Use when deciding what evidence package a COLT (Conference on Learning Theory) paper needs, given that COLT runs no artifact-evaluation track or badges — the proof appendix is the artifact. Covers proof-verification passes, optional code companions for numerics, formalization aids, and post-acceptance release of scripts.
SkillScore83Maintained
Automatic
◆Research pack
Last verified 8/31/2026
What it changes
Passed 19 of 23 test cases with the skill, and 11 of 23 without it.
Without skill48%
With skill83%
Checks it passed that the model alone missed
Uses bracketed line tags in the output.
Uses bracketed tag [Blocking item] in the output.
Uses bracketed tag [Code companion] in the output.
and 5 more in the full report
10 fixed9 passed either way2 still failing2 broke
Pass-rate lift
+35 pp
Tokens
+39%
Agent turns
0%
Response time
-27%
Model gemini-3.6-flashMeasured 8/31/2026v119 of 23 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 10
Prompt
In our COLT submission, Corollary 1 is presented in the main text as a 2-line deduction from Theorem 2. However, Theorem 2 in the ledger is currently unverified. Evaluate this situation using the 5-block bracketed format.
What was checked, without and with the skill
Checked
Without skill
With skill
Flags Corollary 1 as dependent on an unverified blocker.
Passed
Passed
Uses bracketed tag [Blocking item] in the output.
Missed
Passed
case-21 · gemini-3.6-flash · 8/31/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md119 linessingle file
COLT Artifact Evaluation
First, the honest framing: COLT has no artifact-evaluation track, no artifact
badges, and no code-submission requirement. Nothing of the kind appears in the COLT
2026 Call for Papers or conference pages (checked 2026-07-08). Do not import
NeurIPS/ICML-style artifact checklists into a COLT plan; they answer questions this
venue does not ask. What replaces artifact review at COLT is the referee's line-by-line
reading of your proofs — so this skill treats the proof appendix as the artifact and
tells you how to make it verifiable.
What replaces the artifact track
At an artifact-track venue
The COLT counterpart
Who "evaluates" it
Code repository with run scripts
Complete proofs in the same PDF
Reviewers, line by line
Reproducibility checklist form
Explicit assumption and constant bookkeeping
Reviewers and the AC
Artifact badge
Community trust that the argument closes
Readers citing the bound
Dockerized environment
Self-contained lemma dependencies
Anyone re-deriving the result
Benchmark seeds and configs
Scripts for illustrative numerics, if any
Discretionary reader interest
The proof appendix as artifact
A COLT artifact plan is a verification plan:
Map every theorem to its full proof location and confirm no step says "similarly" or
"standard" where the step is neither.
Track constants explicitly through the proof chain; a constant that changes value
between Lemma 4 and Theorem 1 is the theory equivalent of a failing test.
List every external result invoked (concentration inequalities, minimax theorems,
reductions) with a precise citation including theorem number, and check that your
setting satisfies that result's hypotheses — the most common hidden bug.
Have a coauthor who did not write a proof verify it cold, logging time-to-verify.
Sections that take a fresh expert more than an hour per page need restructuring
before a time-pressed referee sees them.
A verification ledger you can actually run
Keep a machine-checkable inventory next to the LaTeX source: