Live Claim Hints Continuation 2026 10 10 English - teamswyg/laya-tools GitHub Wiki
Development comment hints: progress map on 2026-10-10
When text changes, a small local model independently scores 🛠 current-work reports · ✅ completion reports · ❓ response requests. This research preview suggests possible labels or emoji. A reported completion does not verify that the actual task is finished.
Try the preview
Follow the installation and launch guide, then open http://127.0.0.1:8877/. Choose an example or edit the text to update the output. Insufficient confidence produces abstention. Text is sent only to the local server and is not stored. Actual Riido labels and task states are not changed.
The current model is an authored Go feature classifier, not converted Laya weights. Its artifact is 73,988 bytes and runs on CPU without a GPU.
Current stage
| Stage | Status | Meaning |
|---|---|---|
| Local web preview | Running | Observe predictions and abstention while editing text |
| Complete review of new situations | 315/400 declarations confirmed | 305 current passes, ten holds, 85 awaiting qualification |
| Amendments and independent re-review | Ten current holds | Amend official findings or supporting evidence while preserving prior records |
| Updated whole-cohort check | Plan and namespace ready, not executed | Zero metadata findings for two earlier observation changes; later changes need refresh |
| Author reference joins | 186 original declarations saved | 88 independent reviewer references joined; actual findings, timestamps and histories preserved |
| Preparation-check automation | Go guard source reviewed; no actual automatic review | Checks inputs, methods and read budgets; no semantic judgments |
| Bilingual comment construction and QA | Pending | 1,200 communication families, 2,400 comments |
| Matched comparison training | Pending | Exactly two matched fits from the same parent |
| Usefulness on actual work | Unproved | Review counts and internal scores are not accuracy |
Partial success does not certify the whole study. Ambiguity remains explicit. No new fit has run in this continuation. Complete material and reference QA precedes comparison of old training material with new material.
Reading the results
Current model quality is unqualified for production. One newly authored current-work/question sentence produced abstention on all three heads. This separate smoke check used no protected research material and is not an accuracy evaluation. Existing decision thresholds remain unchanged.
Model execution time differs from complete web latency. Go profiles differ from total process and GPU memory. The recent idle server RSS snapshot of 12,416 KiB is neither peak memory nor a minimum-hardware claim. Historical read-accounting and input-scope evidence gaps remain recorded; cumulative bytes do not measure physical disk I/O.
Profiling guide · Earlier 205-review snapshot · Accumulated experiments
Earlier 225-review public English HF snapshot: only three public progress files were added and verified after publication. No new model weights were uploaded.
The matched runner built with Go1.27.1 after exact verification of29 actually compiled Go files and four module files. No protected training/evaluation material or model was read, and no fit or evaluation ran.
The 315 confirmed records cover sources 1–315. Twenty situations with insufficient historical read-accounting and input-scope evidence underwent genuinely new observations of registered native inputs. The new results are seventeen passes and three support-only holds; two historical passes became holds in the new review. Prior failures, findings and timestamps remain preserved, with new actual observations and judgments recorded separately. Historical procedures are not retroactively certified. READ, WRITE, process output and actual model display are distinguished.
Public bilingual HF snapshot at 305 reviews: only two documentation files were added and verified by download from the immutable revision. This is not a new weight release or training result.
Latest ten reviews completed actual whole-evidence judgments. Oversized final metadata was sharded from already-recorded findings and actual timestamps without replaying semantic inputs; the failure remains preserved. Twenty more situations are under parallel review. The 315-review snapshot adds no trained weights.
Public bilingual HF snapshot at 315 reviews: two public documents were added and verified by downloading the immutable revision. No new weights or training results were released.