.

# SUB-017 , PG Operational Flows (Sergey doc) , ingest + cross-check

> Source doc: Sergey, "PG Operational Flows", Confluence eng space, page 802881537. > URL: https://sett-ai.atlassian.net/wiki/spaces/eng/pages/802881537/PG+Operational+Flows > First cross-checked at v13 (2026-06-02) on 2026-06-03. Re-ingested at v18 (2026-06-04) on 2026-06-07 , see "v13->v18 delta" section at the bottom for everything Sergey added. > Method note in doc: 3 lenses , ~90 days #cosmo-support (182 PG threads -> 223 signals), live read-only Cosmo walkthrough 29 May 2026, existing specs + March UX review. Working files under product KB discovery/user-research/cosmo-pg-flows-2026/.

Doc covers 13 PG flows in 5 groups (A Onboarding/Ideation/Knowledge, B Review/QA, C Delivery, D Post-Delivery/Revision, E Observability) + Cross-cutting frictions + Open items.

---

1. CONFLICTS (our knowledge says X, Sergey says Y)

C1. Lifecycle status model , DIFFERENT vocabularies

• Sergey: Ideas -> Building -> In Review -> (PG Review) -> Ready -> Delivered -> Accepted | Revision. Also uses "Ready to Send", "Waiting for Customer", pg_review step.
• Vault (Playable-Schema.md, locked 2026-04-29): 13 canonical statuses , assigned, in-progress, pending-tl, pending-pg, waiting-for-fix, approved, stuck, build-queued, building, build-complete, build-failed, delivered, completed.
• Conflict: Sergey's terminal state is Accepted; vault's is Completed. Sergey's pre-client gate is PG Review / Ready to Send; vault has pending-pg -> approved. Sergey uses Waiting for Customer as a post-send state; vault collapses that into delivered. Note: vault explicitly says legacy stage_status enum (incl. pg_review) is back-compat only , "new work reads status." Sergey's doc still leans on pg_review. Either the live product diverged from the locked schema, or Sergey is describing the live UI labels (which may not match our canonical keys). Needs reconciliation , which is ground truth?

C2. Revision rate , numbers don't match

• Sergey: Revision Rate 27.9% / 28.1% / 28% (target 15%). Avg time to fix 9.6d (target 1.5d).
• Vault (Revisions.md): Revision Rate 20.4%, Avg Resolution 3.3d, Repeat 1.3%. Scorecard.md shows 21.2%.
• Conflict: ~28% vs ~20-21%. Either the metric climbed ~7-8 points since our April capture, or different denominators. Avg-fix 9.6d (Sergey) vs 3.3d (vault) is a 3x gap , big. Verify which is current. If real climb, that is a quality-trend story leadership should hear.

C3. Revision 6-type taxonomy , SHIPPED vs NEVER BUILT

• Sergey: "the 6-type list from the spec was never built, so it's free text only and revision reports can't break down by cause." (stated twice, flows 10 + 11).
• Vault (Revisions.md): documents the 6 types as the Sergey-era spec: Asset Mismatch, Broken/Technical, Design/Layout, UX/Tutorial, Copy/CTA, Cross-Game.
• Conflict: Our vault holds the 6-type spec; Sergey now confirms it was never implemented. Not a hard contradiction (spec existed, code didn't), but our vault reads as if the taxonomy is live. Update vault to flag taxonomy = spec-only, not shipped. Harel's top ask (fix-type catalogue) is therefore still unmet.

C4. Scorecard header numbers , different snapshot

• Sergey: Monthly Quota 530 | Delivered 430 (81.1%) | In Pipeline 1017 | Gap 66 | Daily Pace 14.8 vs 50.
• Vault (Scorecard.md): Monthly Quota 722 | Delivered 151 (20.9%) | In Pipeline 878 | Gap 186 | Daily Pace 10.8 vs 35.7.
• Not a true conflict , different month / different snapshot (ours April, Sergey late-May). Flagged so we don't treat either as static. Confirms Scorecard layout is stable; only data moved.

C5. Delivery tab counts , moved

• Sergey: Revisions sub-tabs In Progress 77 / Pending Allocation 8 / Dead Ends 31.
• Vault (Delivery.md): Ready to Send 51 | Revisions 181 | ... and the famous P29 cross-page mismatch (181 vs 104 vs 34 vs 244).
• Not a clean conflict , Sergey's 77/8/31 = 116 is yet another set of revision numbers. Reinforces P29 (revision count is unreliable across surfaces). The mismatch persists; add Sergey's figures as a 5th data point.

---

2. WEIRD / SUSPICIOUS (flagged by Sergey himself or looks off vs our model)

W1. Onboarding requires manual allocation or game is invisible

• Sergey 🛑: "In cosmo need to go to allocation and allocate the game to PG otherwise they won't see it ideation." Matches our P25 (filter-by-my-games missing) but is worse , it is a hard gate, not just a filter convenience. New.

W2. Pre-core games pollute delivery calculations

• Sergey 🛑: game entity created pre-core (often) shows on all systems, quota commitments show as junk data, "no way to treat a game as pre-core and ignore it." We have the pre-core tag in schema + P18 (pre-core games showing) , but Sergey says the junk-data-into-quota-math consequence is unhandled. Sharpens P18 into a data-integrity problem, not just a display problem.

W3. Manual "Add Prompt" idea is invalid without linked build_id

• Sergey 🛑 (flows 2,3,4): "A new idea is invalid without a linked build_id." Form lets you create a Version without a base build , "need to validate whether this makes any sense." We have NO build_id documentation in vault. New + suspicious (UI permits an invalid state).

W4. Rename inconsistency , zip name sticks, inner files revert after rebuild

• Sergey 🛑 (flow 7): "the ZIP name may stick while the files inside go back to the old name after a rebuild." This is the mechanism behind our SUB-015 Insight #10 (Klondike/Crazy Labs naming). Sergey gives the *cause*; we only had symptoms. Valuable.

W5. Send fails silently / Mark-as-Sent rows with no date

• Sergey 🤔: "Send fails silently fairly often (heard in #cosmo-support, not live-verified)" + Mark-as-Sent creating "Delivered rows with no delivery date." Matches our data-trust crisis; adds the no-date-row mechanism as a Scorecard corruption source. New mechanism detail.

W6. "Rebuild Without Sett AL Analytics" purpose finally explained

• Vault (Delivery.md) had: "No Sett AL Analytics" green badges "(purpose unclear)". Sergey explains it: AppLovin rejects playables with Sett AL analytics embedded; the rebuild strips them so the creative passes network review. Resolves a documented unknown. Move from weird->known.

W7. Quota set on Game page, not Weekly Quota Pools

• Sergey 🤔: "The quota is set on the Game page in Cosmo (not Weekly Quota Pools)... ❓ who sets or edits it isn't clear." Conflicts with our mental model from Demand-chain / Cosmo-Pains where weekly quota = monthly/4 auto-derived. Sergey says monthly quota lives on the Game page and is the root input. Worth confirming where the authoritative quota number lives.

W8. No confirm step before Send/Reject + no undo

• Sergey 🛑/🤔: "no confirm step before Send/Reject fires" and "a mistaken Send or Reject can't be reversed , the PG pings an engineer to move the task back." New, concrete, high-leverage quick win.

---

3. NET-NEW (not in our knowledge bank at all)

N1. The full system topology around Cosmo

Sergey names the surrounding systems and their roles , we had almost none of this:

• Atlas , idea generation UI (New Ideation, Create new ideas via prompt, up to 5/prompt, supports reference screenshots/videos, Review Queue with approve/reject/high-priority). "Slow because it loads containers (1-2 min)", PGs default to Cosmo.
• Apollo , where the build + review happens; PG review gate is in Apollo; revisions reopen in Apollo.
• ACK , holds Production Rules + Game Knowledge Navigation Index (read-only in Cosmo Game page).
• Logos , conversational layer; "can chat about the game's knowledge."
• Atlas (the registry) , game must exist in Atlas/Apollo/Cosmo.
• Monday.com , onboarding checklist board, game-documentation board, core-request board, "onboarding playables" board with signed-game automation that creates the game in the DB.

We had Atlas mentioned thinly (Demand-chain) and nothing on ACK/Logos/Monday topology.

N2. New-game onboarding step-by-step (flow 1)

Full manual sequence: Monday onboarding checklist -> mark signed (automation creates game in DB) -> Logos to create production rules (send naming conv, store links, trigger rules, languages, network; get github link) -> Monday game-documentation board (prod rules, assets, delivery drive) -> Cosmo Game screen -> different Monday for first core request -> Cosmo allocation to PG. Entirely new , we only had "onboarding is a pain point" (Open-Questions/PG.md).

N3. Concept-from-build / Helix core (flow 4)

• "Helix core" , brand-new term, not in vault. Concept creation points a task at a specific game build or Helix core; needs the source repo/core to have run the build pipeline + gotten Slack build approval. Can't create a concept on a game with zero prior playables. New.

N4. Fix menu , full item list (flow 7)

Per-card Fix▾ menu: Rename Zip & Files · Remove HTML Zips · Fix Slack Link · Rebuild · Rebuild Without Sett AL Analytics. We only had a partial list (Remove HTML zips, Rename, Rebuild). "Fix Slack Link" + the bulk Rebuild-without-AL on the Builds page are new.

N5. Send-to-Customer box detail (flow 8)

Exact UI: shows target channel (e.g. #sett-lion-studios-playables), sent as reply in original thread, preview reads "🎉 Your playable [name] is ready!/revised!" + Play now + Google drive links, editable name, Send/Cancel. Channel-less games block Send -> fallback ⚙ Copy delivery link shared by hand (Wrike/Drive/CRM). "Fix Slack Link only patches an existing channel; creating+connecting a new one needs an engineer." New detail.

N6. Mark as Sent semantics (flow 9)

Fix▾ -> Mark as Sent -> popup asks "delivery this month" vs "just a status update" , the choice decides whether it counts toward quota. New , explains a Scorecard-integrity lever we didn't know existed.

N7. AppLovin ad-network rejection flow + partner-ID matching (flow 12)

• AppLovin rejects -> Rebuild Without Sett AL Analytics. New flow.
• Partner-ID <-> SETT-ID matching done by hand outside Cosmo; no stored table; "❓ who owns the ID matching , PG, data team, or eng?" Entirely new gap.

N8. Knowledge-write gap (flow 5)

PG can SEE Game Knowledge (ACK) read-only in Cosmo but has "no clean way to add or upload knowledge inside Cosmo." Knowledge actually lives across Slack, briefs, drives. "❓ where does it land when writing exists , Cosmo Game page, ACK directly, or elsewhere?" New explicit gap.

N9. Cross-cutting friction mechanisms (the "why" behind our symptoms)

• Disappearing playables = failed builds, often the 5 MB limit. Eng quote: "Usually when playables disappear from the delivery page, it's failed in the building process" and "Apollo failed tasks are not presented in Cosmo." Eng added build-failure alerts + daily scan + cleared 82 abandoned builds , but no alert reaches the PG. We had P13/P6 (failed builds) but not the 5MB cause, the 82-build cleanup, or the "Apollo failed tasks invisible in Cosmo" root cause.
• Mis-filed playables: status flip from build event, wrong owner (e.g. "under Harel's name"), or delivery-name (Apollo) vs idea-name (Atlas) mismatch. The name-source split (Apollo delivery name vs Atlas idea name) is a new, precise root cause for search failures.
• "Report a bug" button is said to break, which pushes everything to #cosmo-support. New.

N10. Klondike + Simon's Cat named as strict-naming partners

Sergey names Simon's Cat (new to us) alongside Klondike. Lion Studios appears via the channel example #sett-lion-studios-playables. We had Klondike + Crazy Labs from SUB-015; Simon's Cat is new.

N11. Frequency / cadence data per flow

Sergey gives operational cadence we lacked: onboarding 1-2x/month per PG; ideation+curation weekly on Thursday for the following week (before Monday morning production start); Barzel-style "fresh concepts" rule already in our Demand-chain but Sergey adds the thursday->monday timing. Useful for any weekly-board design.

---

v13 -> v18 delta (re-ingested 2026-06-07)

Sergey expanded the doc from 13 flows to 12 (Observability renumbered 13->12; no flow dropped, the old "13" was just the observability flow). New material per flow, all now in vault:

Flow 1 , Onboarding (new frictions)

• 🛑 No structured way to ingest game briefs + assets into ACK at onboarding.
• 🛑🛑🛑 (Sergey's emphasis) It's possible to create ideas on games without an approved core, and the idea then takes a half-baked build_id on a WIP core. Sharpens W2/W3 , the pre-core hole and the build_id-validity hole are the same root problem at onboarding.

Flow 4 , Concept / Helix (substantially expanded)

• New ways to create a New Concept (NC): Batch prompt in Cosmo can create an NC; in Atlas, just mentioning "new concept" in chat creates one; for a new core, open a Monday ticket with a brief + assets prompt , max once a month.
• Helix friction rewritten: "Missing Helix at scale" , limits explore-at-scale capability. Today PGs talk to Shahaf to check Helix compatibility, and are blocked by an asset manager who sets up the assets for Helix. (Earlier v13 framing , "can't create concept on zero-playable game", "bad repo name needs a dev" , folded into this.)

Flow 5 , Partner knowledge (new frictions + idea)

• 🛑 New ACK knowledge does not apply to ideas generated before it landed , no invalidation/re-apply.
• 🛑 Saving is manual PG translation from client feedback; no automatic learning. Want auto/semi-auto save to ACK from a revision.
• 💡 Sergey's idea: on revision submission, auto-save knowledge or ask the PG whether to save.
• Cadence sharpened: knowledge updates happen "usually right after a revision brings new info."

Flow 6 , PG review (happy/unhappy split + new frictions)

• Apollo review now has two outcomes: Approve -> Ready to Send in Cosmo; Add a feedback -> becomes an internal fix request, goes back to the auditor (distinct from a client revision).
• Explicit happy vs unhappy flow: Approved -> build pipeline -> build success -> Ready to Send. Unhappy: Approved -> build fail/stall -> PG finds it manually on Dead Ends / Build page or by tracking -> escalates in #cosmo-support.
• 🛑 Extra steps to play on phone (💡 wants a Slack link to play on mobile).
• 🛑 Hard to find all the tasks waiting for PG Review in Apollo.

Flow 7 , QA (explicit checklist + Fix-menu annotations)

• Sergey's actual QA checklist: replay the playable (sometimes wrong; easier on mobile via the Cosmo link), check the zip folder structure, the naming convention, that all trigger versions exist (play a few as a sanity test), that the networks exist.
• Fix▾ menu annotated: Rename Zip & Files (Sergey: "why need this?" , questions its purpose); Remove HTML Zips; Fix Slack Link (put it in the right thread); Rebuild (fixes broken zips, AND a hack to return a playable to Ready-to-Send); Rebuild Without Sett AL Analytics (when a network like AppLovin rejects the build for SETT analytics).
• History tab: PG should see only PG + customer revisions there; used to confirm the fix was actually applied even though already reviewed in Apollo.
• Volume sharpened: QA runs "dozens of times per week".
• Display name is edited "in slack in the Send preview" (the editable name targets the Slack message).

Doc-structure / parent change

• Page parent moved (parentId 327799 -> 816906278) , Sergey reorganized where it sits in the eng space tree. Cosmetic, noted for traceability.

What did NOT change (still current, no edit needed)

• Lifecycle string, revision-rate numbers (still 27.9/28.1%, 9.6d), Scorecard snapshot (530/430/1017/66), revision sub-tab counts (77/8/31), all C1-C5 conflicts, Mark-as-Sent semantics, ad-network flow, cross-cutting frictions, the 5 open items. All as captured in the sections above.

New entities for the glossary (added 2026-06-07)

• Shahaf , person, gatekeeps Helix compatibility checks.
• Asset manager , role, sets up assets for Helix (a bottleneck on concept-at-scale).
• NC (New Concept) , the concept artifact; multiple creation paths (Cosmo batch prompt, Atlas chat, Monday ticket for a new core).
• WIP core / half-baked build_id , a core still in progress that an idea can wrongly bind to.
• Add a feedback (Apollo) , internal fix request back to auditor; distinct from a client revision.

---

Open items Sergey himself wants confirmed (carry into our Open-Questions/PG.md)

• Who sets the monthly quota on the Game page (PG vs TL vs eng).
• Whether revision-type taxonomy is planned to ship (analytics depend on it).
• Send silent-failure cause + channel-creation path.
• Exact approve controls in Apollo for the PG review gate (flow 6) , not yet swept.
• Exact Atlas-vs-Cosmo split for idea generation/curation.