公开记录项
A compact candidate is useful only when the page can show why it was produced and what was lost.
公开记录项
JustAnIota separates deterministic WordPress registry demos from the Protocol5 vector experiment track. Both paths must expose mode, score, evidence, Unicode sequence, registry reference, warnings, segment trace, ranking lane, vector evidence, source-evidence atlas, read-only or local-only boundary, and round-trip drift where available.
公开记录项
Approximate semantic retrieval is evaluated by provenance, ranked candidates, measured drift, and negative cases. It is not treated as exact reconstruction, standard conformance, production certification, or model-independent encoding.
Result modes
- DatabaseOnly: deterministic lookup or fallback behavior from reviewed records without live model assistance.
- CachedSemantic: previously computed descriptor or embedding evidence reused with visible version and provenance metadata.
- Hybrid: stored vectors are preferred, and live embedding assistance may fill gaps only when the result discloses that help.
- AiAssisted: live model assistance was used and must be disclosed in the result record.
- RoundTripApproximate: reverse output was compared with the source, and drift remains part of the result rather than being hidden.
Evidence fields
- Source text, locale, direction, normalization, and Unicode scalar sequence.
- 注册表 ID, registry digest or version, source descriptor provenance, and canonical envelope metadata.
- Candidate score, ranked alternatives, candidate ranking lanes, warning list, unsupported cases, and negative examples.
- Segment trace with raw text, lookup key, match level, consumed token indexes, fallback or unknown warnings, and provenance when present.
- Evidence summary counts, source-evidence atlas families, match-level counts, candidate-ranking-lane counts, vector coverage, and unknown rate when an experiment path can provide them.
- Segment vector evidence such as dimensions, magnitude, min/max, mean, and bounded coordinate preview when stored vectors are part of the path.
- Seed-only versus SQL-backed ranking-delta evidence when a hosted experiment compares the packaged seed registry with the Protocol5 category corpus.
- Read-only API status and local-only population boundary when a public endpoint depends on private or desktop-side embedding work.
- live_ai_used flag plus model or embedding metadata when model assistance is part of the path.
Protocol5 evidence promoted into IOTA-1 vocabulary
- 状态 endpoints may report profile, approximation flag, public seed count, SQL corpus reachability, table row/embed counts, vector dimensions, and configured live-AI state.
- Ranking demos should compare seed-only and SQL-backed paths by output changed, top candidate changed, ranking-lane deltas, source atlas counts, vector coverage, and unknown rate.
- Search endpoints should be read-only over reviewed stored vectors or caller-supplied embeddings; they must not populate, mutate, train, or widen support claims from the public web route.
- Population belongs to local Protocol5 tooling such as EmbeddingDesktop or reviewed scripts, with source rows, embedding version, model, dimensions, and hashes recorded outside public mutation routes.
{
"profile": "segment-vector-evidence-v1",
"mode": "database-only",
"segment_trace": [
{
"raw_text": "good help",
"lookup_key": "good-help",
"match_level": "category",
"source_evidence": ["category-corpus", "plain-language-descriptor"],
"found_vector": true,
"vector_dimensions": 1998,
"vector_preview": [{ "index": 12, "value": 0.18, "magnitude": 0.18 }]
}
],
"evidence_summary": {
"candidate_ranking_lanes": [{ "name": "sql-category-corpus", "count": 1 }],
"source_evidence_atlas": [{ "name": "public-unicode-assignment", "count": 1 }],
"vector_coverage": 1.0,
"unknown_rate": 0.0
},
"population_boundary": "local-only; public APIs are read-only"
}[
"segment-vector-evidence-v1",
"database-only",
[
[
"good help",
"good-help",
"category",
[
"category-corpus",
"plain-language-descriptor"
],
true,
1998,
[
[
12,
0.18,
0.18
]
]
]
],
[
[
[
"sql-category-corpus",
1
]
],
[
[
"public-unicode-assignment",
1
]
],
1,
0
],
"local-only; public APIs are read-only"
]字段顺序遵循有键 JSON 示例、已发布的模式顺序以及公共字段注册表。
Review rule
- Public copy must say approximate semantic match when a path is approximate.
- Low-confidence results should be surfaced as warnings, alternatives, or rejection reasons, not rewritten as certainty.
- A compact token, glyph, scalar, hash, embedding vector, or screenshot is evidence only when the surrounding profile and registry record make it inspectable.
- Any future support claim must align page copy, machine artifact, validator behavior, implementation evidence, and release notes.