riidolaya-development-claim-hints-semantic-contrast-v0.1

This is an experimental, CPU-only Go classifier for three claims expressed in English or Korean development comments. One fixed semantic-contrast data continuation reduced mean three-head cross-entropy on a newly sealed AI-reference research evaluation from 0.996490 to 0.860646, compared with a fixed float continuation control. The paired difference was −0.135844, with a 95% lineage-bootstrap interval of [−0.154294, −0.117779]. The prespecified research-progress conditions were met.

The general stratum got worse: its candidate-minus-control CE difference was +0.023867. Completion hints remain unqualified, and these AI-reference results do not establish human agreement, private-work usefulness, or production performance. The model is for explicit, optional research use by people and agents. It grants no authority to change a task or application state.

What the model emits

Ordinary program/runtime state is not by itself a development-work report; surrounding wording must report someone’s work.

Every head has the categorical reference states true, false, and unknown. A completion claim is an attributed textual claim, not a verified completed task. Unknown is a learned semantic target as well as the emitted abstention state; unknown_reason distinguishes semantic_unknown, low_confidence, low_margin, untrained, and no_word_content.

An emitted true or false still requires the unchanged confidence floor 0.9, margin floor 0.05, temperature 1, trained weights, and word content. These numeric gates do not qualify semantic quality. Probabilities are not calibrated guarantees. Only text enters numeric features; reference labels, byte-span evidence, IDs, groups, and review provenance do not enter inference features.

The model uses 2,048 contextual hashed feature bins, three categorical heads, nine output columns, and float32 weights. Its RSC v1 artifact is 73,988 bytes. This is the same float format and footprint as the parent and matched control. It is not a ternary artifact, LoRA adapter, conversion of Laya weights, or a teacher model. There is no LLM runtime or GPU requirement.

Fixed experiment and data provenance

The plan was frozen before new wording and model outputs. Exactly one new float WarmFit copied the original parent into a separate receiver, reset AdamW moments and optimizer step, and ran 40 epochs, batch 32, learning rate 0.001, weight decay 0.01, seed 1729. The candidate records 2,120 inherited steps + 2,120 new updates = 4,240 total steps. The existing research ledger had 20 fits; this continuation brings its disclosed cumulative total to 21 research fits. There was no seed/epoch/parameter sweep, QAT, calibration, selector, or retry on the fresh evaluation.

The fixed training corpus has 1,680 rows, 840 paired English/Korean families, and 280 lineage groups. It retains 1,440 publisher-owned original public training rows and replaces 240 rows with newly authored, AI-reviewed original semantic-contrast text: 120 families in 40 groups. Replacement groups were selected by the first 40 SHA-256 values of claims-semantic-contrast-v1: followed by the original public group identifier, without using labels or model scores. Each lineage contains three families and six rows. Original data and models were preserved.

The independent evaluation contains 360 rows, 180 paired families, and 60 lineage groups: 240 contrast rows in 40 groups and 120 general rows in 20 groups. General labels were not forced into class balance. Each locale's wording was rated independently; paired locale references need not agree. IDs, families, and declared groups are disjoint across training and evaluation. The evaluation body stayed sealed during check mode and was opened only after the sole fit completed. The matched control was loaded after that fit as well.

Authoring and initial reference review used separate AI roles without model queries. The 240 new training rows received initial blind review; wording-repair confirmation had the reviewer's prior reference knowledge. The final evaluation combines 348 initial author/review agreements, 9 adjudicated rows, and 3 independently rerated wording-repair rows. Its initial full review was label-hidden; later adjudication and repair confirmation had disclosed prior-reference exposure. Those repairs are not a fresh, fully blind 360-row review. There is no human gold reference.

A deterministic lexical/anonymized split screen and bounded semantic inspection found no concrete copied scenario or identity-overlap blocker in the pinned inputs. That audit is not proof of complete semantic-template independence. Repeated synthetic language, separate IDs, and paired locales do not establish 60 independent real-world stories. Reference validity and real-world generalization remain uncertain.

Research evaluation

CE is averaged over the three heads within each row. Candidate and matched-control losses are paired within a declared lineage, keeping both locales and all six rows together. The bootstrap independently resamples 40 contrast groups and 20 general groups with replacement, uses fixed contrast/general weights 2:1, 10,000 resamples, and seed 1729. Its 2.5% and 97.5% endpoints use linear interpolation at q × (N − 1) in sorted draws. Target probabilities use a reported CE floor of 1e-15; no target was clipped in this run. The interval is conditional on these synthetic data and grouping assumptions.

Contrast CE difference was −0.215699; general CE difference was +0.023867. The research rule required the paired CI upper endpoint below zero, lower unknown-target CE in both locales, and no increase in gated-true false positives in any of the six locale/head cells. All three conditions passed. This rule measures research progress; it does not change existing qualification or support requirements.

Candidate gated-true diagnostics are against AI references. A false positive means an emitted true on either a false or an unknown reference. Each locale has 180 evaluation rows. One row per family per locale makes correct counts equal to correct-family support; some families share a lineage.

Completion precision pooled across locales is 18/20 = 0.90, compared with 24/26 ≈ 0.923077 for the matched control. English completion true coverage is 5/55 ≈ 9.09%, and Korean completion true coverage is 13/54 ≈ 24.07%. The candidate does not deliver a qualified completion route; perfect observed precision in other small cells also does not establish qualification. Original and existing calibration/test bodies remain sealed, and all qualification gates remain unchanged.

A separate, already exposed fixed diagnostic containing two private comments yielded six of six low-confidence unknown head outputs for both models. No wording, IDs, labels, or per-case probabilities are released. This small result establishes no new private-work usefulness or improvement in gated task states, and was not used to author data or select this fit.

The fresh 360-row evaluation is now exposed. It must not become training data, the next selector, or a reusable source of per-case improvements. A new study would require new independent evidence and its own frozen protocol.

Resources and artifact checks

On the recorded CPU-only Go 1.27.1 run with CGO disabled, the external whole-command wrapper measured 1.07 seconds wall time and 15,089,664 bytes (14.39 MiB) peak resident set size. The command included the fixed continuation, evaluation, artifact save/reload, and enabled private CPU profiling. End-of-run Go heap allocation was 5,434,696 bytes; Go heap is not process peak RSS, and neither measurement is GPU memory. This single run is not a portable cost or latency guarantee.

A separate fixed 16-item public/original-fixture benchmark used 32 warmups and 20,000 calls per model, measuring Predict plus a prediction checksum. Text lengths were 62–172 UTF-8 bytes. Mean time was 6.5826 µs/call for the candidate and 6.1844 µs/call for the matched float control. Both had 73,784 bytes fixed model storage and a 30,744-byte inference workspace. Candidate scoped process Go counters observed zero warm-loop allocations; the control recorded one unattributed 16-byte allocation over the whole loop. Background activity was included, so that observation is not a causal allocation comparison. The benchmark loaded only the selected float models; older PTQ/QAT pins required by its protocol were checked but were not comparison arms. There is no CPU speedup claim, significance claim, or inference SLA.

The candidate was reloaded from its saved artifact and reproduced exact serialized bytes, all 360 evaluation predictions, and all 360 score arrays. Parent and comparator save parity confirmed read-only numeric use. The fit's 0.202676 mean loss is an average of the online training trajectory across all epochs; it is not final evaluation CE.

Pins and source

Source PR #177 implements the fixed runner and synthetic safety tests. Local full race tests, vet, formatting, and redacted source/history scans passed before this fit. CI run 37575092999 was still running at source freeze and later passed every gate; the bot merge preserved the reviewed source tree. The separate benchmark source was commit 799eaee6f0c1ffa424e2d5ecdc4213efe62b5ca7.

This release contains only publisher-owned training text/references, the candidate weight artifact, safe aggregate diagnostics and method, and the explicit Go example. It excludes evaluation prose and per-row references/scores, private comments or task data, optimizer state, CPU profiles, raw resource logs, and local paths. Apache-2.0 covers the publisher-owned release files subject to the publication manifest and notices.

Explicit optional Go example for people and agents

This artifact uses pkg/statehintclaims.Load and Predict. The user-facing command is riidolaya claims --model model/claims.rsc --jsonl. Save the following as claims-example.go in a checkout of the pinned laya-tools source and use Go 1.27.1. The example verifies the candidate's full-file SHA and bounded size before loading. It does not train, calibrate, download automatically, register a tool, or write application state.

package main

import (
    "bytes"
    "crypto/sha256"
    "encoding/hex"
    "encoding/json"
    "flag"
    "fmt"
    "io"
    "os"

    "github.com/teamswyg/laya-tools/pkg/statehintclaims"
)

func main() {
    modelPath := flag.String("model", "", "explicit local claims.rsc")
    text := flag.String("text", "", "text supplied explicitly by a person or agent")
    flag.Parse()
    fail := func() { fmt.Fprintln(os.Stderr, "claims research input or model invalid"); os.Exit(1) }
    if *modelPath == "" || *text == "" || flag.NArg() != 0 {
        fail()
    }
    f, err := os.Open(*modelPath)
    if err != nil {
        fail()
    }
    data, readErr := io.ReadAll(io.LimitReader(f, statehintclaims.ArtifactBytes+1))
    closeErr := f.Close()
    sum := sha256.Sum256(data)
    if readErr != nil || closeErr != nil || len(data) != 73988 ||
        hex.EncodeToString(sum[:]) != "cfd35ee70a23f94a8d470b7dea596244a91ac7f1410475e8a42959693c99864b" {
        fail()
    }
    model, err := statehintclaims.Load(bytes.NewReader(data))
    if err != nil || model.TrainingSteps() != 4240 || model.Temperature() != 1 {
        fail()
    }
    var workspace statehintclaims.Workspace
    prediction, err := model.Predict(*text, &workspace)
    if err != nil {
        fail()
    }
    result := struct {
        Mode              string                     `json:"mode"`
        Qualified         bool                       `json:"semantic_quality_qualified"`
        ApplicationWrites int                        `json:"application_state_writes"`
        Prediction        statehintclaims.Prediction `json:"prediction"`
    }{"research_claim_hints", false, 0, prediction}
    if json.NewEncoder(os.Stdout).Encode(result) != nil {
        fail()
    }
}

Download claims.rsc from the repository at the release's verified Hugging Face commit revision. Then invoke the example explicitly with your own text:

go run claims-example.go --model claims.rsc --text "$CLAIM_TEXT"

The JSON output is suitable for a person's inspection or an agent's read-only research workflow. Optional hints must remain separate from verified task facts and any application's authorized mutation policy.

Riidolaya is an independent experiment in laya-tools. Shout-out to laya.tools for the inspiration; that acknowledgment is not a claim that these weights were converted from Laya or endorsed by it.

Source PR177 passed all CI gates and bot-merged with the exact reviewed tree. Code CI does not qualify the model. Fixed release tag: semantic-contrast-v0.1.

Download the optional model and verified Go example at the fixed release tag:

hf download JooYoon/riidolaya-development-claim-hints-semantic-contrast-v0.1 claims.rsc claims-example.go --revision semantic-contrast-v0.1 --local-dir model
go build -trimpath -o riidolaya ./cmd/riidolaya
printf '%s\n' '{"text":"I am running regression tests now."}' | ./riidolaya claims --model model/claims.rsc --jsonl

Only optional download uses the Hugging Face CLI; CPU inference itself uses Go, with no Python or remote model call. Both the CLI and explicit example identify unqualified research output.