Source-Grounded Research Announcement
Use this method to turn a technical artifact into a public announcement without inflating what the artifact proves.
Non-Negotiables
- Do not publish or trigger a live publishing path unless the human has reviewed a successful dry run and explicitly approved live release.
- Treat every technical claim as unsupported until it has a source anchor: file path, theorem or definition label, declaration name, README section, repository object, release note, paper section, dataset card, benchmark report, or other inspectable artifact.
- Re-verify numeric claims immediately before live release. Counts, dates, benchmark values, model sizes, status labels, and repository totals must not rely on memory.
- Preserve proof and verification status exactly. Do not turn a template, target, route, negative result, witness inventory, prototype, draft, or stated objective into a solved result.
- Keep a visible claim ledger whenever the topic invites overreading.
- Remove generic AI and hype phrasing: "delve", "let's explore", "in conclusion", "fascinating", "journey", "game-changing", "revolutionary", "breakthrough", "changes everything", and similar filler.
Workflow
Collect source material before drafting.
- Ask for or locate the artifact, repo, paper, docs, logs, benchmark output, release notes, or source bank.
- Build a short source note with the artifact anchors, the draftable claims, the non-claims, open verification obligations, and any numeric claims.
- If the evidence is missing, draft only a source request or a narrower announcement.
Define the claim boundary.
- Write
Claimed:bullets for exactly what the source establishes. - Write
NOT claimed:bullets for tempting interpretations the source does not support. - Include proof status, implementation status, benchmark scope, and data scope when relevant.
- Write
Draft three angles when time permits.
- Artifact-led: start from the named source object.
- Constraint-led: start from the boundary, failure mode, proof obligation, or audit rule.
- Research-positioning-led: start from why this artifact matters within a larger line of work, while keeping the claim narrow.
Format for the target platform.
- For X threads, use 5 to 9 tweets by default.
- Separate tweets with lines containing only
---. - Do not use
1/,2/,(1), or numbered tweet prefixes. - Prefer one canonical link or artifact reference over link stacks.
- Avoid tag and hashtag stacks unless the human supplied them.
Attach anchors in drafting metadata.
- For each technical sentence, record the specific anchor that supports it.
- Keep anchors adjacent to the draft in reviewer notes, a table, or a manifest field.
- Mark numeric claims as
verify_before_post: trueuntil rechecked.
Review adversarially.
- BLOCK if any technical claim lacks an anchor.
- BLOCK if the copy implies a result, proof, verification, benchmark win, or production status stronger than the source supports.
- BLOCK live release if numeric claims have not just been re-verified.
- WARN if an anchor is too broad and should be narrowed before publishing.
- WARN if the post sounds promotional rather than artifact-led.
Run dry-run first.
- Render the post, manifest, or publish payload without posting.
- Show the human the exact final copy, source anchors, verification obligations, and reviewer findings.
- Accept only explicit human approval for live release after dry-run success.
Claim Ledger Template
Claimed:
- [narrow claim] - anchor: [artifact path, section, declaration, report, or link]
- [narrow claim] - anchor: [artifact path, section, declaration, report, or link]
NOT claimed:
- [tempting stronger interpretation that is not supported]
- [proof, benchmark, production, or scope claim the artifact does not establish]
Verify before live release:
- [numeric or status claim requiring fresh check]
Reviewer Checklist
- Hook names a concrete artifact, boundary, theorem or definition label, source file, repo, release, or precise result shape.
- The thread or post moves from artifact to claim shape to boundary to significance, without repeated filler.
- Every technical claim maps to an inspectable artifact anchor.
- Non-claims are explicit when the subject is likely to be overread.
- Numeric claims have fresh evidence.
- The final wording avoids hype, generic AI phrases, engagement bait, and unsupported institutional claims.
- The publishing path is still dry-run unless the human has approved live release after seeing the final payload.