# Conflict-wording study: dataset

IsoKron, November 2026. This folder holds the registered design, the five amendments, the harness, every run record, the spend ledgers, the final-message coding sheets, the coded corpus of conflict messages and its independent second coding. The study page is at https://isokron.ai/research/mcp-conflict-errors-ai-agent.

## The corpus (`corpus.json`, `2-corpus.md`)

110 conflict and stale-write messages that real MCP servers, databases, APIs, editors and tools return, each with its project, version or commit, file and line, the text and five codes.

- **The codes describe message content only.** They record what a message tells the caller: whether the write was applied, what changed, the suggested next step, whether it hands back the current value, and its length. They are not grades of a project's quality or safety. A project whose message scores low on these codes may handle conflicts correctly in every other respect.
- **Terse messages may be deliberate.** A short message can be a design choice: for security, for a low-level transport, or because the caller is expected to read a status code or a structured field instead.
- **The sample is exploratory, not a ranking.** Projects were chosen from a brief and from well-known repositories; popularity was not measured, and the corpus does not compare projects with each other. See "Method and limits" in `2-corpus.md`.
- **Each quote links to its source.** `source_url` in `corpus.json`, and the Source column in `2-corpus.md`, link to the file and line at the pinned commit, or to the documentation page. Observations describe that commit or the page on the access date, not necessarily current behaviour.
- **Quotes are trimmed** to under 15 words, with an ellipsis. Length codes measure the full message or template.
- **Silent paths.** The corpus lists six MCP servers with a write path where the server settles the version or forces the write itself, so a stale write is never reported. Two are named: they are covered by IsoKron's published audits. The other four are given as one aggregate line, "4 further open-source MCP servers, not named pending notice to their maintainers", and descriptions of those write paths are omitted elsewhere in this folder. Records quoting those servers' ordinary error messages are included.

## The second coding (`verification/`)

An independent second coder coded all 110 records blind, from the trimmed text and `verification/codebook.md`, and every record was checked against its source.

- `second-coding.json` and `codebook.md` are unchanged since their checksums were taken (`blind-freeze.sha256`).
- `agreement.log` and `agreement-blind.json` give agreement and kappa per code; `disagreements.md` gives every disagreement and how it was resolved.
- `corpus-verified.json` carries the first code, the second code and the resolved code side by side for the 108 records that could be verified (B67 and B68 could not be read raw). 36 records carry a correction. `corpus.json` keeps the first coder's codes as they were.
- `VERIFICATION.md` is the write-up; `verify-results.json` holds the per-record source checks.

## Paths

Local paths are replaced with `$HOME`, `$WORKDIR` and `$TMPDIR`, except in hash-locked files (`3-design.md`, the five amendments, `harness/src/**/*.mjs`, and the two frozen verification files), which are published byte for byte.

## License and third-party material

**License.** Except where otherwise noted, this dataset, *Conflict-wording study: dataset* by IsoKron (Stormhaven Enterprises LLC), is licensed under the Creative Commons Attribution 4.0 International license (CC BY 4.0), https://creativecommons.org/licenses/by/4.0/. Attribute as: "*Conflict-wording study: dataset*, version 1.0, IsoKron, CC BY 4.0", with a link to https://isokron.ai/research/mcp-conflict-errors-ai-agent. The harness code (`harness/**`) is licensed under the Apache License 2.0; see "Licensing" below.

**Third-party material.** The quoted message strings in `corpus.json`, `2-corpus.md` and `verification/` are short excerpts from third-party source code and documentation. They are reproduced to identify and study them, they remain under their own licenses and copyrights, and they are not covered by the CC BY 4.0 license; each links to its source. Project and product names belong to their owners and are used only to identify what was studied. Text generated by AI models in the run records is included under CC BY 4.0 to the extent IsoKron holds any rights in it.

## Licensing

This archive mixes code and data, marked per file in `REUSE.toml` ([REUSE 3.3](https://reuse.software/spec-3.3/)); license texts are in `LICENSES/`.

| Path | License |
|---|---|
| `harness/**` (scripts, config, tests) | Apache-2.0 |
| everything else (design, amendments, run records, coding sheets, corpus, `verification/`, reports) | CC BY 4.0 |

Copyright 2026 Stormhaven Enterprises LLC (IsoKron). Hash-locked files carry no license header so that their published checksums still match; `REUSE.toml` applies. Quoted third-party strings and names are not covered; see "Third-party material" above.
