Evidence First Research

Evidence-first research

Answer the question, and make how you know it inseparable from what you found.

Why this exists. A report where "the gate takes 75 seconds, measured" and "the gate takes about five minutes, roughly" look identical is a report that will eventually be quoted wrong. That happened here: a cost claim of "3-5 minutes per push, about an hour across fifteen pushes" reached a changelog and a PR body. Measured, it was 75 seconds and nineteen minutes — off by four times, and caught only because someone re-derived it by hand. Nothing in the format made the unmeasured number look unmeasured.

So every finding is graded, the grades are a closed set, and the renderer refuses a record that grades badly rather than producing a page that looks finished.

The four grades

| Grade | Means | Needs | |---|---|---| | MEASURED | You ran it and read the result | The command, the machine, the conditions | | READ | You read it in a source that is authoritative for this | path:line, a spec section, a doc URL | | INFERRED | You reasoned to it from things above | Say what from — reasoning is not a citation | | UNVERIFIED | You did not check | Nothing. Saying so is the finding |

MEASURED and READ are refused without an **Evidence:** line. UNVERIFIED is deliberately free: demanding a citation for "nobody looked" pushes writers to leave the gap silent instead, which is the failure this is built to prevent. Full rules in references/provenance.md — read it when grading or when a grade is disputed.

The grade is about you, not the claim. A true fact you did not check is UNVERIFIED. A number someone else measured and you copied is READ, not MEASURED — and if you cannot cite where you read it, it is INFERRED.

Steps

1. Write the question down first, as one sentence, before looking anything up. A question that

changes shape mid-investigation produces findings that answer neither version.

1. Investigate. Prefer running something over reading about it; prefer reading the source over

reasoning about it. Every time you drop a level, that is the grade.

1. Write the record from references/report-template.md into the project's dated-work

directory — /work/YYYY-MM-DD-.md, per the canonical tree. Grade every finding as you write it, not in a pass at the end: a grading pass is where "I think I measured that" happens.

1. Render the view:

python3 .ai-badger/skills/evidence-first-research/scripts/render_report.py /work/YYYY-MM-DD-.md It writes to a temp directory and prints the path. It refuses a target inside the repository — the record is the artefact, and a committed HTML view becomes a second source of truth nobody can date.

1. Read your own ## Still open section before reporting. If it is empty, you either answered

everything or you stopped noticing. The second is more common.

1. Report the grade mix, not just the conclusion. "Four findings, one measured" is a different

answer from "four findings, all measured", and the person acting on it needs to know which.

Charts

Declare one in the record with a fenced block; the renderer turns it into inline SVG. Five kinds, and an unknown kind is refused rather than skipped:

| Kind | For | Body | |---|---|---| | provenance | The grade mix | Automatic — always drawn, never declared | | bars | Comparing measured values | label: value per line | | line | A quantity across a sequence | label: v1,v2,v3 | | matrix | Options against criteria | Header row, then label, v, v | | range | A measurement's interval | label: low..measured..high |

``chart:range title: push seconds by change shape docs-only: 12..14..19 full: 68..75..91 ``

Prefer range to bars for anything timed. A single bar asserts a precision three runs will not support, and the spread is usually the interesting part.

Gotchas

No environment-specific gotchas known.

Red flags — STOP

• A number in the record with no grade on its finding
• MEASURED on something you read rather than ran
• A benchmark without the machine and the conditions
• An empty ## Still open on a question that took more than an hour
• Rendering the HTML into the repository, or committing it
• Reporting the conclusion without the grade mix
• A bars chart of timings, where the spread was never recorded

Verification checklist

• [ ] The question in the record is the question that was asked
• [ ] Every MEASURED finding names the command and the conditions
• [ ] Every READ finding cites path:line or a spec section
• [ ] Every INFERRED finding says what it reasons from
• [ ] ## Still open is non-empty, or its emptiness is defended
• [ ] The renderer ran clean — it refuses badly graded records, so a clean run is a check
• [ ] The HTML is outside the repository and uncommitted

Files

• references/provenance.md — what each grade means, what disqualifies one, worked examples. Read it when grading a finding, or when a grade is disputed.
• references/report-template.md — the record shape the renderer parses. **Read it when

writing the record (step 3).**

• scripts/render_report.py — record → self-contained HTML. No network, inline SVG, no scripts.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Arasz
• Source: Arasz/ai-badger
• License: MIT
• Homepage: https://github.com/Arasz/ai-badger

Install and usage instructions live in the source repository linked above.