Brainery

compliance moat

Every skill load, every decision, every change — sourced.

When agents act on company policy, somebody has to be accountable. Every event below is append-only. Every decision cites the exact skill version + sources used. Policy changes route to the right human for sign-off before agents execute the new version.

Audit entries

12

Agent decisions

3

Pending approvals

2

Avg decision

218ms

approval queue (2)awaiting sign-off

PR-2118

refund-decision-policy.skill.md

v1.2.0 → v1.3.0-DRAFT

routes to

@rachel · CX lead

requested by

brain.coherence (auto-generated)

Raise auto-approve cap from $300 → $500 (matches CX practice)

Coherence engine flagged contradiction: HR policy doc says $300, three #cx-escalations threads (#44021, #45102, #45891) show CX leads approving up to $500 routinely. Brain proposes promoting observed practice to policy.

⚠ blocks skill until resolved · linked: refund-cap-conflict
⊟ view diff
--- a/refund-decision-policy.skill.md (v1.2.0)
+++ b/refund-decision-policy.skill.md (v1.3.0-DRAFT)
  ## Auto-approve thresholds
  
- auto_approve_cap_usd: 300
+ auto_approve_cap_usd: 500
  scope: non-enterprise
  tier_window_days:
    standard: 30
    enterprise: 90  # CEO email 2026-04-22
  
  ## Provenance
- source: notion://policies/refund-v3 (2026-02-12)
+ sources:
+   - notion://policies/refund-v3 (2026-02-12, doc:medium)
+   - slack://#cx-escalations/#44021 (cx-lead, 2026-04-08)
+   - slack://#cx-escalations/#45102 (cx-lead, 2026-04-19)
+   - slack://#cx-escalations/#45891 (cx-lead, 2026-05-02)

PR-2119

pricing-exception-approval.skill.md

v1.0.0 → v1.0.1-DRAFT

routes to

@maria · CRO

requested by

brain.coherence (auto-generated)

Rephrase '25% cap' → '25% CRO sign-off threshold'

@maria (CRO) approved 30% for Quantum Logistics on 2026-03-25. Skill body reads as if 25% is a hard ceiling; replacing wording with the threshold semantics it actually has.

⚠ blocks skill until resolved · linked: pricing-cap-quantum
⊟ view diff
--- a/pricing-exception-approval.skill.md (v1.0.0)
+++ b/pricing-exception-approval.skill.md (v1.0.1-DRAFT)
  ## Decision tree
  
  if discount_pct <= 15: auto-approve
  if discount_pct <= 25: route to deal-desk
- if discount_pct >  25: REJECT (hard cap)
+ if discount_pct >  25: route to CRO sign-off
+   # CRO can authorize larger exceptions;
+   # 25% is a routing threshold, not a ceiling.
  
  ## Precedent
+ - 2026-03-25: @maria approved 30% for Quantum Logistics
+   source: slack://#deal-desk/2026-03-25
decision audit log · last 28h
2026-05-04 21:34:12 UTCdecisionsupport-bot-prodrefund-decision-policy · v1.2.0184ms

Approved $40 refund for customer 12345 (standard tier, 14d).

APPROVED · auto · matched 'tier=standard, days_since=14, amount<$50'

↳notion://policies/refund-v3 (doc:medium)
↳slack://cs-policy/2026-04-22 (ceo-email:high)
2026-05-04 21:33:58 UTCdecisionsupport-bot-prodrefund-decision-policy · v1.2.0226ms

Routed $340 refund (customer 9001) to @rachel for sign-off.

ROUTED · enterprise tier, amount > $300 auto-approve cap

↳zendesk://ticket/9001
2026-05-04 21:33:41 UTCdecisionsupport-bot-prodrefund-decision-policy · v1.3.0-DRAFT312ms

Refused $1,200 refund (customer 7788). Sources contradict — escalated.

ESCALATED · agent stopped instead of guessing

↳notion://contracts/acme-msa-v2 (doc:high)
↳slack://cs-policy/2026-05-04 (lisa-edit:medium · 2hr ago)
2026-05-04 21:32:09 UTClint findingbrain.coherence

New contradiction detected on `enterprise.refund_window_days`.

Filed as approval request PR-2118; blocks refund-decision-policy until resolved.

↳notion://policies/refund-v3 → 30 days (doc:medium)
↳slack://cs-policy/2026-04-22 → 90 days (ceo-email:high)
2026-05-04 21:30:55 UTCskill loadoncall-botauth-jwks-rotation-spike · v1.1.084ms

Loaded auth-jwks-rotation-spike.skill.md after 401 spike crossed 10× baseline.

2026-05-04 21:28:14 UTCapproval@mayadb-connection-pool-exhausted · v1.0.0 → 1.1.0

Approved skill change to db-connection-pool-exhausted.skill.md.

2026-05-04 21:25:02 UTCingestbrain.ingest

Ingested 23 messages from #incidents · 4 entities created · 1 skill updated.

↳slack://incidents/2026-05-04T21:25
2026-05-04 21:20:41 UTCskill change@deepakpricing-exception-approval · v1.0.0 → 1.0.1-DRAFT

Opened PR: pricing-exception-approval.skill.md (rephrase cap → threshold).

2026-05-04 21:18:09 UTCapproval@rachelrefund-decision-policy · v1.1.4 → 1.2.0

Approved skill change to refund-decision-policy.skill.md (added damaged-on-arrival §1, retro deck citation).

2026-05-04 21:14:33 UTCingestbrain.ingest

Ingested 4 Notion pages from policies/ workspace · 2 new entities · 1 doc-doc contradiction surfaced.

↳notion://policies/refund-v3
↳notion://policies/cs-tier-matrix
↳notion://policies/sla-by-tier
↳notion://policies/deal-desk-thresholds
2026-05-04 21:12:18 UTCskill loadcursor-bot · @hamzarefund-decision-policy · v1.2.067ms

Loaded refund-decision-policy.skill.md while Cursor agent was drafting refund handler PR.

2026-05-04 21:10:02 UTClint findingbrain.coherence

Coherence run completed · 7 findings (2 HIGH · 2 MEDIUM · 3 LOW) · 2 blocking · 0 LLM calls · deterministic.

2 new findings since last run. Filed PR-2118, PR-2119 to approval queue.

why this beats Glean / Sierra

Glean returns search hits. Sierra runs an opaque agent. Neither gives the customer's compliance team an audit trail of why the agent made the decision it did. We do — every action cites a versioned skill + the exact sources behind it.

what GBrain doesn't do (and we do)

GBrain logs requests + page edit history. It does not trace agent decisions to skill versions to sourcesin a compliance-grade audit. Single-user systems don't need that. Every business does.