This page tells how the ground slice of the core plugin claudetto checks the claims of a plan file. A user reads it to learn why Claude gets a deny on a plan, and how to fix it. A contributor reads it before an edit of slices/ground/.

Claude can write a number, a file:line reference, or a convention with no source. An example is npx in a project that has bun.lock. The ground slice finds such claims in a plan file and denies the write. The check is a program. It makes no model call, because a check that guesses can block correct text.

The slice adds one tool.call hook on Write and Edit. The hook checks only these files:

• A file under openspec/changes/.
• A file named proposal.md, design.md, or tasks.md.

For Write, the hook reads content. For Edit, it reads new_string. A write of any other file is not checked. A file named *.test.* and a file in a fixtures folder are not checked, also under openspec/changes/.

A claim has a source when it is in one of these:

• A user message of the session.
• An answer of AskUserQuestion, which counts as a user message.
• The text of the handoff, which counts as a user message in any role.
• A tool result of the session.
• CLAUDE.md in the project root.
• The old text of the file that the call changes.

The sources are the messages of the agent that makes the call. For a subagent, the hook reads the transcript of that subagent with $.session.messages({ agentId }). A tool result of the subagent counts for that subagent only. If the session cannot read the transcript, the hook skips the check and the call runs.

A tool result that is a deny of this check is not a source. It names each claim, so a retry would pass on it.

Kind	The check fails when
Number	A number of 3 digits or more is in no user message and no tool result. A year, a number in the name of a standard (ISO 8601, RFC 3339), and a number that the text computes do not count. The text computes a number when it is a binary size in bytes (16 MiB and 16,777,216), a total of a table column (three rows or more) or the difference of two cells of the same row. A source can write the number with _, as 1_000_000.
file:line	The file is found and has fewer lines than the reference.
Attribution	A sentence such as the user chose ... has words that are mostly not in a user message.
Done claim	The text says that tests, checks, lint, or a build pass, or that something is fixed, and the turn has no Bash call. A sentence with must, should, will, if, or a similar word is not a claim.
Convention	The text uses a command of npm or npx (not the word npm in a sentence) in a project with bun.lock or bun.lockb, or pip in a project with uv.lock, and no user rule names the tool.

For a file:line reference, the check looks for the file at the project root first. Then it searches the folders below the root, but it skips node_modules and dot folders and reads at most 2,000 folders. If the check does not find the file, it skips the reference, because a guess would deny a correct one.

• A line with (proposed) or [NEEDS CLARIFICATION: ...].
• A fenced code block.
• A number in a #### Scenario: block of a spec.md file. The block ends at the next heading of level 4 or above. A number there is an input of a case and not a claim.
• An Edit of a spec.md file whose old_string is in a scenario block.

The deny lists each claim and gives three ways out:

1. Name the source: a user message, a tool result of this session, or a file of the project. When a local command with no model call or login can show the claim, run it, and name its result.
2. Mark the claim (proposed) or [NEEDS CLARIFICATION: ...].
3. Ask the user.

Claude then writes the text again.

The slice has no userConfig key and no limit. The hook costs no request and no token. If the hook throws, Claude Code logs the failure and the call runs.

A replay ran the check of the hook on the tool calls of recorded sessions. It made no model call. The replay script is not in 0.3.0, and the shipped plugin has no replay entry point.

• The replay ran the source check on one recorded PreToolUse call at a time.
• The replay read the files of the project as they were on the day of the run. A file:line reference resolved as a path suffix under the cwd, and a reference that did not resolve passed.
• The rate is the denies divided by the events checked. The events checked are the writes of a plan or a spec file.

The run of 2026-10-21 used 1,014 transcript files (952 MB) from ~/.claude/projects. The transcripts now hold plan edits, so the rate is measured. Each deny was graded by hand: a true deny is a claim with no source in the session or the project, and a false deny is a claim that has a source or is no claim.

Run	Calls replayed	Events checked	Denies	True	False	False rate
Before the fixes	25,331	53	9	1	8	15 percent (8 of 53)
After the fixes	25,436	53	1	1	0	0 percent (0 of 53)

The calls differ because live sessions grew between the runs. The denies before the fixes, by cause:

• 8601, 3339, 7763 (4 denies): the number is in the name of a standard (ISO 8601, RFC 3339, RFC 7763).
• 1,000,000 (3 denies): the project writes the number as 1_000_000, and the check matched only 1,000,000 and 1000000.
• 892, 1,005, 178, 287, 47,305, 46,388, 917 (1 deny): the numbers are column totals and differences in a table of line counts.
• 16,777,216 (1 deny): the text says 16 MiB (16,777,216 bytes), and the project states 16 * 1024 * 1024.
• npm in a project with bun.lock (1 deny): the text says "an npm marketplace", which is a name of a Claude Code feature and no command.
• 140 (1 deny, true): "140 pages" of a wiki folder, and the session listed no count.

The 5 fixes are in slices/ground/domain/claims.ts, and each has a test in tests/claudetto/ground/claims.test.ts. A fix does not accept a guess: a binary size needs the exact product, a table total needs the exact sum of the column, and a difference needs two cells of the same row. The sample is 53 events from 5 projects, so the rate has a wide margin. A false deny costs one rewrite, because the deny names each claim and gives three ways out.

• Spec
• Finish
• Controls