; feature contract triple_record_field_context_v1_joint_v1.
• Published preparation/trainer source: 4441c221dfa1b13230d9469d56974b5d7742b174.
• 768 source views: eight expression/body families × six field orders × eight first/second orientations × Korean/English. Whole-family split: 384 train, 192 calibration, 192 held-out test. Exact cross-split contexts/features overlap zero times. The three roles are title copying, ready state and reason suffix.
• Fresh Go-seeded initialization, 240 FP32 plus 240 QAT optimizer updates on MPS. Calibration selects epoch and temperature. Test rows are first used in the separate Go audit. PTQ adds zero updates.
• 192 export/Go logit comparisons; largest absolute error 0.00000304.
• 100 native constructions, each built once and run twice: 80 observations on four held-out source views with new inputs and 20 on paired counter-intents. Five profiles and budgets 1/2/4/8. These are repeated construction observations; the separate goal of 100 distinct experimental approaches remains open.

The compiler's finite selection examples determine completeness. The model context carries complete alternatives and intent, with zero test inputs or expected outputs. Inference occurs once before construction; native execution and repeated execution perform zero model calls. Disconnected construction uses deterministic ordering.

Complete-mask ordering on 192 held-out views

Export	First candidate	Within two	Within four	Warm Go prediction median	Weight file	Resident tensors
FP32	192/192	192/192	192/192	15.958 µs	74,624 B	74,624 B
PTQ ternary	70/192	101/192	161/192	16.000 µs	3,854 B	18,752 B + 8 B scales
QAT ternary	134/192	192/192	192/192	16.000 µs	3,854 B	18,752 B + 8 B scales

The 768/24/8 network has 18,656 parameters. Five trits per byte store matrix weights at about 1.6 bits each, with a theoretical information limit of about 1.58 bits. Go expands ternary matrices to int8; the packed file size and runtime tensor memory are reported separately. Feature parsing creates temporary objects, and total process memory exceeds tensor storage.

Actual execution and the intent gap

At budget one on the four selected normal source views (48 observed record fields), deterministic ordering matched 32/48 fields, the frozen ordinal model 36/48, and each fresh field model 48/48. At budget eight every profile reached 48/48. The four views are a smaller subset of the 192-view ranking audit; PTQ's complete result here coexists with its worse ranking across that full set.

Then we kept the same candidate expressions and changed intent/examples to copy the input state into the title, set wait, and prefix the reason. At budget one, the fresh models matched 12/24 runtime fields, deterministic order 20/24 and the frozen ordinal control 16/24. Full budget eight recovered 24/24 for all profiles. Named output completeness also fell to 8/16 at budget one: unchanged guard cases pass, while active transformations fail.

The current training repeats three normal intents. Its strong normal ordering therefore establishes a narrow assembly shortcut. The counter-intent experiment shows the next training requirement: paired, opposing goals over identical alternatives, varied Korean/English wording and literal/copy/concatenation roles. Keep these observed counter-intents as diagnostics and reserve fresh pairs for subsequent evaluation. No weights or thresholds were retuned after seeing them.

Resources and time

The actual optimization loops took 1.234 s total and 0.584 process CPU s. The whole Python training process took 3.18 s wall and 1.99 CPU s, with peak RSS 429.36 MiB. Sampled MPS tensor peak was 1.98 MiB and driver peak 50.72 MiB. Host GPU utilization and changes to whole-machine CPU utilization were not measured. Allocation is not a utilization percentage.

Across the 100 constructions, maximum compiler-child peak RSS was 84.20 MiB, and maximum generated-program peak RSS was 4.91 MiB. The two native runs per construction consumed 1.522 process CPU s in aggregate. Outer construction commands recorded 26.573 CPU s across all 100 invocations, including waited child process work on this machine. All process identities, times and partial values remain in the raw captures. Prediction medians above include feature projection after train/calibration warmup; fresh CLI prediction timings are retained separately. Whole-command timings include compilation/startup and fixed profile order, so they do not isolate a model speedup.

Run it

From the research repository root, use Go 1.27.1 and the clean pinned compiler:

unzip -q publication/record-field-learning-20261005/evidence.zip -d /tmp/gooo-field-pilot
GOWORK=off GOTOOLCHAIN=local go run ./cmd/record-field-curriculum \
  --audit-prepared /tmp/gooo-field-pilot/prepared \
  --models publication/record-field-learning-20261005/models \
  --output /tmp/gooo-field-audit-new
GOWORK=off GOTOOLCHAIN=local go run ./cmd/record-field-curriculum \
  --audit-prepared /tmp/gooo-field-pilot/prepared \
  --verify-native /tmp/gooo-field-pilot/native \
  --output /tmp/gooo-field-recount-new.jsonl
gooo body-compose --source publication/record-field-learning-20261005/demo.gooo.fixture \
  --cases publication/record-field-learning-20261005/demo-cases.json \
  --model publication/record-field-learning-20261005/models/fp32/model.json

Audit output must be a fresh directory. Numerical parity records are provided beside the model directory, and the full archive also retains that layout. The native study accepts --native-prepared, --compiler, --compiler-source, --frozen-model, --models and a fresh --output to reproduce all 100 observations.

evidence.zip retains complete Go preparation inputs, source/context exports, optimizer journals, all model variants, all prediction rows, and 100 raw native captures with new cases. SHA256SUMS pins every published file. The size report checks the 64 MiB uncompressed pilot budget. No external pretrained tensors initialize these models. Our earlier Laya experiments motivated the local decision approach; this release starts from the recorded own initializer.

The public model package is asketeddy/gooo-record-field-tiny-v1.