Research History 75 - teamswyg/laya-tools GitHub Wiki

laya-tools ยท User Guide / ์‚ฌ์šฉ์ž ์•ˆ๋‚ด

Go ๊ณ„์‚ฐ ๊ฐœ์„ 74 ํ•œ๊ตญ์–ด ยท Go calculation improvement74 English: ๊ฐ™์€ ํŠน์ง•ยท์ ์ˆ˜๋ฅผ ์œ ์ง€ํ•˜๊ณ  ๋กœ์ปฌ ๊ณต๊ฐœ ์˜ˆ์ œ์—์„œ ๊ณ„์‚ฐ ์‹œ๊ฐ„์ด์•ฝ24%ยท20% ์ค„์—ˆ์œผ๋ฉฐ ํ• ๋‹น๋Ÿ‰์€ ๊ฐ™์Šต๋‹ˆ๋‹ค / identical features and scores with approximately24%/20% lower local feature time and unchanged allocations. ๋‹จ๋… ์ €์žฅ reader74 ยท Standalone storage readers74: RSS18.23/18.41MiB๋กœ RAM ์ ˆ๊ฐ์€ ์ž…์ฆ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค / no RAM saving was demonstrated. ์ž…๋ ฅ ์„ค๋ช… ๋ณด๊ฐ• ยท Input description review keeps source truth and visible information separate / ์ฝ”๋“œ ์ •๋‹ต๊ณผ ๋ชจ๋ธ์ด ์ฝ๋Š” ์ •๋ณด๋ฅผ ๋‚˜๋ˆ  ๋‹ค์Œ ํŠœ๋‹์„ ์ค€๋น„ํ•ฉ๋‹ˆ๋‹ค.

์ƒˆ ์ž๋ฃŒ์˜ ์‹ค์ œ ๊ด€์ฐฐ73 ยท New-source observations73: three goals/24 finite inputs matched in one actual Go run; labels/fits0 / 3๋ชฉํ‘œ24์ž…๋ ฅ์€ ์ผ์น˜ํ–ˆ์ง€๋งŒ ์ƒˆ ํ•™์Šต ์ •๋‹ต์€0์ž…๋‹ˆ๋‹ค. ์ €์žฅ ํ˜•์‹73 ยท Storage73 preserves FP64 bits and compares minified/gzip controls; file size is separate from RAM / ์ˆ˜์น˜๋ฅผ ์œ ์ง€ํ•œ ์ €์žฅ ๊ธฐ์ค€์„ ์„ ๋น„๊ตํ•˜๋ฉฐ ํŒŒ์ผ๊ณผ RAM์„ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.

๋‘ ๋ฒˆ์งธ ์ž‘์€ ์ฃผ์žฅ ๋ชจ๋ธ ๊ฒฐ๊ณผ ยท Second tiny claim fit: ํ›„๋ณด ํ™•์ธ27โ†’31โ†’33์œผ๋กœ ๋‘ ํ•™์Šต ๋ฐฉ์‹์˜ ํšจ์šฉ์€ ์‹คํŒจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์ •๋ ฌ์„ ์œ ์ง€ํ•˜๊ณ  ๊ฐœ๋ฐœ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์„ ๋จผ์ € ๋Š˜๋ฆฝ๋‹ˆ๋‹ค. Go ์ž…๋ ฅ ์žฌ์‚ฌ์šฉ ยท Go input reuse์€ ์ •๋ ฌ ํ• ๋‹น81โ†’0์„ ๊ด€์ธกํ–ˆ๊ณ  ์ „์ฒด ํšŒ๊ท€ ๊ฒ€์‚ฌ๋ฅผ ํ†ต๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค.

Actual first fit71 / ์‹ค์ œ ์ฒซ ํ•™์Šต71: ํ•œ๊ตญ์–ด / English. One complete76 FP32 fit completed; validation checks27โ†’31, utility failed. Keep existing controls and test one fixed same-parent ranking term. File32,792B/whole-fit RSS26.98MiB/wall0.805s are separate observations, not LLM savings. / ์ „์ฒด76๊ฐœ FP32 ํ•™์Šต์€ ์™„๋ฃŒ๋์ง€๋งŒ ํšจ์šฉ์ด ์‹คํŒจํ•ด ๊ธฐ์กด ๊ทœ์น™์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜72๊ฐœ ์ค€๋น„ ์ž๋ฃŒ์˜ HF immutable archive๋Š” ๊ณต๊ฐœยท์›๋ณธ ๋Œ€์กฐ๋ฅผ ์™„๋ฃŒํ–ˆ์Šต๋‹ˆ๋‹ค. Earlier entries remain historical / ์•„๋ž˜ ๊ธฐ๋ก์€ ๋‹น์‹œ ์ƒํƒœ๋ฅผ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค.

Actual preparation69โ€“70 / ์‹ค์ œ ์ค€๋น„69โ€“70: ํ•œ๊ตญ์–ด / English. Original72 arrays and saved-array review completed; whole76 roles were reassigned once before fitting. Payload1.69MiB and whole preparation RSS53.95MiB are separate observations, not serving costs. / ์›๋ž˜72๊ฐœ ๋ฐฐ์—ด๊ณผ ๋ณ„๋„ ๊ฒ€์ฆ์„ ์™„๋ฃŒํ•˜๊ณ 76๊ฐœ ์ „์ฒด๋ฅผ ๋‹ค์‹œ ๋ถ„ํ• ํ–ˆ์Šต๋‹ˆ๋‹ค. payload์™€ ์ „์ฒด ์ค€๋น„ RSS๋ฅผ ๊ตฌ๋ถ„ํ•˜๋ฉฐ ๋ชจ๋ธ ์ถ”๋ก  ๋น„์šฉ์ด ์•„๋‹™๋‹ˆ๋‹ค.

Actual claim checks66โ€“69 / ์‹ค์ œ ์ฃผ์žฅ ๊ฒ€์‚ฌ66โ€“69: ํ•œ๊ตญ์–ด / English. Roles66 completed once; original72requests/216captions fit existing input bounds. Loss-eligible groups remain9/3/3. Four finite upstream requests across2source families were observed separately; first projection and fitting remain subsequent steps. / ์—ญํ• 66์„ ํ•œ ๋ฒˆ ์™„๋ฃŒํ–ˆ๊ณ  ์›๋ž˜72์š”์ฒญยท216์„ค๋ช… ๋ชจ๋‘ ๊ธฐ์กด ์ž…๋ ฅ ํ•œ๋„์— ๋งž์Šต๋‹ˆ๋‹ค. ํ•™์Šต ์ ๊ฒฉ ๊ทธ๋ฃน9ยท3ยท3์„ ์œ ์ง€ํ•˜๊ณ , ๋‘ ์›์ฒœ ๊ฐ€์กฑ์˜ ์œ ํ•œ upstream4์š”์ฒญ์€ ๋ณ„๋„๋กœ ๊ด€์ธกํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ํˆฌ์˜ยทํ•™์Šต์€ ๋‹ค์Œ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. Earlier entries below retain their historical state / ์•„๋ž˜ ๊ธฐ๋ก์€ ๋‹น์‹œ ์ƒํƒœ๋ฅผ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค.

Claim preparation64โ€“67 / ์ž‘์€ ํžŒํŠธ ํ•™์Šต ์ค€๋น„64โ€“67: ํ•œ๊ตญ์–ด / English. One actual metadata inventory preserved216 candidates:35 positive/95 negative eligible,23 known masked and63 unknown. Go projection and whole-group execution tools are prepared; original roles and fitting remain pending. / ์‹ค์ œ metadata ๊ฒ€์‚ฌ1ํšŒ์—์„œ ํ›„๋ณด216๊ฐœ๋ฅผ ๋ณด์กดํ–ˆ์Šต๋‹ˆ๋‹ค. ์ ๊ฒฉ ์–‘์„ฑ35ยท์Œ์„ฑ95ยทknown ์ œ์™ธ23ยทunknown63์ด๋ฉฐ Go ํˆฌ์˜ยท์ „์ฒด ๊ทธ๋ฃน ์‹คํ–‰ ๋„๊ตฌ๋ฅผ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์›๋ณธ ์—ญํ•  ๋ฐฐ์ •๊ณผ ํ•™์Šต์€ ์•„์ง ์ง„ํ–‰ ์ „์ž…๋‹ˆ๋‹ค.

Caption content61 / ์„ค๋ช… ๋‚ด์šฉ61: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์ƒ์„ธ ๊ฒ€ํ†  / Detailed findings ยท ๊ธฐ๊ณ„์  ๋Œ€์กฐ / Mechanical checks. All 738 original review entries across the existing 72 requests have readings: 524 consistent, 111 contradictions, 12 omissions and 91 uncertain. These are review entries, not new tasks or model results. The 21 unknown requests remain unknown; roles and fitting remain pending. / ๊ธฐ์กด 72์š”์ฒญ์˜ ์›๋ž˜ ๊ฒ€ํ†  ํ•ญ๋ชฉ738๊ฐœ๋ฅผ ๋ชจ๋‘ ์ฝ์—ˆ์Šต๋‹ˆ๋‹ค. ์ผ์น˜524ยท๋ชจ์ˆœ111ยท๋ˆ„๋ฝ12ยท๋ถˆํ™•์‹ค91์ด๋ฉฐ ์ƒˆ ์ž‘์—… ์ˆ˜๋‚˜ ๋ชจ๋ธ ์„ฑ๊ณผ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. ๋ณด๋ฅ˜ ์š”์ฒญ21๊ฐœ๋Š” ์œ ์ง€ํ•˜๊ณ  ์—ญํ•  ๋ฐฐ์ •ยทํ•™์Šต์€ ์•„์ง ๋ฏธ์™„์„ฑ์ž…๋‹ˆ๋‹ค.

Go preparation62โ€“63 / Go ์ค€๋น„62โ€“63: ์—ญํ•  ๋ถ„ํ•  ๋ชจ๋“ˆ / Whole-group role module ยท ๊ณต๊ฐœ ์›๋ฌธ ๋ฌธ๊ตฌ / Public upstream wording. The bounded Go module preserves connected groups and rejects invalid coverage IDs before sorting. The 25 complete upstream quotes retain MIT notices. Synthetic checks and source matching do not establish training readiness, independent data authorship or measured memory savings. / Go ๋ชจ๋“ˆ์€ ์—ฐ๊ฒฐ ๊ทธ๋ฃน์„ ๋ณด์กดํ•˜๊ณ  ์ž˜๋ชป๋œ coverage ID๋ฅผ ์ •๋ ฌ ์ „์— ๊ฑฐ์ ˆํ•ฉ๋‹ˆ๋‹ค. ์›๋ฌธ25๊ฐœ๋Š” MIT ๊ณ ์ง€๋ฅผ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๊ฒ€์‚ฌ์™€ ์›๋ฌธ ๋Œ€์กฐ๋งŒ์œผ๋กœ ํ•™์Šต ์ค€๋น„ยท๋…๋ฆฝ ์ž‘์„ฑ ์›์ฒœยท์‹ค์ธก ๋ฉ”๋ชจ๋ฆฌ ์ ˆ๊ฐ์„ ์ฆ๋ช…ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

Caption references59 / ์„ค๋ช… ์ฐธ์กฐ59: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage. One frozen Go run connects 288 original text locations and 18 literal expressions; content review, roles and fitting remain pending. / Go ๊ณต์‹1ํšŒ๋กœ ์›๋ฌธ288๊ณณยท๋ฆฌํ„ฐ๋Ÿด ํ‘œํ˜„์‹18๊ณณ์„ ์—ฐ๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์„ค๋ช…์˜ ์˜๋ฏธ ๊ฒ€ํ† ยท์—ญํ• ยทํ•™์Šต์€ ์•„์ง ๋ฏธ์™„์„ฑ์ž…๋‹ˆ๋‹ค. ๊ณต๊ฐœ ์›์ฒœ ์ „์ด ์ค€๋น„ / Source transfer preparation.

Stored-truth utility58 / ์ €์žฅ ์ •๋‹ต ํšจ์šฉ58: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage. The best fixed control needs 91 checks versus an answer-knowing oracle's 73: 19.8% headroom on 72 existing development requests, not achieved model or Codex savings. Retain 21 unknowns; training is not ready. / ๊ธฐ์กด ๊ฐœ๋ฐœ ์š”์ฒญ 72๊ฐœ์—์„œ ๊ฐ€์žฅ ์ข‹์€ ๊ณ ์ • ๊ธฐ์ค€์€ ํ™•์ธ 91ํšŒ, ์ •๋‹ต์„ ์•„๋Š” ์ˆœ์„œ๋Š” 73ํšŒ์˜€์Šต๋‹ˆ๋‹ค. ์•ฝ 19.8%๋Š” ๊ฐœ์„  ์—ฌ์ง€์ด๋ฉฐ ๋ชจ๋ธ ์„ฑ๊ณผ๋‚˜ Codex ์ ˆ๊ฐ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๋ณด๋ฅ˜ 21๊ฐœ๋ฅผ ์œ ์ง€ํ•˜๊ณ  ํ•™์Šต ์ค€๋น„๋Š” ๋ฏธ์™„์„ฑ์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ ์ค€๋น„ / Next preparation.

CI Release Project license Status

Public behavior57 / ๊ณต๊ฐœ ๋™์ž‘57: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage.62 actual Go API observations:59 matches/3 disagreements; preserve two reversed large-number comparison discrepancies and one stronger glob policy difference. Two source families/six finite contracts; no model, final or savings claim. / ์›๋ณธ Go API62๊ด€์ธก์˜59์ผ์น˜ยท3์ฐจ์ด๋ฅผ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค. ํฐ ์ˆซ์ž ๋น„๊ต์˜ ๋ฐ˜์ „์Œ๊ณผ ๊ฐ•ํ•œ glob ๊ฒ€์‚ฌ ์ •์ฑ…์˜ ์ฐจ์ด๋ฅผ ๊ตฌ๋ถ„ํ•˜๋ฉฐ ๋‘ ์›์ฒœ ๊ฐ€์กฑยท์—ฌ์„ฏ ์œ ํ•œ ๊ณ„์•ฝ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ ์„ฑ๋Šฅยท์ตœ์ข… ํ‰๊ฐ€ยท์ ˆ๊ฐ ์„ฑ๊ณผ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. Laya ํšŒ๊ณ KO / EN.

Finite claim truth56e / ์œ ํ•œ ์ฃผ์žฅ ์ •๋‹ต56e: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage. After independently reviewing the stated input scope, one frozen audit connects36 actual observations to12 derived rows:11 finite-answerable/1 no-answer. Preserve general pending/unknown captions; independent parents/fits/models/utility claims0. / ์š”์ฒญ์— ๋ช…์‹œํ•œ ์ž…๋ ฅ ๋ฒ”์œ„๋ฅผ ๋…๋ฆฝ ๊ฒ€ํ† ํ•˜๊ณ  ๋ด‰์ธ ๋’ค36์‹ค์ œ ๊ด€์ธก์„12ํŒŒ์ƒ ํ–‰์— ์—ฐ๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค.11ํ–‰์€ ์œ ํ•œ ๋ฒ”์œ„์—์„œ ๋‹ตํ•  ํ›„๋ณด๊ฐ€ ์žˆ๊ณ 1ํ–‰์€ ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜ ๋ฌธ๊ตฌ์˜ pending/unknown์„ ๋ณด์กดํ•˜๋ฉฐ ๋…๋ฆฝ ๋ถ€๋ชจ ์ฆ๊ฐ€ยทํ•™์Šตยท๋ชจ๋ธ ์„ฑ๊ณผ0์ž…๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์›์ฒœ์˜ ๋‹ค์Œ ๊ณ„์•ฝ / Next source contracts ยท CI/merge completion.

Resident costs56d / ์ƒ์ฃผ ์ฒ˜๋ฆฌ ๋น„์šฉ56d: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage. One frozen actual Go replay completed24 rows and25,080 repeated requests. Child peak RSS9.22โ€“10.52MiB and per-row wire p95 0.022โ€“0.050ms; first response577.13ms remains unexplained. Collector33.11MiB peak is separate. Repeated48 inputs are not2400 final requests; all narrow-rule inputs used BM25 fallback. No model inference, new fit, cache or LLM savings claim. / ์‹ค์ œ Go ์‹คํ–‰24ํ–‰ยท๋ฐ˜๋ณต25,080ํšŒ๋ฅผ ์™„๋ฃŒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒ˜๋ฆฌ๊ธฐ peak RSS9.2210.52MiBยท์™•๋ณตp95์•ฝ0.0220.050ms์ด๋ฉฐ ์ฒซ ์‘๋‹ต577.13ms ์›์ธ์€ ๋ฏธํ™•์ธ์ž…๋‹ˆ๋‹ค. ์ธก์ •๊ธฐ33.11MiB peak๋Š” ๋ณ„๋„์ž…๋‹ˆ๋‹ค. 48์ž…๋ ฅ ๋ฐ˜๋ณต์€2400์ตœ์ข… ์š”์ฒญ์ด ์•„๋‹ˆ๋ฉฐ ์ข์€ ๊ทœ์น™๋„ ์ „๋ถ€ BM25 fallback์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์ถ”๋ก ยท์ƒˆ ํ•™์Šตยท์บ์‹œยทLLM ์ ˆ๊ฐ ์ฆ๋ช…๊ณผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. CI completion / CI ์™„๋ฃŒ ๊ธฐ๋ก.

Choose your language / ์–ธ์–ด๋ฅผ ์„ ํƒํ•˜์„ธ์š”: English โ€” Start here ยท ํ•œ๊ตญ์–ด โ€” ์ฒ˜์Œ ์‹œ์ž‘ํ•˜๊ธฐ

riidolaya helps you find code, compare model-selection policies, and preview which repository owns a task. Start without a model; add local Laya inference only when you want to measure its benefit. Codex integration is optional.

riidolaya๋Š” ์ฝ”๋“œ ์ฐพ๊ธฐ, ๋ชจ๋ธ ์„ ํƒ ์ •์ฑ… ๋น„๊ต, ์ž‘์—…์„ ๋‹ด๋‹นํ•  ์ €์žฅ์†Œ preview๋ฅผ ๋•์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์—†์ด ๋จผ์ € ์‹œ์ž‘ํ•˜๊ณ , ํšจ๊ณผ๋ฅผ ์ธก์ •ํ•  ๋•Œ ๋กœ์ปฌ Laya ์ถ”๋ก ์„ ์ถ”๊ฐ€ํ•˜์„ธ์š”. Codex ์—ฐ๋™์€ ์„ ํƒ ์‚ฌํ•ญ์ž…๋‹ˆ๋‹ค.

Experimental: actual Codex cost savings have not been demonstrated. Repository results are previews, not automatic work assignments.

์‹คํ—˜ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์‹ค์ œ Codex ๋น„์šฉ ์ ˆ๊ฐ์€ ์•„์ง ์ž…์ฆํ•˜์ง€ ์•Š์•˜์œผ๋ฉฐ, ์ €์žฅ์†Œ ์„ ํƒ ๊ฒฐ๊ณผ๋„ ์ž๋™ ์ž‘์—… ๋ฐฐ์ •์ด ์•„๋‹Œ preview์ž…๋‹ˆ๋‹ค.

Latest owned task results / ์ตœ๊ทผ ์‹ค์ œ ์ž‘์—… ๊ฒฐ๊ณผ: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ55 / English results55. The existing humanize/UUID requests ran once per Sol6/Luna low request profile: both ordinal candidates passed; only the Sol6-requested UUID candidate passed. All four bind zero exit, whole usage and durable budget/cleanup receipts. Actual coverage is20 records,7 distinct requests,5 attempted families and3 repositories. Laya truncated and abstained on both full inputs; cold CPU RSS was about1.54/1.58GB. No training, final scoring, general ranking or savings claim. / ๊ธฐ์กด humanizeยทUUID ์š”์ฒญ์„ Sol6/Luna low์— ๊ฐ๊ฐ ํ•œ ๋ฒˆ ์‹คํ–‰ํ•ด ์„œ์ˆ˜๋Š” ๋‘˜ ๋‹ค ํ†ต๊ณผํ–ˆ๊ณ  UUID๋Š” Sol6 ์š”์ฒญ๋งŒ ํ†ต๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค. ๋„ค ๊ธฐ๋ก์˜ ์ข…๋ฃŒยท์ „์ฒด ์‚ฌ์šฉ๋Ÿ‰ยท์˜ˆ์‚ฐ ์˜์ˆ˜์ฆยท์ •๋ฆฌ๋ฅผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ๋ˆ„์ 20๊ธฐ๋กยท7๊ณ ์œ  ์š”์ฒญยท๋‹ค์„ฏ ์‹œ๋„ ๊ฐ€์กฑยท์„ธ ์ €์žฅ์†Œ์ž…๋‹ˆ๋‹ค. Laya๋Š” ๋‘ ์ „์ฒด ์ž…๋ ฅ ๋ชจ๋‘ ์ž˜๋ ค ๋ณด๋ฅ˜ํ–ˆ๊ณ  cold CPU RSS ์•ฝ1.54/1.58GB์˜€์Šต๋‹ˆ๋‹ค. ํ•™์Šตยทfinalยท์ผ๋ฐ˜ ๋ชจ๋ธ ์ˆœ์œ„ยท์ ˆ๊ฐ ์ฆ๋ช…์€ ์—†์Šต๋‹ˆ๋‹ค. KO source ยท EN source.

Prior training / ์ด์ „ ํ•™์Šต: ํ•™์Šต 4,456๊ฐœยท๊ฒ€์ฆ 2,599๊ฐœ ๊ฒฐ๊ณผ / 4,456 training and 2,599 validation examples. All ten tiny-head fits completed; usefulness gates failed. / ์ž‘์€ ๋ชจ๋ธ 10๊ฐœ ํ•™์Šต์€ ์™„๋ฃŒํ–ˆ์ง€๋งŒ ์œ ์šฉ์„ฑ ๊ธฐ์ค€์€ ์‹คํŒจํ–ˆ์Šต๋‹ˆ๋‹ค.

Go short-claim preview56a / ์ž‘์€ ์ฃผ์žฅ Go preview56a: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage. Model-free riido-shortclaim --stream preserves candidates and proposes verification order. Eight-candidate full-request warm p95 was0.063โ€“0.079ms, repeated child peak RSS10.25โ€“10.56MiB. The48 authored cases form11 groups below15; fits0. The18.64% oracle gap is an upper bound, not achieved model performance or LLM savings. / ๋ชจ๋ธ ์—†๋Š” Go ๋„๊ตฌ์™€ ๋ฐ˜๋ณต ์ŠคํŠธ๋ฆผ์ด ํ›„๋ณด๋ฅผ ๋ณด์กดํ•˜๋ฉฐ ๊ฒ€์‚ฌ ์ˆœ์„œ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. 8ํ›„๋ณด ์˜ˆ์ œ warm p95์•ฝ0.0630.079msยท๋ฐ˜๋ณตchild RSS10.2510.56MiB. 48์‚ฌ๋ก€๋Š”11๊ทธ๋ฃน์œผ๋กœ ํ•˜ํ•œ15๋ฏธ๋‹ฌยทํ•™์Šต0์ž…๋‹ˆ๋‹ค. Oracle18.64%๋Š” ์ƒํ•œ์ด๋ฉฐ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด๋‚˜ LLM ์ ˆ๊ฐ์ด ์•„๋‹™๋‹ˆ๋‹ค. Historical preparation / ์—ญ์‚ฌ์  ์ค€๋น„ ยท Next acquisition / ๋‹ค์Œ ํ™•๋ณด.

Golden sets / ๊ณจ๋“ ์…‹: 24๊ฐœ์™€ 2,400๊ฐœ์˜ ์˜๋ฏธ ๋ฐ ๋„๋ฉ”์ธ๋ณ„ ์ •๋‹ต / Scale, held-out evaluation and domain-specific evidence. At least 2,400 distinct final requests per evaluated domain remains the target, with related cases grouped. Search evidence and development attempts do not validate final model routing. / ์˜์—ญ๋ณ„ ๊ตฌ๋ถ„๋˜๋Š” ์ตœ์ข… ์š”์ฒญ ์ตœ์†Œ 2,400๊ฑด์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ ๊ด€๋ จ ์‚ฌ๋ก€๋Š” ๊ทธ๋ฃน์œผ๋กœ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์™€ ๊ฐœ๋ฐœ ์‹œ๋„๋กœ ์ตœ์ข… ๋ชจ๋ธ ๋ผ์šฐํŒ…์„ ๊ฒ€์ฆํ–ˆ๋‹ค๊ณ  ๋ณผ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.

Acquisition and budgets / ํ™•๋ณด์™€ ์‹คํ–‰ ์˜ˆ์‚ฐ: ํ•œ๊ตญ์–ด ๊ณ„ํš / English plan. Follow staged source collection, independently checked task contracts, grouped splits, and bounded comparisons; repeated model executions do not increase unique-request counts. / ์›์ฒœ ํ™•๋ณดยท๋…๋ฆฝ ์ž‘์—… ๊ณ„์•ฝยท๊ทธ๋ฃน ๋ถ„ํ• ยท์ œํ•œ๋œ ๋น„๊ต ์‹คํ–‰์„ ๋‹จ๊ณ„์ ์œผ๋กœ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ™์€ ์š”์ฒญ์„ ์—ฌ๋Ÿฌ ๋ชจ๋ธ๋กœ ์‹คํ–‰ํ•ด๋„ ๊ณ ์œ  ์š”์ฒญ ์ˆ˜๋Š” ๋Š˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ฒ€ํ†  ์›๋ณธ ยท Reviewed source.

Historical source survey53 / ์—ญ์‚ฌ์  ์›์ฒœ ์กฐ์‚ฌ53: ํ•œ๊ตญ์–ด / English. This preserved snapshot surveyed120 candidates from8 repositories with0 execution-eligible contracts then;54 prepared2 contracts and55 first executed those2 requests. / ๋‹น์‹œ8์ €์žฅ์†Œยท120ํ›„๋ณดยท์‹คํ–‰ ์ž๊ฒฉ0์„ ์กฐ์‚ฌํ•œ snapshot์„ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„54์—์„œ2๊ณ„์•ฝ์„ ์ค€๋น„ํ•˜๊ณ 55์—์„œ๊ทธ๋‘์š”์ฒญ์„์ฒ˜์Œ์‹คํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. KO source ยท EN source.

Historical contract preparation54 / ์—ญ์‚ฌ์  ๊ณ„์•ฝ ์ค€๋น„54: ํ•œ๊ตญ์–ด / English. At54,2 of120 candidates received independent contracts, preserving humanize MIT and UUID BSD source/license bytes and module identities; external model attempts were0 then. Actual55 results are above; training/final eligibility remains0. / 54๋‹น์‹œ120ํ›„๋ณด ์ค‘2๊ฐœ์˜ ๋…๋ฆฝ ๊ณ„์•ฝ์„ ์ค€๋น„ํ•˜๊ณ  humanize MITยทUUID BSD ์›๋ณธ/๋ผ์ด์„ ์Šคยท๋ชจ๋“ˆ ์ด๋ฆ„์„ ๋ณด์กดํ–ˆ์œผ๋ฉฐ ์™ธ๋ถ€ ์‹คํ–‰์€ ๋‹น์‹œ0์ด์—ˆ์Šต๋‹ˆ๋‹ค. ํ›„์†55์‹คํ–‰ ๊ฒฐ๊ณผ๋Š” ์œ„์— ์žˆ๊ณ  ํ•™์Šตยทfinal ์ ๊ฒฉ์€ ๊ณ„์†0์ž…๋‹ˆ๋‹ค.

Fair comparison preparation 55 / ๊ณต์ • ๋น„๊ต ์ค€๋น„ 55: ํ•œ๊ตญ์–ด ์•ˆ๋‚ด ยท English guide.

The v2 IDs clarify the same two external requests without adding distinct tasks. Full input, original Go language and native predictions were frozen before4 actual attempts. Two child plans shared4 ordered no-refund reservations in one durable ledger, concurrency1, with4 terminal receipts; this is not a host/account/provider-wide cap. Development observations do not replace the separate target of at least2,400 distinct protected final requests per domain.

v2 ID๋Š” ๊ธฐ์กด ๋‘ ์™ธ๋ถ€ ์š”์ฒญ์˜ ํ‰๊ฐ€ ๋ฒ„์ „์ด๋ฉฐ ๊ณ ์œ  ์ž‘์—… ์ˆ˜๋ฅผ ๋Š˜๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ „์ฒด ์ž…๋ ฅยท์›๋ณธ Go ์–ธ์–ดยท์‚ฌ์ „ native ์˜ˆ์ธก์„ ๊ณ ์ •ํ•œ ๋’ค ๋„ค ์‹œ๋„๋ฅผ ์‹คํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œ durable ledger์—์„œ ๋„ค ์˜ˆ์•ฝ์„ ํ™˜๋ถˆ ์—†์ด ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์”ฉ ์‚ฌ์šฉํ–ˆ๊ณ  ์ข…๋ฃŒ ์˜์ˆ˜์ฆ๋„ ๋„ค ๊ฐœ์ž…๋‹ˆ๋‹ค. ํ˜ธ์ŠคํŠธยท๊ณ„์ •ยท๊ณต๊ธ‰์ž ์ „์ฒด์˜ ํ•œ๋„๋Š” ์•„๋‹™๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ ๊ด€์ธก์€ ๋„๋ฉ”์ธ๋ณ„ ์„œ๋กœ ๋‹ค๋ฅธ ๋ณดํ˜ธ๋œ ์ตœ์ข… ์š”์ฒญ ์ตœ์†Œ2,400๊ฐœ ๋ชฉํ‘œ๋ฅผ ๋Œ€์‹ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

Usage and task completion / ์‚ฌ์šฉ๋Ÿ‰๊ณผ ์ž‘์—… ์™„๋ฃŒ: ํ•œ๊ตญ์–ด ์‚ฌ์šฉ๋ฒ• / English guide. Existing telemetry and independent acceptance are separate; missing evidence stays unknown. Authored examples are tool checks, not model outcomes. / ๊ธฐ์กด ๊ธฐ๋ก์˜ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ๋…๋ฆฝ ์ž‘์—… ๊ฒ€์ฆ์„ ๋”ฐ๋กœ ํ™•์ธํ•˜๋ฉฐ, ๋ˆ„๋ฝ์€ ๋ฏธํ™•์ •์œผ๋กœ ๋‚จ๊น๋‹ˆ๋‹ค. ์ง์ ‘ ์ž‘์„ฑํ•œ ์˜ˆ์ œ๋Š” ๋„๊ตฌ ๊ฒ€์‚ฌ์ด๋ฉฐ ๋ชจ๋ธ ์ž‘์—… ๊ฒฐ๊ณผ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค.

Explicit task execution / ๋ช…์‹œ์  ์ž‘์—… ์‹คํ–‰: ํ•œ๊ตญ์–ด / English. An opt-in Go executor binds a public task's CLI invocation, usage, candidate and independent checks; failures stay visible. The development pilot is separate from final routing evaluation. / ์„ ํƒ์  Go ์‹คํ–‰๊ธฐ๋กœ ๊ณต๊ฐœ ์ž‘์—…์˜ CLI ์‹คํ–‰ยท์‚ฌ์šฉ๋Ÿ‰ยทํ›„๋ณดยท๋…๋ฆฝ ๊ฒ€์‚ฌ๋ฅผ ์—ฐ๊ฒฐํ•˜๊ณ  ์‹คํŒจ๋„ ๊ธฐ๋กํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ ํŒŒ์ผ๋Ÿฟ์€ ์ตœ์ข… ๋ผ์šฐํŒ… ํ‰๊ฐ€์™€ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.

Preserved prelaunch stop / ์‹คํ–‰ ์ „ ์ค‘๋‹จ ๊ธฐ๋ก: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ 50 / English results 50. The earlier packaging defect stopped before any main coding-model process. This historical result is retained separately. / ์ด์ „ ํŒจํ‚ค์ง• ๋ฌธ์ œ๋Š” ์ฝ”๋”ฉ ๋ชจ๋ธ ๋ณธ ์‹คํ–‰ ์ „์— ๋ฉˆ์ท„์œผ๋ฉฐ ๊ทธ ๊ฒฐ๊ณผ๋ฅผ ๋ณ„๋„๋กœ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค.

Search helper screen / ๊ฒ€์ƒ‰ ๋ฐฉ์‹ ๋น„๊ต: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results. Three fixed helpers failed the 5% necessary headroom gate on 4,456 training requests; fitting stopped. Full runs used 106.7โ€“107.3MiB peak RSS. / ํ•™์Šต์šฉ 4,456๊ฐœ์—์„œ ์„ธ ํ›„๋ณด ๋ชจ๋‘ 5% ๊ฐœ์„  ์—ฌ์ง€ ๊ธฐ์ค€์— ๋ชป ๋ฏธ์ณ ์ถ”๊ฐ€ ํ•™์Šต์„ ์ค‘๋‹จํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ์‹คํ–‰์˜ ์ตœ๋Œ€ RSS๋Š” 106.7โ€“107.3MiB์˜€์Šต๋‹ˆ๋‹ค.

Typed behavioral truth56b / ๋™์ž‘ ์ •๋‹ต ๊ฐ์‚ฌ56b: ํ•œ๊ตญ์–ด ๊ฒฐ๊ณผ / English results ยท ์‚ฌ์šฉ๋ฒ• / Usage ยท ์‚ฌ์ „ ๊ณ„ํš / Plan. The Go maintainer auditor checks error identity, state, ownership, lifetime and graph behavior. New24 plus legacy48 authored cases form17 total and16 labeled groups, passing both minima of15. Clear labels cover34 answerable and17 no-answer requests;21 remain unknown. The524 source assertions are not independent requests. Roles, fits, model calls, weights, final scoring and new performance measurements remain0; this is one synthetic English authoring process. / ์˜ค๋ฅ˜ยท์ƒํƒœยท์†Œ์œ ๊ถŒยท์ˆ˜๋ช…ยท๊ทธ๋ž˜ํ”„๋ฅผ ๊ฒ€์‚ฌํ•˜๋Š” ์œ ์ง€๋ณด์ˆ˜ Go ๋„๊ตฌ์ž…๋‹ˆ๋‹ค. ์ƒˆ24+๊ธฐ์กด48์€ ์ „์ฒด17ยท๋ผ๋ฒจ16๊ทธ๋ฃน์œผ๋กœ ๋‘15 ํ•˜ํ•œ์„ ํ†ต๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ต ์žˆ์Œ34ยท๋‹ต ์—†์Œ17ยทunknown21์ด๋ฉฐ, ์›์ฒœ ๊ฒ€์‚ฌ524๊ฐœ๋ฅผ ๋…๋ฆฝ ์š”์ฒญ์œผ๋กœ ์„ธ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋™์ผ ํ•ฉ์„ฑ ์˜์–ด ์ž‘์„ฑ ํ๋ฆ„์˜ ์ž๋ฃŒ๋กœ ์—ญํ• ยทํ•™์Šตยท๋ชจ๋ธ ํ˜ธ์ถœยท๊ฐ€์ค‘์น˜ยท์ตœ์ข… ํ‰๊ฐ€ยท์ƒˆ ์„ฑ๋Šฅ ์ธก์ •์€0์ž…๋‹ˆ๋‹ค.

Scoped proposal preparation56c / ์ž‘์€ ์†์„ฑ ์ œ์•ˆ ์ค€๋น„56c: ํ•œ๊ตญ์–ด ๊ธฐ๋ก / English record ยท ์‚ฌ์šฉ๋ฒ• / Usage ยท ์‚ฌ์ „ ์„ค๊ณ„ / Design. Go riido-scopeprep generated12 proposal rows from4 existing parents across3 properties, preserving candidate order and pins. All36 captions remain pending; these add no independent requests or approved truth. Official candidate evaluations, groups, fits, models, rankings, performance and protected-final access remain0. / ๊ธฐ์กด ๋ถ€๋ชจ4๊ฐœ์—์„œ ์†์„ฑ3๊ฐœยท์ œ์•ˆ12ํ–‰์„ ์ค€๋น„ํ•˜๊ณ  ํ›„๋ณด ์ˆœ์„œยทํ•€์„ ๋ณด์กดํ–ˆ์Šต๋‹ˆ๋‹ค. ์„ค๋ช…36๊ฐœ๋Š” pending์ด๋ฉฐ ๋…๋ฆฝ ์š”์ฒญ์ด๋‚˜ ์Šน์ธ๋œ ์ •๋‹ต์„ ๋Š˜๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ณต์‹ ํ‰๊ฐ€ยท๊ทธ๋ฃนยทํ•™์Šตยท๋ชจ๋ธยท์ˆœ์œ„ยท์„ฑ๋Šฅยท๋ณดํ˜ธ ์ตœ์ข… ์ ‘๊ทผ์€0์ž…๋‹ˆ๋‹ค. ๋ณ„๋„ ์ƒ์ฃผ ์ž์›56d ์„ค๊ณ„ / Separate resident resource design56d was unexecuted at56c; actual56d is recorded above / 56c ๋‹น์‹œ ๋ฏธ์‹คํ–‰ ์„ค๊ณ„์ด๋ฉฐ ์‹ค์ œ56d๋Š” ์œ„์— ๋ณ„๋„ ๊ธฐ๋กํ–ˆ์Šต๋‹ˆ๋‹ค.

What do you want to do? / ๋ฌด์—‡์„ ํ•˜๊ณ  ์‹ถ๋‚˜์š”?

Goal / ๋ชฉ์  English ํ•œ๊ตญ์–ด
Install and get your first result / ์„ค์น˜ยท์ฒซ ์‹คํ–‰ Getting started ์ฒ˜์Œ ์‹œ์ž‘ํ•˜๊ธฐ
Find code, plan models, preview repositories / ๊ธฐ๋Šฅ๋ณ„ ์‚ฌ์šฉ Workflows ๊ธฐ๋Šฅ๋ณ„ ์‚ฌ์šฉ๋ฒ•
Connect an agent or Go application / ์—์ด์ „ํŠธยทGo ์—ฐ๋™ Agents and Go ์—์ด์ „ํŠธยทGo ์—ฐ๋™
Understand memory, speed, and errors / ์„ฑ๋Šฅยท๋ฌธ์ œ ํ•ด๊ฒฐ Performance and troubleshooting ์„ฑ๋Šฅ๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ
Reuse or redistribute / ํ”„๋กœ์ ํŠธ ๋ชฉ์ ยท๋ผ์ด์„ ์Šค Project and licensing ํ”„๋กœ์ ํŠธ์™€ ๋ผ์ด์„ ์Šค

What runs locally? / ์–ด๋””์„œ ์‹คํ–‰๋˜๋‚˜์š”?

Operation / ๋ช…๋ น Behavior / ๋™์ž‘
search --lexical, plan without a task, repo-preview without --laya Local Go logic; no model download required / ๋ชจ๋ธ ์—†์ด ๋กœ์ปฌ Go๋กœ ์ฒ˜๋ฆฌ
setup Downloads public model/runtime assets / ๊ณต๊ฐœ ๋ชจ๋ธยท๋Ÿฐํƒ€์ž„ ๋‹ค์šด๋กœ๋“œ
Laya-assisted search, routing, preview Local inference; roughly 1.4 GiB measured process RSS / ๋กœ์ปฌ ์ถ”๋ก , ์ธก์ •๋œ ์ „์ฒด ๋ฉ”๋ชจ๋ฆฌ ์•ฝ 1.4GiB
codex --dry-run Shows a proposed command; does not launch Codex / ์‹คํ–‰ ๊ณ„ํš๋งŒ ํ‘œ์‹œ
codex Starts your installed Codex; its normal provider and usage apply / ๊ธฐ์กด Codex ์‹คํ–‰, ๊ธฐ์กด ๊ณต๊ธ‰์žยท์š”๊ธˆ ์ •์ฑ… ์ ์šฉ

CI/release badges describe build status and versions, not proven savings or accuracy. The license badge describes project code; upstream notices still apply.

CIยท๋ฆด๋ฆฌ์Šค ๋ฐฐ์ง€๋Š” ๊ฒ€์‚ฌ ์ƒํƒœ์™€ ๋ฒ„์ „์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ ์ •ํ™•๋„ยท์ ˆ๊ฐ ํšจ๊ณผ์˜ ์ธ์ฆ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๋ผ์ด์„ ์Šค ๋ฐฐ์ง€๋Š” ์ž์ฒด ์ฝ”๋“œ ๊ธฐ์ค€์ด๋ฉฐ ์™ธ๋ถ€ ๊ตฌ์„ฑ ์š”์†Œ ๊ณ ์ง€๋„ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.

Repository ยท Downloads ยท Detailed bilingual documents / ์ƒ์„ธ ๋ฌธ์„œ ยท Report an issue / ๋ฌธ์ œ ์ œ๋ณด

Page cost experiment 12 / ํŽ˜์ด์ง€ ๋น„์šฉ ์‹œํ—˜

ํ•œ๊ตญ์–ด ยท English. 2,948 distinct queries: fewer emitted candidates does not establish lower CPU cost. PR #35 tracks CI and merge status.

Reusable search session / ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ์žฌ์‚ฌ์šฉ ์„ธ์…˜

ํ•œ๊ตญ์–ด ยท English. Use --session --limit 20 to compute once and read subsequent pages using a cursor. See the associated PR for CI and merge status.

Page-saving claim training 14 / ํŽ˜์ด์ง€ ์ ˆ๊ฐ ์ฃผ์žฅ ํ•™์Šต

ํ•œ๊ตญ์–ด ยท English. 24 model exports over 2,948 queries all match always-helper; the fixed gate fails. PR #37 tracks CI and publication readiness.

Helper-call budget experiment 15 / ๋ณด์กฐ ํ˜ธ์ถœ ์˜ˆ์‚ฐ ์‹œํ—˜

ํ•œ๊ตญ์–ด ยท English. Validation-only calibration: the simple score-gap rule beats the learned heads, but all fixed gates fail. Reuses the immutable experiment-14 models.

CoNaLa distinct-question audit 16 / ์‹ค์ œ ์งˆ๋ฌธ ์ˆ˜ ๊ฐ์‚ฌ

ํ•œ๊ตญ์–ด ยท English. 2,879 curated rows contain 2,074 question IDs / 2,089 normalized original intents, insufficient alone for 2,400 distinct original questions. No model scoring or training.