Frozen inputs and completed reference review do not establish training-data acceptance. Full scope was retained; model creation, new Fits, model serialization and actual application writes remain 0. Human accuracy and operational usefulness remain unproven.

1. Retain the current failure. The pinned Go QA tools checked the whole scope once. The comparer recorded exit code 2 and BLOCKED_BEFORE_ALL_FITS. We remain at data validation.
2. Diagnose the causes. Review source and creation-trace evidence, and definitions that separate absent assertions from unclear context in bilingual judgments. Do not decide which rater is wrong or relabel first outputs.
3. Prepare the next separate study. Source/provenance, bilingual context definitions and relevant primary research are under review. No new protocol is selected. There is no budget reset, further wording edit, easier subset or lowered gate.

We are preparing a small Go model that offers ❓ response-request · 🛠️ progress-report · ✅ completion-report hints for Riido development comments. Hints describe a comment's meaning; they do not verify factual task completion. The default remains OFF, with no actual application writes.

Current full-QA record · HF preparation snapshot · before whole QA · 한국어

The HF link is an immutable preparation snapshot from before whole QA. At that draft stage, all 1,200 first references in each locale were locked, while separate bilingual review covered 780 / 1,200 pairs with 420 remaining. Bilingual review is now complete, and subsequent whole QA recorded FAIL; the table above shows the current status. This is not a raw-data or new-model release.

• Go checked references and correction records for 120 scenarios / 360 bilingual concept families. Actual training messages: 0.
• The 71 pending hypotheses, including scaffold links, form four candidate review components; the largest contains 117 scenarios. These are not confirmed lineages or independent sample counts.
• A prospective plan reviews 300 compact and 100 general candidates before wording and allocation. Actual assignments: 0; all existing criteria remain unchanged.

80-concept design and audit · Preservation-verified HF followup

• The third 40-group tranche resolves six scope/attribution/link findings with one retained 26-field correction; independent semantic readback passed.
• Go prior-only reference and array-preservation bugs were fixed before actual input access. A closed format-mismatch attempt is preserved; the same pinned inputs then passed. No model/training qualification is claimed.

120-concept join and repairs · Preservation-verified HF 120 followup

All 120 inputs/60 bilingual families/20 groups, separate-locale first votes and frame verdicts are sealed. The Go join validates pins, structure, ordering, all 360 evidence records and zero normalized duplicates. The fixed protocol fails at 56/60 accepted frames: 1 assertion and 3 naturalness holds. All inputs, holds and disagreements remain; no result-driven rewrites occurred.

Designated≤16/17–32-codepoint coverage is KO 18/18, 17/18 and EN 15/18, 18/18. Both general groups meet the same accepted pair with both locales ≥65 rule. Eleven style-flagged and four frame-held families do not overlap, affecting 15/60 families. Locale label agreement and AI frame QA are neither model accuracy nor human gold. Unknown counts are KO R1/A2/C2 and EN R2/A0/C0.

The 120 rows are excluded from training, model evaluation, gold and final 2400 counts. The full target remains 400 groups / 1,200 bilingual families / 2,400 rows, with DEV 20-family/10-group support per locale, attribute and state, original style and 0.9/0.05/T1 gates. The first provisional 40-group / 120-family concept tranche is complete with preliminary AI metadata QA; the remaining 360 groups and full registry review are unfinished. all 400 concepts, lineage, bilingual brevity, bands and splits require review/freeze before new wording.

The next proposed procedure allows one label-blind input editorial pass and at most one retained revision, then a single all 2,400 input seal, fresh locale first votes and separate frame/QA. Editing is planned as AI review, not native-human validation. Human criteria and operational usefulness still need separate evidence. New Fits 0, total 23, model queries/application writes 0; qualification unproven. Source/publicCI remain at #179.

Only verified owned reproducible Go build caches released 274.23 MiB. Original data, votes, models, profiles and proofs remain; owned cache is now about 227 MiB. This is disk cleanup, not inference-memory measurement.

Issue result · Verified bilingual HF aggregates. All 6 files / 28,269 bytes, full bytes/SHA and tag short-core-pilot-v0.1 verified. No raw rows, per-item votes or models published.

The A0 corpus is not admitted for training. Both versions are complete: 2,400 Korean/English rows, 1,200 families and 400 registered groups; 2,400 role-separated blind first votes and 1,200 whole-family frame audits are sealed. TRAIN has 1,680 rows; DEV/CAL/TEST each have 240. Initial votes/audits are preserved and held families stay in all primary denominators.

Author–AI reviewer head disagreements are TRAIN 60/5,040 and protected 14/2,160. Overall agreement is 7,126/7,200=98.97%; KO 3,566/3,600=99.06%; EN 3,560/3,600=98.89%. All three labels agree across a language pair in author 1,198/1,200=99.83% and reviewer 1,184/1,200=98.67% families. AI workflow correlation can hide shared errors. These are not model accuracy, human gold or meaning preservation; role separation and 400 registered groups do not prove full independence.

Frozen failures: TRAIN ≤16 code points KO 167/168 and EN 126/168; EN 17–32 241/252. All six DEV language×head unknown cells fall below 20 families for each reference source; the full-cohort 10-group minimum passes. English short DEV has ≤16 1/30 families and 17–32 20/30. Source-pin, structure and UTF-8 evidence checks pass, without proving semantic correctness. Label equality cannot release frame/style holds, and clear subsets cannot replace full denominators. No label error or impossibility of short English is inferred.

The next short-first approach is a conceptual proposal only: recruit minimal identifiable development-task situations before wording, bin and hash assignment, preserving general context, negation, ambiguity and quoted scope. Require a fresh whole-group version and prospective blind QA. New authoring/replacements have not started; minimums and labels are unforced.

New Fits 0; existing research total 23; qualification false; application writes 0. Models/profiles and protected raw-root barriers remain. Public source has the existing PR #179 tree; no new source/CI result is claimed. Copied-source checks on owned fake fixtures are private structural verification.

The HF failure diagnostic is published and verified in JooYoon/riidolaya-development-natural-short-preparation-v0.1; tag natural-short-preparation-v0.1; revision 7d27ff32c4fdc00e102a265b993ea3d67c6ccd9d. The result below records complete remote byte/SHA and tag verification.

Verified immutable HF diagnostic · Issue result. All six files (43,809 bytes), full bytes/SHA and resolved tag were verified. No raw rows, per-item votes or models were published.

We research tiny Go hints for response/activity/completion reports in development comments. 23 actual research Fits; 0 new Fits this stage; no operationally qualified model. The preceding MLP failed its frozen conditions, with completion precision 24/32, and was not promoted.

The Go audit inspected all 1,680 public training rows. Per locale, activity unknown is 57/840(6.8%) and completion unknown 66/840(7.9%). No text has at most 16 Unicode code points; 17–32 counts are 24 Korean / 0 English. Bytes, code points, tokens and semantic complexity differ. Operational priors and causes of model failure remain unproven.

A frozen hash order selected 40 whole groups / 240 rows without labels or model outputs. A fresh AI role sealed all original-target-blinded first judgments before comparison: 716/720 head agreement (99.44%),236/240 exact three-head row agreement (98.33%). This measures AI reference consistency, not model accuracy or human gold. Four discrepancies and both vote sets remain unchanged. Zero exact normalized duplicates do not prove semantic independence.

PR179 passed full Linux/macOS CI and public-data reproduction, then bot-merged with the same tree. All 8 owned HF aggregate files (40,854 bytes), full bytes/SHA and resolved release tag were verified. Usage and aggregates · Verification record.

bash scripts/verify-claims-training-audit.sh

The script downloads our immutable public corpus and compares Go output with the frozen record. It performs no model inference, training or app writes. One whole audit observed 0.61s wall/0.04s user CPU/13,205,504B peak RSS. Raw profiles, real-work text/IDs and per-item references stay private.

The marking criteria treats ❓/🛠️/✅ as candidate request/activity/completion-report marks. Subtask completion is not whole-task completion. Attribution, scope, reply context and event version belong to the adapter; current models accept text only. Helpful marks still need human evidence.

A new 2,400-row data-preparation plan was frozen before wording: group split 1,680/240/240/240train/dev/cal/test, fresh dev 180 short-focused plus 60 general rows. The later proposal is one matched WarmFit per sibling from the same parent, old versus new training, with no general-comment CE regression. New wording and future Fits have not started. Preserve exposed-evaluation non-reuse, calibration/test seals, existing .9/.05 precision/support rules, default OFF and zero app writes. Prior experiments and failures remain below as history.

We aim to provide inexpensive hints about question, progress and completion-report claims in development comments for humans and agents. A completion report is not proof of completed work. Production defaults remain OFF, application writes remain zero, and no model is operationally qualified.

Cumulative actual research Fits are 23. We fitted fresh linear and 16-hidden-unit MLP arms once each, using the same unchanged 1,680 original synthetic training rows and 2,048 contextual features. Sample order, 40 epochs, batch 32, AdamW, seed 1729 and 2,120 updates were fixed for both. Separate roles authored, blindly reviewed and adjudicated 360 bilingual evaluation rows, 180 paired families and 60 declared related groups. Both fitted models were saved before evaluation was read. The 240 contrast and 120 general-comment rows are AI references, not human gold.

Mean CE delta, MLP minus linear, was +0.190955, with a declared-group paired-bootstrap 95% interval of [+0.125168, +0.258100]. General-comment CE worsened by 0.051024. Request/activity/completion gated-true false positives rose from zero to 1/2/2 in Korean and 1/2/6 in English. Every frozen research-progress condition failed: better mean CE, lower unknown CE in both languages, no per-head false-positive increase, and no general-stratum regression. Zero linear completion proposals establish neither 0% precision nor 100% safety. The prior continuation model's CE of 0.878159 is descriptive only: its 4,240-update history is not an equal-budget control.

The whole command took 2.22 seconds in a single run, with peak process RSS 18,169,856B and Go heap at reporting 7,385,360B. Go 1.27.1 required no GPU/MPS. MLP artifact/model structure/caller workspace sizes are 131,972/131,776/31,000B. File size, static structures, Go heap and whole-process memory are distinct. Single warm observations over the same 16 original synthetic inputs and 20,000 calls measured MLP 6.923µs and linear 7.265µs per call. They establish no statistical speedup or application SLA. Both process allocation counters increased by 16B/one allocation; that cannot be attributed to Predict.

PR178 passed all CI gates and bot-merged with the reviewed source tree. Model bodies remain on the immutable HF research version, not GitHub. Usage · aggregate results · protocol and verification record.

Humans and agents can explicitly select the model with riidolaya claims --model ./claims.rcm --jsonl to inspect research JSONL output. Hints do not automatically change application labels, emoji or task states. Existing models and policies are not automatically replaced.

Protocol deviations remain recorded. Root opened an interrupted 240-row partial-review file to count lines, so we do not claim zero pre-fit file IO. Text and labels were neither displayed nor parsed; a fresh independent full 360-row review did not reuse partial votes. Initial admission failed on ID syntax with zero actual Fits. A compatibility mapping then changed identifiers only and preserved every other byte.

Repeated synthetic scaffolds and correlated AI judgement remain; declared groups are not proven semantically or statistically independent. This original float Go classifier is not a pretrained Laya fine-tune or a BitNet/1.58-bit LLM. Shout-out to laya.tools for prior low-cost semantic tools. Human gold, product usefulness and cost savings remain unproven.

Next PDCA clarifies human/product marking criteria and independent references, then audits annotations, class distributions and missing context using training material only. A separate next hypothesis and fresh evaluation must be frozen before further Fits. This exposed evaluation will not become training, selection or per-case tuning material: no grid, third Fit or automatic quantization on it. Existing .9/.05, precision/support rules and calibration/final-test seals persist. Public assets contain own synthetic training, own weights and aggregates only, with no real-work text, private IDs, per-row scores or raw profiles. Prior parent/ternary models, failures and the historical records below remain preserved.

All 11 owned HF files (1,747,361 bytes), immutable revision, resolved tag and exact Go output on original public examples were verified. Publication verification is separate from operational qualification.

A tiny Go classifier assists with development progress, completion-report and question hints. Completion is a claim about identifiable development work, not proof that real work is complete. Ordinary program/runtime state alone is distinct from a work report. Application writes remain zero and production defaults remain OFF.

There are 21 actual research Fits. This study ran one fixed 40-epoch continuation, adding 2,120 updates from the original float parent, and compared it with the previously frozen equal-update float control. It retained 1,440 original synthetic training rows and substituted 240 new original contrasts. Independent roles authored, blindly reviewed and adjudicated 360 fresh evaluation rows; their body remained sealed until the candidate Fit completed.

The predeclared research-progress conditions passed. CE delta was −0.135844, with a 95% paired lineage-bootstrap interval of [−0.154294, −0.117779]. General-stratum CE worsened by 0.023867. Candidate pooled completion precision of 18/20, 90%, still fails operational requirements; English completion coverage is 5/55, about 9.09%. Both models still abstained on all six head decisions in two private snapshot comments. Human gold, product usefulness and cost savings remain unproven.

This is a 73,988-byte float RSC artifact, not ternary, LoRA or converted pretrained Laya weights. The whole fit/evaluation/save/bootstrap command took 1.07 seconds with about 14.39 MiB peak RSS. Single warm runs on 16 fixed original inputs measured 6.58µs for the candidate and 6.18µs for the control. Both retain 73,784-byte model storage and a separate 30,744-byte caller workspace. These observations establish no speedup or application SLA.

PR177 passed every CI gate and bot-merged with the reviewed tree. All 13 files, 1,622,281 bytes, of the immutable HF release, tag semantic-contrast-v0.1, were read back with exact bytes/SHA, tag resolution and Go example output parity. Parent/control/ternary assets and failure history remain preserved.

Humans and agents can explicitly use Go1.27.1 and riidolaya claims --model model/claims.rsc --jsonl. Usage · aggregates · verified record.

Fixed .9/.05, precision/support requirements and original/existing calibration/test seals persist. The newly exposed 360-row evaluation cannot become later training, selection or per-case tuning material. Only original synthetic fit data, own weights and aggregates are public; no real-work text, private IDs, per-row scores or raw profiles. This is research progress, not operational qualification.

The target is inexpensive development progress, completion-report and question hints. Completion is a textual claim, not proof that real work is complete. No application state is written and production defaults remain OFF. This experiment uses the project's Go classifier trained on original synthetic data; pretrained Laya encoder comparisons remain separate history.

One matched float continuation and one QAT continuation bring the prior 18 research Fits to 20. PTQ adds zero Fits. Float artifacts are 73,988B versus ternary 4,023B; per-model storage is 73,784B versus 3,880B. Count the separate 1,215B shared decoder and 30,744B caller workspace. Whole-process peak RSS was about 10.0–10.7MiB, distinct from a roughly 19-fold model-storage reduction.

Single before/after runs over 16 fixed original synthetic inputs, 20,000 iterations, measured QAT at 17.70→7.21µs and PTQ at 17.64→7.12µs; original float took 7.19µs in the latter round. Fixed artifacts and exact output checksums stayed unchanged, with no extra refit. This shows decoder improvement, not superior speed to float, GPU/MPS acceleration or an application latency guarantee.

No operationally qualified model. Exposed 240-row diagnostic mean CE was original float 0.416372, matched float 0.352619, PTQ 0.558685 and QAT 0.428378. QAT underperforms matched float; Korean completion has zero proposals and English completion4/4 lacks minimum support. All four models abstained on all six head decisions in two private comment observations. Abstention is not semantic correctness or demonstrated usefulness.

PR176 passed every CI gate and bot-merged with the reviewed tree. The immutable HF version, tag ternary-claims-v0.1, has14owned files/150,732B, all read back with exact bytes/SHA and Go example output parity. The standard Hub .gitattributes was inspected separately. Parent assets, failures, .9/.05 gates and old/new calibration/test seals persist. Weights are only on HF; no real-work text, per-row scores or raw profiles are public.

After optional download, use Go1.27.1 and riidolaya claims --model model/qat.rqt --jsonl. Humans and agents receive unqualified research JSON; CPU inference requires no Python/GPU or remote model call. Model usage · public aggregates/scopes · verified record.

The next semantic experiment is separate from compression. Its Plan was fixed before wording/model outputs; independent roles prepare240training and360fresh evaluation rows. Training and26wording repairs passed label-hidden AI review, evaluation review is underway, and new Fits remain0. Independent human gold and real product usefulness remain unproven.

The following historical figures belong to their original targets and splits. The current three-claim development-work corpus is managed separately from the older eight-intent corpus.

The target is inexpensive progress, completion-report and question hints for development comments. A completion report is a textual claim, not verified work. Production defaults remain OFF and all actual application writes remain 0. No learned candidate has passed the prospective quality gates or entered current calibration/final-test selection or promotion.

The full original 1,200-family/2,400-row corpus is complete: 840 train and 120 each validation/calibration/test, 840 declared connected groups. Originals, holds, votes and group-only corrections remain preserved. Disclose the first 41 ID-cue risk, later 319 opaque IDs, selecting-role authorship of 13 non-test cases and correlated AI history. These are not human gold or 2,400 independent product observations. Current calibration/final test remain sealed.

Seven learned Go studies / 17 actual fit arms retain confidence .9 / margin .05 and all support/precision/cost requirements. Earlier internal research choices do not override prospective quality failures. The shared probe uses original AI-reviewed labels, not teacher pseudo-labels or human gold; it performs no backbone fit. Source PR171, PR172 and PR173 all passed CI and were bot-merged. Code CI is separate from model qualification. Read the latest safe aggregate record.

The published shared probe, tag laya-shared-probe-v0.1, contains a 4,320-byte RSP feature head. Its Go cached-feature fit/evaluation phase took 1.72 s with peak RSS 24,182,784 bytes. Fresh prose still requires the large frozen Laya encoder; extraction made 1,680 MPS calls, with whole-command 128.79 s and peak RSS 2,976,038,912 bytes. Small head size/Go-phase RSS do not establish low-memory novel-text inference or a serving SLA.

The separate frozen Laya reference report dataset, tag laya-reference-v0.1, preserves 240 actual MPS calls: whole-command 24.24 s, peak RSS 2,887,172,096 bytes. Base 89/240 and published delta 88/240 had no Korean gated proposals; no qualified teacher was established and formal eligibility was unassessed. Historical temperatures/instructions differ. These resource scopes are not a matched speed/memory comparison; sampled MPS allocation overlaps RSS and is not an additive transient peak.

Earlier immutable research releases remain available: dev-comment domain, features, completion contrast, capacity and smoothing. Follow their bilingual Go 1.27.1 examples and typed loaders; no model is automatically selected by the v1 CLI/SDK. The shared RSP loader consumes frozen feature vectors, not text. Models and report assets live on HF, not as binaries in Git.

The historical zero-comment snapshot did not mean zero body observations: 54 body observations already existed. A new authorized trial processed two real comments without training on private data. AI reference judgments were completed before the reviewer saw model outputs: neither comment contained an explicit question; one was a reminder with an unresolved response-request intent, and the other an attributed deployment/verification report with hedged completion. Neither verified task completion. The pinned parent v0.2 output completion_report at confidence 0.9091 for the reminder-like comment and progress at 0.5282 for the hedged report. Both primary-intent predictions disagreed with the AI reference. The unlearned literal/event assertion filter produced 0 candidates. There is no human gold, and a sample of 2 is not an accuracy/precision estimate or proof of filter safety.

These are one-run observations, not steady inference SLAs or a guaranteed speedup. PR174 passed Linux, macOS, secrets and quality checks and was automatically merged. It is a source-only unlearned assertion shadow, with no model, Fit or probabilities. It remains distinct from learned semantic hints. No labels, emoji, annotations or task states were actually written. Real names, texts, IDs, URLs, private paths and raw profiles are absent from this public record.

Most prior fits are standalone Go linear/MLP supervision; the shared probe trains a Go head on frozen Laya representations. No LoRA, ternary or backbone tuning is claimed. Shout-out to laya.tools; model licenses, provenance and limitations remain explicit. The previous preparation records below are unchanged.

PR175 adds the Go SDK, explicit riidolaya claims JSONL preview and fixed study runner. Response requests, current development work and completed work are separate true/false/unknown heads, with learned unknown distinguished from confidence abstention. The 73,988-byte RSC format is neither process memory nor 1.58-bit weights. Full local and GitHub CI checks passed; the exact tested tree was automatically merged.

Preserve 1,520 initial broad-code-domain annotations without fitting. One 840-row packet lacked response-unknown cases, and some annotations confused program events with work reports. The clarified target is presence of development-work reports. Runtime behavior and conflicting facts do not themselves establish completed work or unclear current-work claims.

New original data contain 1,200 bilingual families / 2,400 rows / 400 related story groups, split before authoring into 280 training and 40 each validation, calibration and test groups. Three distinct comment situations in a story share its partition. Independent label-hidden AI review covered 840 training families / 1,680 rows and 120 validation families / 240 rows. Eleven training-row differences were adjudicated with originals preserved; validation had no differences. Calibration/final-test bodies remain sealed. AI references are not human gold; prior data, seals, models and gates are unchanged.

One actual fresh Fit made 2,120 updates over 40 epochs. Fit/evaluation/save/reload took 0.83s whole command with peak RSS 13,975,552 bytes (about 13.3MiB), a single observation rather than an inference SLA. Gated completion proposals were 0 Korean and 3/4 correct English: 75% completion precision and insufficient support failed screening. Unknown activity/completion references had only 3 families per locale, below the predeclared 5. All six attributes in two exposed private-comment diagnostics abstained on low confidence; abstention is not semantic success.

The HF research model, tag development-claims-v0.1, was published with all 15 uploaded files byte/SHA-verified after redownload and SDK output parity. Standard Hub .gitattributes was reviewed separately. Only original training prose is packaged, with no evaluation/private prose or profiles. This is one separate-target Fit after the prior 17 studies, bringing actual Fits to 18; no operational qualification is claimed. See the actual result and checklist.

The train corpus contains 840 fictional scenario families, 1,680 Korean/English rows and 708 declared related-case groups, with 105 families per intent. Of 931 authored families, 91 are held; six further reserve designs have no wording. AI authorship and label-visible AI review are not human gold or independent product observations.

HF preview005 is published. All 35 files / 5,260,256 bytes were downloaded and matched against original bytes and SHA-256. The downloaded training corpus also passed the actual Go structural check. Tag v4-preparation-005 resolves to the same commit. Read the English usage card. All previous 001–004 assets and tags remain fixed. Releases are cumulative: do not concatenate them.

Train ancestry is frozen across 937 original, held and reserve designs. Separate heldout cases are being checked against that ledger. Authoring, design review and exact-prose review with labels hidden initially are separate roles; AI role separation does not establish statistical independence or human truth.

The full 1,200-family/2,400-row study still needs 360 validation/calibration/test families and 720 wordings. New V4 fitting, prediction and operational mutations remain zero. Four fixed learning conditions and the 0.9/0.05 display gates are unchanged. See the review record. The sections below retain historical snapshots.

The train-only corpus now contains 593 fictional scenarios, 1,186 bilingual rows and 537 declared lineages. Of 684 original candidates, 91 remain held for semantic, scope, technical-coherence or novelty concerns. AI reviewers read exact texts model-blind with expected labels visible; anchoring and synthetic-author bias remain.

HF preview004 is published: all 29 files / 3,268,472 bytes were downloaded and matched against original bytes and SHA-256. The fixed tag resolves to that commit. Read the English usage card. All prior assets and tags remain unchanged. Releases are cumulative snapshots; do not concatenate them.

Three pre-fit reinterpretations retained selected-range review closure, input-readiness closure and explicit stop-now purpose. Original text, labels, disagreements and full field changes remain preserved; relabeling created no new scenarios. Author self-reconsideration is not a second independent vote. See the review record.

The full 1,200/2,400 corpus and 840/120/120/120 partitions remain incomplete: 247 more train families and 360 separately prepared validation/calibration/test families are needed. New V4 fitting/forwards, heldout authoring and operational changes remain zero. Sections below preserve historical snapshots.

The reviewed train-only subset now contains 450 fictional scenarios, 900 bilingual rows and 420 declared lineages. Of492 authored candidates/984rows,42families/84rows remain held for semantic, scope, technical-coherence or novelty concerns. AI reviewers read exact texts with expected labels visible; anchoring and correlated judgement remain. This is not human gold or product accuracy.

HF preview003 is published: all23files/1887186bytes were downloaded and compared against original bytes and SHA-256. Tag v4-preparation-003 resolves to that commit. Read the English usage card. Prior001/002assets and tags remain fixed; do not concatenate these cumulative snapshots.

All192texts in the latest batch were reviewed and90scenarios retained. Six direct same-scope contradictions were included as unclear, while six scenarios remain held, including one describing explicitly different operations. Related functional contract contrasts remain grouped conservatively. See the latest review.

Complete1200/2400data, grouped840/120/120/120partitions and execution locks remain pending. New V4 fitting/forwards, heldout authoring and operational changes remain0. Existing models and.9/.05gates stay fixed while original authoring continues in parallel. The002/001sections below preserve historical snapshots.

The reviewed train-only subset now contains 275 fictional scenarios,550 bilingual rows and269 declared lineages. Of300 original candidates/600rows,25families/50rows remain held for semantic, technical-coherence or repeated-template concerns. An AI reviewer read every exact text with expected labels visible; anchoring and correlated AI judgement remain. This is not human gold or product accuracy.

HF preview002 is published: all17files/855062bytes were downloaded and matched against original bytes and SHA-256. Tag v4-preparation-002 resolves to that commit. Read the English usage card. Preview001 assets and tag remain fixed; the two previews are cumulative snapshots and must not be concatenated.

All four PR165 CI gates passed and the bot merged the source-identical tree. The real Go structure checker also passed on550rows. Complete1200-family/2400-row data, splits and execution locks remain pending. New V4 fitting/forwards, heldout authoring and operational effects remain0. See the latest checks.

The numbers and CI-pending statements below preserve the earlier preview001 snapshot.

The goal is inexpensive assistance for progress, completion reports and questions in development prose using a tiny Go model. A completion report is an author's claim, not verification that a task is done. This work prepares judgments for labels or emoji; it creates no display and writes no task state at this stage. Product cost savings and quality gains remain unverified.

The comparison parent is the previously trained Go v0.2 model, 32,960 bytes (about32KiB). The new fitting driver also uses Go cross-entropy and AdamW. It needs no Python or GPU and is separate from Laya neural-network, LoRA and ternary-weight training. Existing models and display gates are not automatically replaced.

The current data covers newly authored fictional development scenarios, all in the training partition.

Originals and disagreements are retained. Names or filler changes do not create new scenarios. The current subset is class-unbalanced, and the full1200-family/2400-row corpus is unfinished. This is AI-authored and AI-reviewed material, not a human-verified golden set or independent product observations. New V4 fits:0; authored heldout rows:0; state writes:0.

People and agents use the same Go command. First check the structure of one supplied training file.

riidolaya corpus-check --partition train --rubric-sha256 "$RUBRIC_RECEIPT_SHA256" < train.jsonl

It checks bilingual pairs, duplicate IDs, declared lineages and input bounds, returning counts only. A pass does not establish semantic truth, translation correctness or training rights. Explicitly supplying a final-test file does read it; this command does not enforce the fitting driver's final-test seal. See the structure-check guide.

A complete ready-to-run plan is still pending: it must pin the full originals, splits, reviews, source and executable. The following maintainer check is for when that plan exists; the example paths do not refer to files already created.

go build -trimpath -o .cache/statehint/v4-fit ./cmd/riido-statehint-tune-v4
.cache/statehint/v4-fit --plan .cache/statehint/v4/PLAN.ready.json --check

Use Go1.27.1 and a verified clean source commit. --check reads the parent and train/validation/calibration files, with no model forwards, fitting or final-test text access. Actual fitting is a separate explicit --fit procedure. The fixed proposal compares three parent warm starts with one fresh start, emphasizing wrong completion proposals in validation. The .9/.05 display gate stays fixed. Separate calibration and a saved lock must precede final-test access. The driver guide and limits currently point to source under review.

At this draft's snapshot, PR164 passed CI and merged at c0d8b16; its merged tree was checked against the reviewed tree. PR165, containing the new fitting driver, is awaiting CI. Code checks are separate from a new fitting run or product-quality evidence. Dataset publication is handled separately and the verified preview link is recorded below.

This page adds V4 preparation status while preserving existing Wiki context. Earlier measurements, failures and status-at-the-time are not rewritten. The V3 record and chronological research history remain available.

한국어

The public preparation dataset preview001 contains 190 train rows. All 11 files / 233373 bytes were reread at the pinned commit; all 10 explicit assets matched their original bytes and SHA-256. Tag v4-preparation-001 resolves to that same commit. This is not a model or a completed 2400-row training result.