JOB 130 PUBLICATION PROSE — SOURCE AUDIT

Scope: tool introduction, news, and blog/reflection HTML bodies. No private conversation excerpts, no new scientific experiment, no online Carrier edits.

Publication routes (root agent to verify and package):
https://www.unifiedstate.us/carrier/handoff-desk/
https://www.unifiedstate.us/carrier/handoff-desk/downloads/carrier-handoff-desk.zip
Three experiment routes use retrieval-lab/, retrieval-lab-2/, retrieval-lab-3/.
User reports all three published. Root agent subsequently verified all three page routes returned HTTP 200 and expected titles; this does not verify live functionality. The editorial evidence audit used the local public artifacts.

Read and checked:
- 130/READ-ME.txt
- 130/tools/core.js and app.js implementation
- 130/data/seed.json (6 source-backed notes, 5 meanings, 2 sources)
- 130/checks/result.json (20 passed checks; Node/minimal-DOM, mocked Blob download)
- 127/upload/carrier/retrieval-lab/downloads/technical-report.html
- 128/upload/carrier/retrieval-lab-2/downloads/technical-report.html
- 129/upload/carrier/retrieval-lab-3/downloads/technical-report.html
- 127/posts/blog-body.html and 128/posts/blog-body.html for detailed methods and qualifications.

Key checks:
R01 MRR .914778 lexical vs .915184 graph; tiny effect, no established reliable benefit. Definition-conditioned synthetic questions.
R02 pooled nDCG .6741 dense vs .6707 typed; answer .9086 vs .9225 but no established benefit; incomplete model-judged pool, sparse typed overlay, shared model family.
R03 retrieval pooled nDCG decreases under requirement reranker; all-task coverage improvement does not persist on historical final36. Reused public tasks, not fresh confirmatory test.
R03 handoffs: 768 scored fields, 128 outputs, 8 base templates, 4 representations, all at ceiling. Definitions supplementary; no superiority, equivalence or general hallucination-reduction proof.

Prose deliberately distinguishes negative ranking results from nonconfirmatory or ceiling findings. It does not say Carrier was proven useless or all methods shown equivalent.

Desk claims verified against code:
- Frozen 6257 concept catalog; no automatic live gate fetch.
- Checks exact selected definitions and literal evidence text; no semantic entailment or authenticity claim.
- Claim kinds stay distinct; claimed decision requires authority text but does not authenticate it.
- Returned JSON comparison does not mutate current draft/checkpoint.
- Import is separate and replaces draft/checkpoint.
- JSON and self-contained AI text exports; no automatic saving or draft transmission.
- Current format is experimental JSON/text, not USL1/CVP2.

The posts omit a final verification count because publication integration may change the code or checks. The download guide should retain the actual final checks and their limits. Do not claim physical-browser, mobile, accessibility or live WordPress validation from Node DOM checks.

Reflection is authored public editorial opinion by Lumin. Honesty is treated as a practice, not a quantitatively established project outcome or a substitute for methodological limits.
