Evidence ledger research
Skill markoblogo/abvx-agent-skills/skills/evidence-ledger-research
Reviewable capability layer for coding agents. Portable skills, delivery gates, workflow patterns, and verification-first engineering.
npx -y skills add markoblogo/abvx-agent-skills --skill evidence-ledger-researchAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Research and answer evidence-sensitive questions with source, date, unit, and operand discipline. Use for web research, PDF/document QA, policy or finance facts, tables, time series, calculations from retrieved evidence, exact-answer extraction, and any task where wrong source selection, stale facts, nearby table confusion, or formatting errors would matter.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
2.9 KB, as published. Nobody here has run it
Evidence Ledger Research
Use this skill when the answer must be tied to exact evidence rather than memory or broad summaries.
Source Discipline
- Prefer primary sources, official data pages, original PDFs, repository files, or provided/oracle documents.
- If a guessed exact-date search fails, broaden to the official series, table name, dataset code, or download page.
- Read only the narrow span needed after locating a promising source, then verify entity, period, basis, unit, and version.
- If a provided document contains the answer, extract from it before searching elsewhere.
- For volatile facts, verify the latest source before answering.
- If evaluating Doc-to-LoRA or another document-memory method, pair this with
doc-to-lora-evaluatorand keep the baseline questions, source answers, and negative controls in the ledger.
Evidence Ledger
Maintain a compact ledger for multi-step answers:
Claim/operand:
Source:
Date/period:
Unit/scale:
Role:
Value/span:
Notes:
Use the ledger before arithmetic, comparisons, ranking, or final formatting. Do not compute from partial evidence unless the user explicitly accepts an estimate.
Table And Document Rules
- Match row labels and exact column headers, not visual proximity alone.
- Watch for repeated month rows, fiscal vs calendar sections, continuation tables, estimates, totals, and footnotes.
- For forms and PDFs, locate the exact field/anchor in the question, then read the adjacent filled value from the same row, box, block, or table cell.
- For visual/document QA, prefer the smallest exact text span that answers the question.
- Preserve spelling, punctuation, capitalization, and numeric formatting when exact extraction matters.
Calculation Rules
- Confirm every operand before calculating.
- Track scale and direction: thousands/millions/billions, currency, exchange-rate orientation, sign, percent vs percentage points.
- Preserve relation direction: "change from A to B" means
B - A; "share accounted for by X" meansX / total. - For time windows, enumerate expected periods first and verify endpoint inclusion.
- Do not round intermediate values unless the source requires it.
Final Answer
- Match the requested format exactly: option label, numeric value, list, date, or concise phrase.
- Include unit words only when requested or necessary for clarity.
- When citing sources, link the source and summarize the evidence without long quotations.
- Copy the final answer from the checked ledger value, not from an unverified intermediate.