evidence-boundary-discipline
Skill3 starsChanged 53 days agoMITView on GitHub
State only what a dated source actually supports: fix the tense to the dateline, carry one provenance label across every surface, give every factual paragraph its own citation, and never let absence escalate into "does not exist". Use when writing or revising a knowledge page, monitor delta, entity section, or delivery table built from reports, press releases or partial legal text. Not for collecting or grading sources (that is claim auditing), and not for prose whose every claim is already primary-sourced.
In this repository59 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-harness-settings-paritySkill
- agent-operating-contract-evolutionSkill
- ai-coding-agent-delegationSkill
- arch-improvement-planSkill
- contract-bound-deliverablesSkill
- cross-agent-repair-routingSkill
- define-goalSkill
- delegated-work-acceptanceSkill
- design-an-interfaceSkill
- design-mdSkill
- discord-rich-messageSkill
- dotnet-repo-workflowSkill
- evidence-backed-analytical-writingSkill
- find-nearbySkill
- gddSkill
- generated-runtime-repairsSkill
- ghc-searchSkill
- git-auto-sync-pipelinesSkill
- github-code-reviewSkill
- github-repo-managementSkill
- grounded-citationsSkill
- harness-engineeringSkill
- hermes-context-file-loadingSkill
- hermes-modelsSkill
- hermes-updatesSkill
- hoa-introspect-distillSkill
- hoa-introspectSkill
- hoa-skill-reviewSkill
- hoa-skill-workSkill
- html-artifactSkill
- jupyter-live-kernelSkill
- kanban-workflowsSkill
- m365-copilot-cdpSkill
- measurement-campaign-evidenceSkill
- microsoft-officeSkill
- notes-to-prdSkill
- ocr-and-documentsSkill
- official-document-verificationSkill
- omp-headless-orchestrationSkill
- ops-claim-verificationSkill
- package-release-verificationSkill
- plan-to-issuesSkill
- project-phase-closeoutSkill
- report-claim-auditSkill
- research-stream-consolidationSkill
- review-revision-candidatesSkill
- rss-feedsSkill
- skill-library-curationSkill
- skill-porting-and-publishingSkill
- spec-convergenceSkill
- structural-validity-gatesSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- teams-cdpSkill
- tmux-agent-controlSkill
- whisperSkill
- writing-plansSkill
- writingSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 51
If the delivery table says a source was not re-read this round, the page cannot call the item fact with no locator.
- Distinctive: line 89
Run the checks with the project's own retired-qualifier list and page set:
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Evidence boundaries for claims built from reported sources
When to Use
- Writing or revising a knowledge page, monitor delta, entity/concept section, or delivery table whose claims rest on news reports, press releases, partial legal text, or sources read in an earlier round.
- A reviewer returned a page or candidate with "this overstates what the source says".
- For the mechanics of a rejected-candidate round (versioned patches, immutability, numstat reporting), pair this with the project's own revision-candidate procedure.
The rule underneath everything here: a page states what a named source said at a dated moment, and never what the world is like now unless a later artifact was actually read.
1. Fix the tense to the dateline
A source writing will be developed in the coming weeks records a plan as of that report. Days later it does
not license "still not formed", "not public yet", or "no draft exists". Write:
The report of
<date>records a pre-session drafting plan (the original uses the future tense); as of this cutoff no text was obtained and the current completion/public state is unknown.
A negative that was true when the source wrote it ("no checkable text before the session") is a statement about that moment, not a standing property of the artifact.
2. Hold one provenance label across all surfaces
| Surface | Carries |
| page body (bullet, table row) | fact / report / judgment / unknown |
| delivery or evidence table | same label, plus re-read-this-round vs attributed |
| runtime receipt / log | run summary only, never a research conclusion |
If the delivery table says a source was not re-read this round, the page cannot call the item fact with no
locator. Preferred fix: keep the Fact and tag the boundary in place — fact; carried from the first run's
in-page citation; not re-verified this round; attributed this round. Deleting previously verified facts is
not required; silently upgrading an unrechecked item to Fact is the defect.
Labels are the project's vocabulary (EVIDENCE_ATTRIBUTION_LABEL, EVIDENCE_RECHECK_LABEL); the parity
rule is universal even when the words differ.
3. Every new factual paragraph carries its own citation
- Full clickable URL, version/date, locator (section, paragraph, figure, line), short verbatim quote.
- A wikilink to a sister page, a media name, or a bare domain is not a citation; the backlink supplements.
- Sections appended to a concept or entity page are read standalone. Give them the same citation set even when the sibling query page already holds the full ledger.
4. Absence never escalates
| Observed | Allowed | Forbidden |
| searched this version, nothing found | "not found in the version read on <date>" | "does not exist", "no channel" |
| visible actions happened in forum A | "this round's visible action was in A" | "there is no pathway in B" |
| no text obtained yet | "not obtained this round; status unknown" | "unobservable until <event>" |
A monitoring window closes on its stated end date or an explicit retirement decision, never early because the evidence looks thin. When a later artifact might exist, say what you searched and when, not what cannot exist.
5. Read the definition before mapping an observation onto it
- A sub-criterion requiring sustained multi-year output, durable instruments, or shaped great-power conduct is neither satisfied nor refuted by a single announcement.
- An actor class defined as "moderate-material-weight, non-great-power" does not require the usual cohort; "the author is not the group this page normally covers" is therefore not a counterexample.
- A method label (
codification,implementation,pilot) does not prove equivalence with, or independence from, the prior arrangement. Name the level (bill section vs proposed code subsection) and say what was not compared.
6. Self-check before delivering
Run the checks with the project's own retired-qualifier list and page set:
# every banned qualifier the reviewer retired, including softened variants
grep -n -f "${EVIDENCE_QUALIFIER_FILE:?set the retired-qualifier list}" <candidate pages>
# each appended section must show its own external source
grep -c "https://" <appended-section-range>
# label parity: rows marked attributed in the receipt must appear tagged on the page
grep -n "${EVIDENCE_ATTRIBUTION_LABEL:-attributed}\|${EVIDENCE_RECHECK_LABEL:-not re-read}" <page>The qualifier file is content, not a hard-coded policy: when a reviewer retires a new phrase, append it there so the next round's grep catches the softened variant too.
Report as separate lines what was read this round, what is attributed from an earlier round, and what could not be obtained from this client/path/date. Reviewers accept the latter two when labelled and reject a flat list that hides which is which.
Configuration
| Key | Env | Default |
page_glob | EVIDENCE_PAGE_GLOB | none — the pages named by the delivery table |
receipt_path | EVIDENCE_RECEIPT_PATH | none — the project's delivery/evidence table |
qualifier_file | EVIDENCE_QUALIFIER_FILE | none — must be supplied per project |
attribution_label | EVIDENCE_ATTRIBUTION_LABEL | attributed |
recheck_label | EVIDENCE_RECHECK_LABEL | not re-read |
qualifier_pattern | EVIDENCE_QUALIFIER_PATTERN | none — prefer the file; a pattern is the fallback |
Resolution order and the config-file path are defined once in skill-porting-and-publishing § Configuration; this skill does not restate them.
Pitfalls
- Restating an older future-tense plan as present status; the invented observation is what gets rejected.
- Marking a claim Fact on the page while the delivery table calls it attributed.
- Appending an observation whose only citation is a sister-page wikilink and a media name.
- Treating one event as satisfying or refuting a sub-criterion that requires sustained output or a named actor class.
- Letting a run summary in a log or receipt carry a research conclusion that later rounds then inherit.
- Quoting the reviewer's rejected formulation inside a "not established" list without marking it as withdrawn, so a later grep cannot tell assertion from retrieval.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

