numbers-traceable
Skill0 starsChanged 3 months agoMITView on GitHub
Every number in the paper must trace to a file on disk. Use while drafting or editing any sentence containing a measurement, an accuracy, a speedup, a p-value, or a table entry. Prevents the quietest and most damaging failure, which is a plausible number nobody produced.
In this repository14 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-cost-controlSkill
- citation-disciplineSkill
- delivery-contractSkill
- env-isolationSkill
- experiment-disciplineSkill
- figure-integritySkill
- latex-compileSkill
- literature-surveySkill
- page-fittingSkill
- paper-machineSkill
- paper-structureSkill
- research-integritySkill
- reviewer-mindsetSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 1 is distinctive, and 4 are 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.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
Every quantitative claim in the paper points at a specific value in a specific file under results/.
- Very specific: line 34
Copy exactly, then round once: If the file says 0.8234, write 0.82.
- Very specific: line 35
Never write 0.85 because it reads better.
- Distinctive: line 40
If you only have your own number, say only your own number.
- Very specific: line 53
If
results/is empty, there is no Experiments section: Do not write one. Describe the method, state that the evaluation is pending, and …
≈ “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
Numbers Must Trace
The rule
Before you write a number, open the file it came from and read it.
Every quantitative claim in the paper points at a specific value in a
specific file under results/. If you cannot point at the file, you cannot
write the number.
Why this is the dangerous one
A fake citation can be caught by a database lookup. A fake number cannot be caught by anything except the artifact that should have produced it. It survives compilation, review by a hurried reader, and often review by a careful one. It is the failure that ends careers.
Agents produce fake numbers without intending to. The mechanism is mundane: the draft needs a number, the result file is a few tool calls away, and the model completes the sentence with something reasonable. It is autocomplete, not deception, and the paper is equally wrong.
Practice
Read, do not recall. Even if you ran the experiment yourself in this same session, open the file again when you write the number. Your memory of the value is a generated guess.
Copy exactly, then round once. If the file says 0.8234, write 0.82.
Never write 0.85 because it reads better. Never round in the direction
that favours your method.
Never write a comparison you have not computed. "Our method improves over the baseline by 12 percent" requires both numbers and the arithmetic. If you only have your own number, say only your own number.
Leave an evidence trail. A LaTeX comment beside the number costs nothing and makes the claim auditable later:
Our method reaches 0.82 macro-F1
% [evidence] results/exp1_results.json -> metrics.f1_macro = 0.8234This turns "trust me" into something a reader, a reviewer, or a later verification pass can check mechanically.
If results/ is empty, there is no Experiments section. Do not write
one. Describe the method, state that the evaluation is pending, and stop.
An honest paper with no results is a draft. A paper with invented results
is fraud.
Self-check before delivery
Walk the paper and list every number. For each, name the file and the field it came from. Anything you cannot source is either removed or explicitly marked as an estimate in the text.
Report the result plainly: "14 numeric claims, all traced to results/" or
"12 traced, 2 removed because the runs did not complete". Both are fine.
Saying nothing is not.
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.