Skip to content

Decisions Evidence

Jump to bottom
xsyetopz edited this page Oct 13, 2026 · 2 revisions

Decisions: Evidence

This page holds the decisions D27, D29, and D30. Read it to learn why dotclaude takes facts from the host and the workspace, not from model reports. Each decision gives the problem, the choice, the alternatives, and the evidence. The index is Decisions. The decisions that 1.0.0 reverses are in Decisions-Reversed.

D27 Measured Facts over Model Reports

Problem: A model can report a false state. GitHub issue #35357 records a Claude model that printed context warnings with false values. The issue is not in this repository, so this page does not check it.

Choice: dotclaude does not take a state figure from the model. It reads the measured value from the host. The 1.0.0 parts of this choice are these:

  1. No prompt text asks the model to estimate or announce its context share. The texts in src/plugins/dotclaude/slices/prompt/texts.ts/ hold no such request.
  2. The context bound reads measured usage. triggerOf in src/plugins/dotclaude/slices/handoff/adapters/mod.ts reads $.session.usage(). CONTEXT_PERCENT in src/plugins/dotclaude/slices/handoff/limits.ts is the bound. Its value is 60. When the measured share reaches the bound, withHint adds a hint with the reason to the turn.complete output of the main session. onToolCall in the same file adds a report-now note to a tool result of a subagent that nears the bound. The bound and its default are in Context Bound.
  3. Compaction is off. The prompt slice sets DISABLE_COMPACT to 1 in src/plugins/dotclaude/slices/prompt/slice.ts. The transform in src/plugins/dotclaude/slices/prompt/transforms/context_management.ts replaces the compaction line. The new text says that the model does not need to wrap up early or hand off in the middle of a task.
  4. A claim of a passed check has no hook. The prompt rules ask for a run before a claim of done. The rules are in src/plugins/dotclaude/slices/prompt/texts.ts/verify-line.md and verification.md. A Stop prompt hook that judged such claims is reversed, see Decisions-Reversed.

Alternatives: A state figure that the model reports, which is rejected for the issue above. A Stop prompt hook that judges the claims of a reply, which is reversed.

Evidence: The code paths of the parts above. The test src/plugins/dotclaude/slices/prompt/lab/section.test.ts checks that DISABLE_COMPACT changes the context_management section.

D29 Semantic Line Breaks

Problem: The text that Claude writes to files and to commits must hold one sentence on each line. The model does not always keep this rule. This is an inference from the design, not a measured result in this repository.

Choice: dotclaude rewrites the text before it is written, with no model token. The code is in src/plugins/dotclaude/slices/sembr/.

  • The tool.call hook in slice.ts rewrites a Markdown Write or Edit. It also rewrites the quoted heredoc of a Bash call to git commit or gh pr create.
  • domain/rewrite.ts puts each sentence on its own line. domain/split.ts splits the sentences.
  • Edit rewrites an old_string that spans two sentences.
  • The hook has the phase rewrite. The generator orders this phase before the check phase of the guard slice (slice.ts, line 4).
  • The design uses no PostToolUse rewrite, because that can make the next Edit fail. This reason is the design text, and this page does not test it.
  • The chat reply is not rewritten. No code in src/plugins/dotclaude/slices/sembr/ or in the output styles rewrites it. See Semantic Line Breaks.

Alternatives: The markdownlint-rule-sembr package as a gate. The design rejects it for the runtime. It needs extra packages, and each block costs tokens.

Evidence: The rewrite test in src/plugins/dotclaude/slices/sembr/lab/sembr.test.ts. The source research/sembr.md is not in this repository, so this page does not check it.

D30 Hold Work That Is Right Under an Unsupported Claim

Problem: An agent can undo finished work that is right, after a claim of fault that it does not check. The claim can come from the user, from a subagent report, or from project context. The source is the paper arXiv 2609.32616. The paper is not in this repository, so this page does not check it. The paper tests Claude-Opus-5 and Claude-Sonnet-5. The mapping to Opus 5.5 and Sonnet 5.5 is inferred.

Choice: A claim of fault, from any source, does not undo finished work. The work stays until a check in the workspace shows the fault. The rule has three parts. The first two parts ship in 1.0.0.

  1. The evidence line. Its text is in src/fragments/evidence-line.md. It is in the shared actions section, src/plugins/dotclaude/slices/prompt/texts.ts/actions.md. It is also in the block of each variant, as EVIDENCE_LINE in src/plugins/dotclaude/slices/prompt/domain/variants.ts.
  2. The subagent rule. onSpawn in src/plugins/dotclaude/slices/prompt/adapters/mod.ts adds the block of the model of the subagent to the task prompt. A subagent gets no shared section, so its block carries the evidence line. The agent bodies in src/plugins/dotclaude/slices/agents/agents/ do not repeat the rule.
  3. The gate, which is not in 1.0.0. src/ has no UserPromptSubmit hook and no accusation flag. The gate is for the case that parts 1 and 2 fail the eval.

Alternatives: The gate of part 3, which is not built.

Evidence: The eval cases are these:

  • src/plugins/dotclaude/slices/prompt/evals/accusation-user/case.yaml, a false claim of fault from the user.
  • src/plugins/dotclaude/slices/prompt/evals/accusation-subagent/case.yaml, a false claim in a reviewer report.
  • src/plugins/dotclaude/slices/styles/evals/accusation-warranted/case.yaml, a true claim, where the agent must change the work.

The 0.28.0 eval suite recorded that the evidence rule came back. See Evals-0.28.0-Ported-Suite.

Open Items

  • The paper tests Claude-Opus-5 and Claude-Sonnet-5. The mapping to Opus 5.5 and Sonnet 5.5 is inferred, and Haiku 5.5 has no data.
  • The paper has no evidence-line text, no false-positive rate, and no token cost.
  • Rates in the paper are worst case, because its tasks hide the evidence. A real claim is often checkable. Then the right action is to check.
  • No eval result in this repository shows whether parts 1 and 2 pass on 1.0.0. So the gate of part 3 stays unbuilt.

See System Prompt.

Related Pages

Home | README | CHANGELOG | Issues

Clone this wiki locally