lit-anchor
Use when assembling or adding to a bibliography, citing prior work, or making a novelty or absence claim — especially with AI assistance, which fabricates plausible-but-fake DOIs, titles, and authors. Governs DOI-verified, small-step literature anchoring so no unverified citation enters the record. Trigger phrases: "add a citation", "find references for", "cite prior work", "is this novel", "no one has done", "the literature says", "build the bibliography", "related work".
- License 许可
- MIT
- 最近更新
- 2026-09-01
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- lit-anchor
- tier
- extended
- description
- Use when assembling or adding to a bibliography, citing prior work, or making a novelty or absence claim — especially with AI assistance, which fabricates plausible-but-fake DOIs, titles, and authors. Governs DOI-verified, small-step literature anchoring so no unverified citation enters the record. Trigger phrases: "add a citation", "find references for", "cite prior work", "is this novel", "no one has done", "the literature says", "build the bibliography", "related work".
03 SKILL.md 正文
lit-anchor
Every citation is a claim that a specific work exists and says a specific thing. AI assistants produce plausible-but-fabricated references — real-looking DOIs that resolve to nothing, or to a different paper; correct-sounding titles with invented authors or years. A bulk "give me twenty references" response is a draft to verify, never a result. This skill is the discipline that keeps a fabricated citation out of your paper.
When to use
- Assembling or extending a bibliography / related-work section.
- Adding a single citation to support a claim.
- Making a novelty or absence claim ("no prior work does X", "we are the first to…").
- Any moment an AI hands you a formatted, confident-looking citation.
When NOT to use
- Summarizing or reasoning about a paper you already hold — that is reading, not sourcing.
- Internal cross-references to your own artifacts (use
layer-sync/decision-log).
Runbook
- Treat every AI-produced identifier as a hypothesis. A DOI, title, author list, year, or venue emitted by a model is unverified until resolved against an authoritative source.
- Verify one citation at a time. Resolve each DOI/URL against a real source (the DOI resolver, the publisher page, an indexed search) and confirm title + authors + year + venue all match. Do not batch — batching is where fabrications hide.
- Main agent synthesizes; sub-agents fetch small pieces. Never delegate "produce the verified bibliography" to a sub-agent that returns a finished, verified-looking table — that is exactly the pattern that launders fabrications. Sub-agents retrieve single documents; the main agent assembles and vouches.
- Mark unverified and abstract-only sources explicitly. An abstract-only source carries a documented "full text before submission" debt; an unresolved citation is flagged, never silently included.
- Back absence claims with a documented search. "No prior work does X" requires the actual queries run and their date recorded — not a model's confidence.
Rules
- No fabricated identifiers. Every DOI/URL in the record resolves to the exact claimed work.
- Bulk AI citation output is a draft. It enters the record only after per-item verification.
- Every citation carries provenance: where it was verified and the as-of date; trained knowledge about a paper is a hypothesis about the present, not an observation.
- Absence/novelty claims are search-backed, with the search recorded.
- A citation that will not resolve after a documented search is dropped or flagged — never quietly kept because it "looks right".
Configuration
{{bibliography}}— where verified citations live (e.g.docs/references.bib, a notes file).{{claim_log}}— optional: where novelty/absence claims and their backing searches are recorded.
Provenance & maintenance
Generalized from a working computational-research repository's DOI-verification protocol, adopted
after a bulk literature-review sub-agent was caught emitting fabricated references: the fix was
small-step manual retrieval with one-by-one resolution and main-agent synthesis, plus an
explicit abstract-only debt on sources not yet read in full. See the pack's examples/ directory
for the worked exemplar mapping.
Re-verify in your project:
- Resolve every identifier in
{{bibliography}}and confirm the resolved metadata matches the entry (a stale or redirected DOI is a silent corruption). - Confirm each novelty/absence claim in
{{claim_log}}still has its backing search recorded.
04 同仓库更多 Skills
skill-library-generator
Use when adopting the Aletheia discipline in a new or existing research repository — discovering the project's actual conventions, interviewing the maintainer to fill the config bindings, extracting the project's own recurring failure modes into project-local skills, and scaffolding the record surfaces (decision log, build log, evidence dirs) after approval. Trigger phrases: "adopt aletheia here", "set up the research discipline", "generate skills for this repo", "bind the pack to this project", "run the adoption interview".
statistical-reporting
Use when a numeric result is about to be reported, compared, or put in a paper — enforcing replication counts, uncertainty intervals, paired comparisons on matched instances, and visible denominators. Also use when designing how many runs a claim needs. Trigger phrases: "is this difference real", "how many replications", "report with confidence intervals", "single run result", "compare method A vs B", "is N=3 enough".
numerical-determinism
Use when the same code and same seed still produce different numbers — pinning thread counts, parallel reduction order, GPU kernels, and engine-level nondeterminism — and when a claim requires bit-reproducibility or a timing result requires hardware context. Also defines the hardware-context fields every run's meta file should record. Trigger phrases: "same seed different result", "pin the threads", "deterministic mode", "results differ across machines", "record the hardware", "timing needs the machine specs".
negative-results-ledger
Use when an idea, approach, or experiment fails, dead-ends, or gets abandoned — recording it in one line so it is never re-attempted from scratch and never silently buried. Also use when a negative or infeasibility result might itself be a finding, and when auditing a project for unrecorded dead ends. Trigger phrases: "that didn't work", "abandon this approach", "record the dead end", "why didn't we try X" , "we tried that before", "the result was negative".
external-positioning
Use when a claim is about to face the outside world — classifying what is novel versus known, scoping claims to what the evidence supports, holding the figure-regeneration bar (every kept figure rebuilds from one command), and holding the reproducibility bar before anything enters a draft. Trigger phrases: "is this novel", "position against the literature", "are we over-claiming", "regenerate the figure", "can we rebuild all figures", "ready for submission", "what's our contribution". (Citation verification lives in `lit-anchor`.)
run-provenance
Use when a kept run must be attributable to exactly what produced it — the environment it ran under (interpreter, locked packages, domain engines, system layer, native architecture), the input bytes it consumed (datasets, external files, config/constants), and the random draws it made (seeds recorded and re-derivable). Everything `results/<name>_<date>/meta.json` must record about a run, and how to make each field true. Also use before any campaign whose results will be kept, when two machines disagree on a result, and when asking "did the data change between these two runs". Trigger phrases: "pin the environment", "lockfile", "which versions produced this", "works on my machine", "environment drift", "hash the inputs", "did the data change", "which dataset version", "fingerprint the config", "data provenance", "record the seeds", "which seed produced this", "unseeded randomness", "make this run reproducible", "what produced this number".