# Copy-ready arXiv metadata

## Title

From Seed Lotteries to a Shippable Local Tab Namer: A Staged Study of Data, Decoding, and Evaluation

## Authors

MD Ishtiaque Hossain (PorkiCoder Research, Vancouver, Canada)

**Identity check before submission:** preserve this form only if `MD` is part of the author's publication name. ArXiv treats the author field as identity metadata, not decorative text or a degree field.

## Abstract

Naming a coding-agent terminal tab is a small generation task with hard product constraints: the title must be locally generated, useful, short enough to scan, and cheap enough to run on every new interaction. This paper consolidates a ten-study development campaign that began with a 12.7M-parameter T5 student and ended with a 76.96M-parameter task-tuned FLAN model selected for shipment. The campaign separates model weights from candidate construction, decoding, deterministic post-processing, data selection, and evaluation. An initial checkpoint appeared to reduce repeated-token titles by 37%, but a deterministic 31-run matched control did not support the proposed literary sequence-order mechanism. On a Stack Overflow board, a deterministic centroid highlighter and constrained four-beam search turned an unchanged 35M model into a strong relative system: on a sealed 1,000-row comparison it scored 6.20 versus 5.45 for title-tuned FLAN with the same highlighter, a paired difference of +0.752. However, an independent Codex audit rated only 19 of 40 outputs useful, below a preregistered 60% absolute gate. The same 35M stack then trailed FLAN by 1.17 points on a separate terminal-task board. Selective judged supervision produced small replicated gains, while 175,642 legal synthetic labels reduced cross-entropy to 0.91 without improving judged utility. A 24.5M pointer system reached 3.88 ms per title and beat the 35M decoder by +0.14 on one packet, but a task-tuned 77M FLAN model beat it by +0.50. Finally, leak-checked continuation training improved the frozen prior ship from 7.38 to 7.59 and from 7.34 to 7.55 on two never-seen 1,000-row holdouts. The central finding is that narrow local generation is a system-design problem: data selectivity, domain, decoding, glue, capacity, and evaluation protocol can each change the conclusion, and their effects are strongly competence-dependent.

**Length:** 1,909 ASCII characters, below arXiv's 1,920-character limit.

## Comments

18 pages, 2 figures, 5 tables. Consolidates ten public staged studies of a privacy-preserving local tab-title generator. Public research notes and claim boundaries are linked in the references.

## Suggested categories

- Primary: `cs.CL`
- Cross-list: `cs.LG`
- Optional: `cs.SE`

## Suggested keywords

local inference; short-form generation; constrained decoding; synthetic supervision; small language models; LLM-as-a-judge; on-device NLP; coding agents

## Journal reference and DOI

Leave blank for the initial preprint unless another published version already exists.

## Source upload

Upload `porkicoder_tab_namer_arxiv_source.zip`, not the complete working package. Its root contains `main.tex` and the `figures/` directory only.

## Final submission checklist

1. Confirm the author's exact identity spelling and current affiliation.
2. Select an arXiv license deliberately.
3. Paste the ASCII metadata above rather than text copied from a PDF viewer.
4. Use `cs.CL` as primary unless an endorser or moderator recommends otherwise.
5. Preview the arXiv-generated PDF and compare the figures, tables, URLs, and page count with `porkicoder_tab_namer_paper.pdf`.
6. Do not describe the Codex audit as human evaluation.
7. Do not merge SO-board and Terminal-board numbers in the abstract or announcement text.
8. Do not claim that the final public installer already runs Session 11; the note says the public installer still served 6t at publication time.
9. Keep the closed holdouts and private terminal text private.
10. After arXiv assigns an identifier, add it to the PorkiCoder research index and, in a later version, to the paper's references if appropriate.