colt-workflow

COLT Workflow

Use this as the project-management layer for a COLT paper. Anchor everything to the

live timetable: for the 39th edition the CFP (checked 2026-07-08) set one submission

deadline — February 4, 2026, Anywhere on Earth, via CMT — for a conference running

June 29 to July 3, 2026 in San Diego. Notification and camera-ready dates were not on

the checked pages (待核实); take them from the CFP and acceptance email of your cycle.

Stage map

• Fit decision: confirm the contribution is a theorem a learning theorist wants

(see colt-topic-selection); if the result is 80% proved, the clock question is

whether the remaining 20% is verification or invention — only the first is

schedulable.

• Result freeze: theorem statements stop moving. At COLT this is the single most

important date you control, because every downstream artifact (proof appendix,

overviews, comparisons) quotes the statements.

• Verification: independent re-derivation of every load-bearing proof

(colt-reproducibility). Schedule it as real work with an owner per theorem — it

finds bugs at a rate authors never want to believe.

• Writing: body-first is a mistake here; write the appendix proofs first, then

distill overviews into the body, so the body never promises what the appendix

cannot deliver.

• Submission week: CMT record early, anonymity sweep, single-PDF assembly

(colt-submission).

• Rebuttal: reviews arrive before decisions; block calendar capacity for the

window even before you know its dates (colt-author-response).

• Aftermath: acceptance → PMLR camera-ready and talk prep (colt-camera-ready);

rejection → triage between fix-and-resubmit (next COLT / ALT) and re-routing.

Backward plan from the deadline

| Weeks before deadline | Milestone | Failure smell |

|---|---|---|

| 10+ | All main results proved at whiteboard level | "The lower bound should follow" — it does not, yet |

| 8 | Theorem statements frozen; ledger of claims opened | Statements still absorbing scope changes |

| 6 | Full proofs written in LaTeX, appendix-structured | Proofs alive only in notebooks and heads |

| 4 | Independent verification pass complete; bugs triaged | "We'll verify while writing" (you will not) |

| 3 | Body drafted: setup, results, proof overviews, known-vs-new table | Overview promises diverging from actual proofs |

| 2 | Internal mock review by an uninvolved theorist | No one external has read the model section |

| 1 | Anonymity sweep, notation audit, cross-reference check | Renumbering theorems this late |

| 0 | Upload, re-download, verify the stored PDF; stop editing | Compiling at the AoE boundary |

Weeks are heuristics; the deadline is not. AoE means when it is February 4 anywhere

on Earth — compute your local cutoff and subtract a safety day.

Verification is the critical path

For a theory paper the long pole is never the writing; it is discovering at week 2

that Lemma 6 is false. Mitigations:

• Verify in dependency order, most-load-bearing first — a bug in the master lemma

invalidates weeks of downstream polishing.

• When a bug appears, immediately re-derive the weakest still-true statement and

check whether the paper survives with it; many COLT papers ship the weaker constant

that week-3 debugging forced.

• Maintain the claim ledger (see colt-artifact-evaluation) as the single source of

verification status; standups read the ledger, not vibes.

Team coordination template

# colt-plan.txt
DEADLINE (AoE):        2026-02-04  -> local cutoff 2026-02-05 11:59 UTC-12
RESULT FREEZE:         2025-12-10  owner: all
PROOF APPENDIX DRAFT:  2025-12-22  owner: A (Thm 1), B (Thm 2, lower bound)
VERIFICATION PASS:     2026-01-09  owner: swap (B checks Thm 1, A checks Thm 2)
BODY + OVERVIEWS:      2026-01-16  owner: A
MOCK REVIEW:           2026-01-21  reader: C (not a coauthor)
ANONYMITY + PDF QA:    2026-01-28  owner: B
UPLOAD TARGET:         2026-02-02  (two days of slack, CMT record made 01-25)
REBUTTAL WINDOW:       dates TBA with reviews -> block capacity on release

Workflow failure modes specific to theory papers

• The moving theorem: a coauthor strengthens a statement during the writing

phase without re-opening its ledger entry; the appendix now proves the old claim.

Rule: statement edits after freeze require a ledger re-verification, no exceptions.

• The optimistic lower bound: upper bound done, lower bound "clearly similar to

[X]" and scheduled for week 2. Lower bounds are inventions, not chores — either

prove it before freeze or descope the paper to the upper bound honestly.

• The single-verifier trap: the only person who checked Theorem 2 is the person

who proved it. Swap verification assignments explicitly (the plan file above does).

• The deadline compile: first full single-PDF build in the final 48 hours,

discovering the restatement macros and cross-references were never wired. Build the

full document skeleton — body, appendix, references — by week 3 even while proofs

are placeholders.

After the decision

• Accepted: camera-ready per instructions; book San Diego-style summer logistics

early (single-track conference, one venue hotel); decide who talks.

• Rejected with deep reviews: the reviews are a free verification report — patch and

target the next cycle or ALT (deadline typically in autumn; verify).

• Rejected for significance: the mathematics may be sound but mis-framed; rework the

motivation and known-vs-new ledger before blaming the result.

• Either way, reconcile the arXiv version with what review taught you.
• Debrief the ledger: which claims did reviewers actually verify, and did your

internal verification pass catch what they caught? Calibrating that gap improves

the next paper's schedule more than any template.

Cycle-volatility warnings

• Every date above except the verified 2026 deadline and conference dates is

cycle-specific; notification, rebuttal window, camera-ready, and registration

timing are 待核实 until announced.

• Whether an Open Problems track (with its own deadline) runs in your cycle: check

the CFP; the 2025 edition ran one with a later, non-anonymous, 4-page format.

Output format

[Stage] proving / freeze / verification / writing / submitted / rebuttal / decision
[Critical path] <the unverified claim or unwritten proof gating everything>
[Calendar] <next official date + source, or 待核实>
[Owner map] <theorem/task -> person>
[Slack] <days between planned upload and AoE cutoff>