Skip to main content

Decisions

This register covers the supplied 2026-10-26/27 release decisions and their supporting work records. Decision approval is not deployment evidence. Initial questions and superseded answers are retained so changes in direction remain visible. Where rejection reasons are absent, this page says so rather than attributing a motive to Ron.

Blockquotes reproduce the recorded decision file. Only wording explicitly quoted by that file is claimed verbatim speech. A shared source line may contain several decisions; these are split into separate entries below.

1. Daily PR and mutation wait for Ron​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:3.

  1. Daily PR (UX.3, P2.18, UX.4) + full mutation run: only when Ron says so ("b").

Alternatives: Immediate daily PR was deferred by Ron.

Consequences: Do not infer release authorization from finished code.

2. PDF preview inside an app modal​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:4.

  1. P2.19 PDF preview: a window inside the app (Modal).

Alternatives: A separate browser window was not selected; the record only states modal.

Consequences: Preview remains inside the app.

3. PDF preview scope: Standards Library​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:5.

  1. P2.19 scope: (a) Standards Library only (clients never have the library and never upload to it).

Alternatives: Customer library access and uploads are excluded.

Consequences: Customer attachments retain their separate flow.

4. Headings on borders​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:6.

  1. UX.5 headings on borders: OK (one CSS fix + geometry rule).

Alternatives: Alternative reasons are not recorded.

Consequences: One CSS fix and a geometry rule are approved.

5. One shared Choose file button​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:7.

  1. P2.20 Choose file: (a) one shared file-picker button built once and reused everywhere; it looks exactly like every other button.

Alternatives: Separate picker styles were not selected.

Consequences: Reuse the shared control.

6. Stray and removed client-file cleanup​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:8.

  1. Stray / removed client files: (a) separate item after current work, both fixes (stray files after 1 day, removed files after the 1-year restore window). Ron: strict per-customer file separation required — follow-up pending (how a file is tied to a customer, enforced beyond the app).

Alternatives: Leaving stray files indefinitely is not the chosen design.

Consequences: Follow-up must cover one-day strays and the one-year restore window.

7. Staff edits held until Return​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:9-10.

  1. Staff edits on a Submitted questionnaire: B — held until Return; the client sees them only after Return to client; the staff admin gets clear feedback that the changes reach the client only when returned.

Alternatives: Immediate publication was not selected; the record calls the chosen option B.

Consequences: Customer sees held edits only after Return; staff must be told this.

8. Read-only restore investigation​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:11.

  1. Read-only look at the restore points behind Ron's #6 report: approved.

Alternatives: No write investigation is authorized by this entry.

Consequences: Inspect the reported case without changing it.

9. Changed-question restore explanation requested​

Date: 2026-10-26. Status: NOT VERIFIED. Source: A:.set/work/decisions-2026-10-26.md:12.

  1. Restoring an answer whose question later changed: Ron asked what it means — pending.

Alternatives: No decision in this initial entry.

Consequences: Superseded by the later answer at lines 35-36.

10. P2.21 phase explanation requested​

Date: 2026-10-26. Status: NOT VERIFIED. Source: A:.set/work/decisions-2026-10-26.md:13.

  1. P2.21 phases: Ron asked what it means — pending.

Alternatives: No decision in this initial entry.

Consequences: Superseded by delivery of both phases at line 37.

11. Defer AI upload scanning​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:14-15.

  1. AI expert scanning gate: not needed now (library uploads are staff-only; the chat accepts no uploads); P4.9 scanning stays in the pipeline for later. Risk acceptance to be written into the threat model.

Alternatives: Scanning now was declined because uploads are staff-only and chat has no uploads.

Consequences: P4.9 remains later; record the risk acceptance.

12. Expert reads one standard square​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:16-18.

  1. AI expert: an "AI expert" button in each standard square (e.g. IPL 13 Privacy). The model may read ONLY that square's files - never another square's (e.g. ISO 27001) - enforced by the app, the database AND the file server (S3), not by the prompt. Approach for v1: whole standard in the prompt (deputy recommendation, follows from Ron's description).

Alternatives: Prompt-only isolation is explicitly insufficient. Whole-standard input is recorded as the deputy recommendation.

Consequences: App, DB and S3 boundaries must all match the square.

13. EU-only AI processing​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:19.

  1. AI expert: EU-only processing — yes.

Alternatives: Other processing regions are not selected; rationale beyond this constraint is not recorded.

Consequences: Verify actual routing before calling the expert live.

14. Sonnet candidate with comparative evaluation​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:20-21.

  1. AI expert model: Claude Sonnet 5.5 (EU-only profile), to be confirmed by a ~50-question Sonnet-vs-Opus eval (half Hebrew) in the approved test; research: tmp/actions/model-research.md.

Alternatives: Opus remains the comparison model; it was not selected without evaluation.

Consequences: Model choice remains subject to the approved evaluation.

15. Measure spending before production caps​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:22.

  1. AI expert spend caps: learn the real spending first (from the test in 16), then set caps.

Alternatives: Guessing operating caps was not chosen.

Consequences: Evaluation limits do not establish production limits.

16. Initial small AI test approval​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:23.

  1. AI expert small test (model access + under $10 of test calls): OK.

Alternatives: Initial ceiling under $10; later wave-3 answers change the total budget.

Consequences: Read together with the later $20 total / $10 daily decision.

17. Friendly errors and reference IDs​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:24-25.

  1. Clearer upload error: NO. New permanent rule: never expose backend errors in the UI; show a friendly message plus an error reference id (UUID) that support can look up in the backend logs. To be written into CLAUDE.md + agent orders.

Alternatives: Displaying clearer backend errors was explicitly rejected.

Consequences: Support uses a reference; never expose backend error text.

18. Operator actions one at a time​

Date: 2026-10-26. Status: NOT VERIFIED. Source: A:.set/work/decisions-2026-10-26.md:26-27.

18-21. Ron's own actions (WorkOS roles, Cloudflare events, MFA S1/S2 + Cloudflare health checks, Grafana fix, rule-book PR #9, D:\tmp-fixbatch): step-by-step guidance, one at a time.

Alternatives: The record groups decisions 18-21 and does not map each number to a completed action.

Consequences: WorkOS, Cloudflare, MFA, Grafana and rule-book actions still need evidence of completion.

19. Strict per-client S3 sessions and keys​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:30-32.

  • 6b. Strict per-customer file separation: YES — keys tenants/<tenant>/clients/<client>/answers/<file>, per-request short-lived S3 credentials scoped to that client's prefix (session policy), server derives keys from the DB row, one-time move of existing objects.

Alternatives: Tenant-only key prefixes are replaced in the approved design.

Consequences: DB-derived client keys, short-lived sessions and a one-time move.

20. Per-square standards keys and no tools​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:33-34.

    1. Confirmed: standards keys per square platform/standards/<standard>/<file>; the AI expert gets DB text for that square only and an S3 session scoped to that square's prefix; no tools.

Alternatives: Other-square access and model tools are excluded.

Consequences: The expert receives only square-scoped text and storage access.

21. Removed and type-changed question restore​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:35-36.

    1. OK: a removed question's answer restores as history ("question removed"); a type-changed question's answer is refused with a message, the rest restore.

Alternatives: Blindly restoring incompatible values is refused.

Consequences: Removed questions restore as history; changed types are refused while the rest restore.

22. Deliver both P2.21 phases​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:37.

    1. Deliver ALL of P2.21 (phase 1 + phase 2) in today's delivery, not just phase 1.

Alternatives: Phase 1 only was rejected.

Consequences: Wave 1 includes the full approved lifecycle scope.

23. Wave 1 sign-off​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:39-40.

Sign-off (Ron, 2026-10-26): "1.A approved, 1.B approved, 1.C approved, 1.D Approved, 1.E Approved. 2.approved, 3. a" - Wave 1 approved: P2.19 PDF preview, UX.5 headings, P2.20 Choose file, P2.21 restore + submitted, UX.6 error references.

Alternatives: No alternative rationale is recorded in the sign-off.

Consequences: Approves P2.19, UX.5, P2.20, P2.21 and UX.6.

24. Threat-model section 14 approval​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:41.

  • Threat model phase-2-questionnaires §14 approved.

Alternatives: Alternative risk treatments are in the threat-model work, not this line.

Consequences: Approval is recorded; it does not verify deployment.

25. Block stale held edits before Return​

Date: 2026-10-26. Status: MERGED, NOT DEPLOYED. Source: A:.set/work/decisions-2026-10-26.md:42-43.

  • Held edits vs customer Reopen: (a) fix — a held edit on an answer the customer changed since is marked stale; Return is blocked until staff confirm or discard each marked edit (W1 fix lap 2). §14 residual 3 updated accordingly.

Alternatives: Silently overwriting the customer change is not accepted.

Consequences: Staff must confirm or discard each stale edit.

26. One PR after all waves​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:46.

  • PR: NOT now. The PR will include all waves and everything discussed today (UX.3, P2.18, UX.4, wave 1, wave 2, wave 3).

Alternatives: Immediate PR was declined.

Consequences: Include earlier UX work and all three waves.

27. Investigate restore through logs​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:47.

  • #6 restore check: there are only ~2 questionnaires on staging; follow the logs to find the case (read-only).

Alternatives: The record requests read-only investigation, not a repair.

Consequences: A report that staging has about two questionnaires is not accepted as live evidence here.

28. Licensed standards may be sent to AI​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:48.

  • Standards licence: Ron CONFIRMS Binary Networks is allowed to send licensed standards (ISO etc.) to an AI model for processing.

Alternatives: No competing option or rationale is recorded.

Consequences: Records Ron’s authorization; not an independent legal determination.

29. Combined release excludes virus scanning​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:50.

  • PR scope confirmed: one PR after wave 2 AND wave 3 are finished; it includes everything. No virus scanning (P4.9) in this PR.

Alternatives: P4.9 scanning is explicitly deferred.

Consequences: Wait for waves 2 and 3 before the combined PR.

30. Wave 2 session reuse​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: One session per operation was not selected.

Consequences: Reuse client sessions up to 15 minutes; user attribution further refines the cache in the design.

31. Wave 2 STS session policy​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: A broad role without the narrowing session policy is not the chosen design.

Consequences: Scope each issued session.

32. Wave 2 sweep in purge family​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: No separate scheduled family is selected.

Consequences: Sweep gets its own role.

33. MinIO for separation tests only​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: Persistent MinIO infrastructure was not authorized.

Consequences: Remove the test instance afterwards.

34. Move standards to per-square keys now​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: Deferring the standards move was not chosen.

Consequences: Unblocks the expert storage boundary.

35. Keep tombstone rows​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: Deleting all metadata rows is not selected.

Consequences: Preserve the tombstone lifecycle.

36. Deletion tripwire 200 objects / 10 percent​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: No numerical alternatives or rejection rationale are recorded here.

Consequences: Implement and review the exact population/approval semantics; do not assume it is live.

37. Research broker alternatives​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:52-56.

Wave 2 answers (Ron, 2026-10-26): 1 reuse a client's S3 session up to 15 min; 2 STS session policy: yes; 3 sweep in the purge family with its own role: yes; 4 MinIO allowed ONLY for the separation test, removed afterwards; 5 asked what "old copies" means (pending); 6 asked what "stray files" means here (pending); 7 standards move to per-square keys now: yes; 8 tombstone rows: yes; 9 tripwire 200 objects / 10%: yes; 10 OK, but research alternatives incl. a separate pass-issuing (broker) service and compare trade-offs (research running).

Alternatives: A separate pass-issuing service must be compared, not silently dismissed.

Consequences: The later Q10 choice settles the approach. Questions about old copies and strays are answered at line 58.

38. Move recovery and legacy strays​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:58.

Wave 2 answers cont. (Ron): 5 keep old copies 7 days + 30 days recoverable - ONLY for the one-time move; 6 stray legacy files found by the move: delete automatically (one-time, during the move).

Alternatives: Permanent old copies were not selected.

Consequences: Move-only seven-day soak plus 30-day recoverability; automatic one-time legacy stray deletion.

39. In-process broker plus four hardening controls​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:60.

Wave 2 Q10 (Ron): (a) option A (in-process STS + session policy) PLUS the four free hardening steps now (VPC-only bucket access, SourceIdentity = user id, many-clients alarm, honest residual in the threat model); per-tenant compute (F) later with P0.12.

Alternatives: Separate broker, presigned URLs, per-tenant KMS and access points were compared in P2.23; per-buyer compute is deferred to P0.12.

Consequences: VPC restriction, user SourceIdentity, many-clients alarm and an honest residual now; API compromise remains a boundary limit.

40. Expert formats​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:62.

Wave 3 answers (Ron): 1 text PDF + CSV + JSON now (DOCX/XLSX/scanned later); 2 300k tokens per question for now; 3 eval budget $20 total, $10/day hard; 4 AI expert disabled on inactive squares; 5 eval on ISO 27001 + IPL 13 Privacy with 50 Ron-approved questions; 6 Ron objects to reading right after upload (thinks it wastes tokens) - clarification pending (text extraction uses no AI tokens).

Alternatives: DOCX, XLSX and scanned files are deferred.

Consequences: Text PDF, CSV and JSON now.

41. Expert input budget​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:62.

Wave 3 answers (Ron): 1 text PDF + CSV + JSON now (DOCX/XLSX/scanned later); 2 300k tokens per question for now; 3 eval budget $20 total, $10/day hard; 4 AI expert disabled on inactive squares; 5 eval on ISO 27001 + IPL 13 Privacy with 50 Ron-approved questions; 6 Ron objects to reading right after upload (thinks it wastes tokens) - clarification pending (text extraction uses no AI tokens).

Alternatives: Alternative token budgets are not recorded.

Consequences: 300k tokens per question for now.

42. Evaluation budget​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:62.

Wave 3 answers (Ron): 1 text PDF + CSV + JSON now (DOCX/XLSX/scanned later); 2 300k tokens per question for now; 3 eval budget $20 total, $10/day hard; 4 AI expert disabled on inactive squares; 5 eval on ISO 27001 + IPL 13 Privacy with 50 Ron-approved questions; 6 Ron objects to reading right after upload (thinks it wastes tokens) - clarification pending (text extraction uses no AI tokens).

Alternatives: Supersedes the earlier under-$10 total test wording.

Consequences: Hard $20 total and $10/day evaluation limits.

43. Inactive square disables expert​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:62.

Wave 3 answers (Ron): 1 text PDF + CSV + JSON now (DOCX/XLSX/scanned later); 2 300k tokens per question for now; 3 eval budget $20 total, $10/day hard; 4 AI expert disabled on inactive squares; 5 eval on ISO 27001 + IPL 13 Privacy with 50 Ron-approved questions; 6 Ron objects to reading right after upload (thinks it wastes tokens) - clarification pending (text extraction uses no AI tokens).

Alternatives: Using an inactive square is not selected.

Consequences: Enforce active state at the API too.

44. Evaluation corpus and questions​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:62-63.

Wave 3 answers (Ron): 1 text PDF + CSV + JSON now (DOCX/XLSX/scanned later); 2 300k tokens per question for now; 3 eval budget $20 total, $10/day hard; 4 AI expert disabled on inactive squares; 5 eval on ISO 27001 + IPL 13 Privacy with 50 Ron-approved questions; 6 Ron objects to reading right after upload (thinks it wastes tokens) - clarification pending (text extraction uses no AI tokens). - Eval material (Ron): IPL 13 ONLY (no ISO 27001), using the files Ron uploaded to staging today (3 uploads succeeded at ~10:19 UTC into standard ae78a1f9-...; the 4th PDF failed). Q6 still pending (a/b).

Alternatives: ISO 27001 plus IPL 13 was superseded by IPL 13 only.

Consequences: 50 Ron-approved questions; the reported uploads are not independently live-verified here.

45. Extract text immediately after upload​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:64.

  • Wave 3 Q6 (Ron): (a) text extraction runs right after upload (no AI tokens; tokens only when a question is asked).

Alternatives: Delaying all reading until the question was considered after a token-cost concern.

Consequences: Extraction uses no AI tokens; model work starts when asked.

46. Alerts and health in this release​

Date: 2026-10-26. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:66.

Observability (Ron, 2026-10-26): add Grafana alerts for the new features (waves 1-3), for the risks in the threat models, and for overall environment health - so it is visible that everything runs and on-call does not need to develop or fix anything. Part of this release.

Alternatives: Leaving monitoring for later was not chosen.

Consequences: Cover new features, threat-model risks and overall health with operations-only responses.

47. Staging monitoring cost option B​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:68.

PS.14 + infra answers (Ron, 2026-10-27): 1 cost option (b) for staging (5-minute evaluation, ~$3-4/month), (a) for prod later;

Alternatives: Option A reserved for production later; stated cost is an estimate.

Consequences: Use five-minute staging evaluation.

48. Page Ron means Slack mention​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:69.

2 "page Ron" = Slack mention in staging; 3 expert kill switch = spend cap 0 (tested, Terraform variable); 4 6-hour WorkOS

Alternatives: No alternate paging mechanism is recorded.

Consequences: Use the staging Slack mention route once verified.

49. Expert kill switch is spend cap zero​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:69.

2 "page Ron" = Slack mention in staging; 3 expert kill switch = spend cap 0 (tested, Terraform variable); 4 6-hour WorkOS

Alternatives: A separate enable flag is not selected in PS.14.

Consequences: Require an enforcement test and a staging drill.

50. Role drift recheck every six hours​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:69-70.

2 "page Ron" = Slack mention in staging; 3 expert kill switch = spend cap 0 (tested, Terraform variable); 4 6-hour WorkOS role-drift recheck: yes; 5 ship thresholds, review after one week - REMINDER set for 2026-11-03; 6 WorkOS Radar: not now;

Alternatives: Boot-only check is extended.

Consequences: Observe configuration drift without waiting indefinitely for restart.

51. Ship thresholds and review after one week​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:70.

role-drift recheck: yes; 5 ship thresholds, review after one week - REMINDER set for 2026-11-03; 6 WorkOS Radar: not now;

Alternatives: Waiting for a perfect baseline was not selected.

Consequences: Review date recorded as 2026-11-03; reminder delivery is not verified here.

52. WorkOS Radar not now​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:70.

role-drift recheck: yes; 5 ship thresholds, review after one week - REMINDER set for 2026-11-03; 6 WorkOS Radar: not now;

Alternatives: Radar is deferred.

Consequences: Do not describe Radar as a current control.

53. VPC-only storage also restricts Ron​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:71.

7 VPC-only storage access (incl. Ron's console/CLI): OK; 8 AI spend budget alert email: ron.b@iontwrks.com; 9 allow the

Alternatives: Console/CLI bypass is not approved by this line.

Consequences: Recovery operations must account for endpoint restrictions.

54. AI budget alert email​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:71.

7 VPC-only storage access (incl. Ron's console/CLI): OK; 8 AI spend budget alert email: ron.b@iontwrks.com; 9 allow the

Alternatives: No alternative address is recorded.

Consequences: Use Ron’s recorded recipient; verify delivery without exposing credentials.

55. Extraction RunTask and PassRole exception​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:72.

ecs:RunTask + iam:PassRole line for the extraction task (waived semgrep rule, this one task family only): OK.

Alternatives: Exception is limited to this task family, not a blanket rule waiver.

Consequences: Keep the IAM scope and review evidence tied to extraction.

56. Documentation is permanently part of done​

Date: 2026-10-27. Status: APPROVED, IN BUILD. Source: A:.set/work/decisions-2026-10-26.md:74.

Documentation (Ron, 2026-10-27): everything must be documented all the time - architecture, decisions (with alternatives), mermaid diagrams - in openwiki and the SETDocs rule book (via PR), not only CLAUDE.md. Permanent rule: an item is not done until its docs and diagrams are updated; QA checks it.

Alternatives: Updating CLAUDE.md alone is insufficient.

Consequences: Update architecture, decisions, diagrams and rule-book docs; QA checks them.

Supporting work and unresolved details​

APPROVED, IN BUILD — P2.23 compares A in-process STS, B separate broker, C presigned URLs, D tag-based access, E per-tenant KMS, F per-tenant compute and G access points/Object Lambda. A was chosen with hardening; F remains LATER with P0.12. A separate broker only improves the boundary if it independently authenticates scope; keys or presigned URLs alone do not contain a compromised issuer. Cost/latency estimates in that research are not measured staging facts. A:.set/work/P2.23.md:426-544

NOT VERIFIED — the precise tripwire approval workflow and all original numbered 18-21 operator outcomes. The P2.23 design still labels questions at lines 398-400 open. Later approvals must be reconciled against the final merged implementation, not inferred from a general sign-off. A:.set/work/P2.23.md:396-400

Source conflict: the P2.22 status says building but the final-design header says nothing built. The supplied app tree contains neither expert routes nor its migrations; this register therefore uses APPROVED, IN BUILD, not merged or live. A:.set/work/P2.22.md:3-10

Later evidence update: waves 2 and 3 now have code on separate build branches, while the integrated source retains design files. Their status remains APPROVED, IN BUILD. The word “design” in older records must not be read as “no developer has written code anywhere.” E:evidence.md:17-18