uai-workflow

UAI Workflow

Use this as the project-management layer over the other UAI skills. UAI is a conference run by the AUAI: there is no editor-in-chief and no rolling submission — one deadline per year, per-edition chairs, and PMLR publication funded by registration rather than author fees. The 2026 calendar below is the verified anchor (checked 2026-07-08); substitute the current year's official dates before planning anything.

The 2026 shape of a UAI year

Backward plan from the paper deadline

Offsets assume a probabilistic-methods paper with theory and experiments; shift for your mix, and re-anchor to current official dates:

• T−10 weeks: contribution sentence frozen; identifiability/guarantee target named; ladder of experiments designed (exact-truth, stress, real).
• T−8: core theorem believed and sketched; assumption set labeled A1…An and no longer moving.
• T−6: exact-truth experiments done; the theorem's prediction visibly matches or the paper's story changes now, not at T−1.
• T−4: full draft in the UAI template; appendix skeleton mirrors main-text numbering; baselines tuned and logged.
• T−3: internal mock review against the four criteria (correctness, novelty, backing, clarity) by someone outside the project.
• T−2: stress and real-data experiments locked; determinism ledger and manifests in the artifact; supplementary ZIP assembled from a clean export.
• T−1: anonymity sweep of PDF and ZIP; page/size caps verified mechanically; OpenReview form fields drafted.
• T−0: upload early in the day AoE-wise; re-download and cold-read.

Role assignments that prevent the classic failures

# uai-cycle-owners.yaml — one name per row, no shared ownership
anonymity_sweep:      # PDF + ZIP + metadata; leaks are unfixable after the deadline
evidence_map:         # every claim -> exhibit; updated when any result changes
appendix_integrity:   # body statements == appendix statements after every edit
artifact_build:       # clean-export ZIP, manifest files, size cap
openreview_form:      # profiles, conflicts, subject areas, abstract == PDF abstract
response_marshal:     # archives submitted PDF; drafts objection bank before reviews
camera_ready:         # PMLR forms, de-anonymization, public repo tag
presentation:         # poster, spotlight, registration, travel/visa timeline

The two most damaging single points of failure at UAI specifically: an unowned anonymity sweep (appendices inside the reviewed PDF widen the leak surface), and an unowned appendix-integrity check (deadline-week edits to body theorems silently orphan their proofs).

Between decision and conference

• Accepted: the June-to-July window carries camera-ready, artifact publication, and registration simultaneously — assign them to different people on day one.
• Rejected: run the axis triage from uai-review-process within a week while context is fresh; the natural re-route targets (AISTATS, NeurIPS/ICML, CLeaR, or a journal) have their own calendars, and reviews in hand are an asset that depreciates.
• Either way: at least one author may be called to review for UAI — the volunteer agreement signed at submission — so protect reviewing bandwidth in March–April.

The cycle archive

One deadline per year means institutional memory is the compounding asset. At each transition, file into a shared, dated directory:

• The exact uploaded PDF and ZIP (deadline day) — responses must quote these.
• The OpenReview form contents as submitted (screenshot or export).
• All reviews, your responses, and the meta-review (decision day).
• The camera-ready as published, plus the PMLR entry once the volume is live.
• A one-page retrospective: which criterion drew fire, which anticipated objection materialized, what the next cycle should do differently.

Groups that skip the retrospective re-learn the same reviewer objection annually.

Contingency table

First-cycle notes

For a group's first UAI attempt: budget one extra week for OpenReview profile setup and conflict entry across all authors; expect the two-column template to shrink an apparently-fitting draft by half a page; and have the mock reviewer score the four posted criteria explicitly — first-time submissions most often lose on backing, not on correctness.

Standing warnings

• Never inherit dates from this file or from memory; the official important-dates page is the only calendar.
• Registration fees, in-person requirements, and remote-presentation options were not fully verifiable for 2026 at pack-check time (待核实) — confirm before booking travel or promising remote attendance.
• One-deadline venues punish schedule slips brutally: a week late here is a year late.

Output format

[Cycle position] fit / build / submitted / response / decided / camera-ready / conference
[Next hard date] <official date + source URL>
[Critical path] <three tasks gating the next transition>
[Owner gaps] <rows in the owner map without a name>
[Re-route plan] <standing answer to "if rejected, then where">