The Mandate and the Diary: Two Cores, Ninety Days, What Stays

From:    Office of the CEO, Banco Meridian S.A.
To:      Integration Management Office — War Room (IMO-WR)
Subject: Directive 01/2026 — Aseguradora del Centro integration
Class:   Internal — Restricted

1. Policy administration and claims of Aseguradora del Centro
   (Mexico City) will operate on integrated rails no later than
   12 April 2026. The window is 90 days. This date is fixed.
2. Effective immediately: Group-wide headcount freeze. No net
   additions, no backfills. The program executes with current
   headcount plus contracted virtual squads.
3. Cloud and inference spend: flat to FY25 approved run-rate.
   Any increment requires CFO sign-off with an offsetting
   retirement of equivalent cost.
3.1 The freeze and the spend cap are one constraint, not two.
4. All production inference runs on Provider 1 under existing
   enterprise terms. Admission of a second provider is permitted
   only under equivalent governance — same manifest, same gates,
   same audit span. No shadow endpoints.

   R. Vasconcellos
   Chief Executive Officer

1. Day 0 — Mon 2026-01-12, São Paulo

   The minutes will argue what the memo doesn't

   The memo does not argue its deadline; these minutes will. The first 12–18 months post-close are what determine deal success [McKinsey, 2021], and 50–60% of synergy initiatives are strongly IT-dependent [McKinsey, 2011]. A 90-day window inside that reality is not a plan, it is a bet. This diary exists so the committee can audit the bet week by week.

   Opening position, minuted before any spend — negatives first. Experienced developers working unadapted AI on their own repositories shipped 19% slower while believing they were 20% faster [METR, 2025]; METR's follow-up could no longer measure cleanly because developers now refuse to work without AI [METR, 2026]. Same breath, the market's pairing: >40% of agentic AI projects canceled by end-2027 and ≥15% of day-to-day work decisions autonomous by 2028 [Gartner, 2025]. Doctrine follows from that record, not from the brochure: every squad write goes through the gate; nothing touches adjudication tables on model judgment alone.

2. Day 4 — Fri 2026-01-16 (Week 1 close)

   The core stays; the core gets a governed interface

   Survey pass 1 from the Juárez desk: policy-admin is COBOL batch behind a SOAP _svc layer, 214 endpoints (illustrative); claims intake is a 2011 JSF application; no API gateway. Context for the committee: organizations average ~897 apps with only ~29% integrated [MuleSoft, 2025]. Under §2 and §3 there is no replacement budget. The core stays; the core gets a governed interface.

   Week 1 closes with the routing manifest generated — the governance object the whole program will version from here on. As filed:

   LEGACY SURFACE SURVEY — AsegC core (partial, pass 1 of 3)
     policy-admin (COBOL batch + _svc SOAP layer)   214 endpoints (illustrative)
     claims intake (JSF application, circa 2011)      63 screens / 38 flows (illustrative)
     reporting (Oracle replica, 22 marts)             nightly file-drop bridges (illustrative)
     auth: SSO on the BR side; static credentials on 61% of _svc endpoints (illustrative)
     data residency: MX-only cluster; claims PII dual-flagged LGPD + LFPDPPP
     API gateway: none
   
   ROUTING MANIFEST — generated, rev 1
     providers:
       provider-1:  share 1.00  (existing enterprise terms)
     routes:
       contract-extraction:  provider-1/claude-opus-5
       claims-triage:         provider-1/claude-haiku-4.5
     gates:
       write-scope:  default: deny   approvers: [arch-lead-br, arch-lead-mx]
       evidence:     span: [provider-1]   scopes: [LGPD, LFPDPPP]
   

3. Day 8 — Tue 2026-01-20

   The first contract change

   First contract change enters review, drafted by the squads against an OpenAPI façade held in front of the legacy _svc layer. Diff as committed:

   --- a/specs/claims-core/openapi.yaml   (v1.4.1)
   +++ b/specs/claims-core/openapi.yaml   (v1.5.0-rc1)
   @@ paths @@
   +  /policies/{id}/claims:
   +    post:
   +      summary: Open a claim against an issued policy
   +      operationId: openClaim
   +      parameters:
   +        - $ref: '#/components/parameters/PathPolicyId'
   +        - name: X-API-Version
   +          in: header
   +          required: true
   +          schema:
   +            type: string
   +            enum: ['2026-01']
   +          description: Contract version; gateway rejects unversioned calls
   +      requestBody:
   +        required: true
   +        content:
   +          application/json:
   +            schema:
   +              $ref: '#/components/schemas/ClaimOpenRequest'
   +      responses:
   +        '201':
   +          description: Claim registered, queued for adjudication
   +          headers:
   +            Location:
   +              schema:
   +                type: string
   +              description: Relative URL of the created claim resource
   +        '422':
   +          $ref: '#/components/responses/UnprocessableClaim'
   @@ components.schemas.Policy @@
        policyholder:
          type: object
          properties:
   +        second_surname:
   +          type: string
   +          nullable: true
   +          description: Required on MX policies (LFPDPPP scope);
   +            null on BR-ported policies. Nullable from day one —
   +            do not backfill.
   

4. Day 23 — Wed 2026-02-04, Mexico City

   The quota wall

   Night batch stops; Provider 1 is at its ceiling. Gateway excerpt as captured:

   The quota-increase ticket with Provider 1 quotes a provisioning window measured in weeks (illustrative); 12 April does not have weeks. Directive §4 is invoked the honest way round: admit Provider 2 under the same manifest, the same gates, the same audit span. The fallback — stepping extraction down a model class and accepting quality risk on claims documents — is documented, and nobody wants it.

   2026-02-04T02:17:44-06:00  meridian-llm-gw  region=mx-central
     route=contract-extraction   model=provider-1/claude-opus-5
     status=429   retries=3/3   backoff=exhausted
     upstream body:
     {"type":"error","error":{"type":"rate_limit_error","message":
      "Your account has exceeded its concurrent request rate limit
       for this model. Retry after 42s or reduce parallelism."}}
     queue_depth=1317   projected SLA breach 04:00 CST   tasks affected: extraction
   

5. Day 31 — Thu 2026-02-12, São Paulo

   The manifest is the change

   Provider 2 clears admission review. The routing manifest is the change — diff as merged (rev 88 → 89):

   What did not change: the deny-default, the two approvers, the single governance object — the evidence span widens to both providers instead of forking. The economics that justify the shuffle are external: ~2.5x price-performance spread between flagships and ~100x across the full catalog [Artificial Analysis, 2026]. Under a flat spend mandate, the routing table is a budget instrument before it is a technical one.

   --- a/manifests/routing.production.yaml   (rev 88)
   +++ b/manifests/routing.production.yaml   (rev 89)
   @@ providers @@
    providers:
      provider-1:
   -    share: 1.00
   +    share: 0.60
   +  provider-2:
   +    endpoint: vertex-ai://us-east1
   +    share: 0.40
   +    admitted_by: gate-arch-2026.02.11
   +    terms: enterprise, DPAs MX+BR on file
   @@ routes @@
    routes:
      contract-extraction:
   -    model: provider-1/claude-opus-5
   +    primary: provider-2/gemini-3.1-pro
   +    fallback: provider-1/claude-opus-5
      claims-triage:
        model: provider-1/claude-haiku-4.5
   @@ gates @@
    gates:
      write-scope:
        default: deny
        approvers: [arch-lead-br, arch-lead-mx]
      evidence:
   -    span: [provider-1]
   +    span: [provider-1, provider-2]
   

6. Day 35 — Mon 2026-02-16

   The gate against our own squad

   First approval gate on file — the gate working against our own squad. Gate log, verbatim:

   Requested scope, reviewer, decision — three log lines, no committee. Filed. No exception granted; none requested.

   2026-02-16T14:03:11-03:00  GATE write-scope
     actor=squad-claims-02
     request=UPDATE policy_claims.adjudication_status
     decision=DENY  reason=no write scope on adjudication tables (read-only route)
     reroute=human-approval queue  approver=arch-lead-mx
   
   2026-02-16T15:41:52-03:00  GATE write-scope
     actor=arch-lead-mx
     request=APPROVE squad-claims-02 patch pc-114
     decision=APPROVED  scope=adjudication_status, single column, TTL 6h
   

7. Day 42 — Mon 2026-02-23, Mexico City

   Cutover stage 1 — the canary

   Cutover stage 1 — claims intake canary. Checklist as executed:

   The bar we hold the program to is the DORA-elite line — <1-day lead time, on-demand deploys, 5% change-fail, <1h recovery [DORA, 2024]. On a 2011 JSF front end, "elite" is not a mood; it is a checklist with a rollback point.

   CUTOVER stage 1 — claims intake (canary)
     [x] contract v1.5.0 published to gateway, both regions
     [x] traffic: 5% of MX intake -> /policies/{id}/claims
     [x] rollback point rb-2026.02.23-01 armed (gateway switch, under 60s)
     [x] gate log streaming to evidence pack, dual scope LGPD + LFPDPPP
     [ ] stage-2 gate: 48h at canary share with change-fail within the 5% bar [DORA, 2024]
   

8. Day 56 / Day 70 — staged cutovers

   Staged cutovers, checklists as executed

   The audit trail now spans both providers continuously; nobody merged anything to make that true — the manifest did it at rev 89 (illustrative). The legacy core is still there, still COBOL, still nobody's rewrite project. It has a governed interface now, and every write against it has a name on it.

   CUTOVER stage 2 — 2026-03-09   intake share 25% (illustrative)
     [x] 48h at canary clean; change-fail within the 5% bar [DORA, 2024]
     [x] rollback drill fired deliberately mid-window; switch-back
          completed inside the armed window
     [x] drill logged as evidence, not as incident
   
   CUTOVER stage 3 — 2026-03-23   intake share 100% (illustrative)
     [x] all new claims intake on the integrated rail
     [x] SOAP/file-drop bridges on the intake path retired (illustrative count)
     [x] claims-contract lead time under one day since stage 2 [DORA, 2024]
   

9. Day 77 — Mon 2026-03-30

   The labor question, minuted

   The committee asks the question committees ask: how many engineers did the squads replace? Answer filed:

   None — and the question is the wrong unit. Headcount was frozen before the first squad was contracted (§2); there is no counterfactual of hires not made to count against. The program was priced in outcomes: since stage 2, the claims-contract path has run inside the DORA-elite bar — <1-day lead time, on-demand deploys, 5% change-fail, <1h recovery [DORA, 2024]. For context only, the labor market we did not hire in: US senior fully loaded $250–350K/yr against LATAM nearshore at $65–72K [Arc.dev, 2026; Howdy, 2026]. That arbitrage is the floor, not the pitch — it is what made the mandate survivable; it is not what anyone bought.

   What we designed against, the record first (Day 0 stands): unadapted AI on expert work measured −19% [METR, 2025], measurement itself collapsed by February 2026 because developers refuse to work without AI [METR, 2026], and a market where >40% of agentic projects get canceled by end-2027 even as ≥15% of day-to-day decisions go autonomous by 2028 [Gartner, 2025]. The one result built like our situation is a call-center study: a gen-AI assistant lifted throughput +14% on average, +34% for novices, ~0% for experts [Brynjolfsson, 2023]. Read that as an instruction, not a comfort — AI does not make your experts measurably faster; it makes your experts' judgment distributable. So the engagement never bought assistants for experts. It encoded how experts decide — routing, write-scopes, eval bars — into the manifest and the gates, and ran ninety days of operations through that encoding. The operations were the training run. What is left at Day 92 is trained intelligence.

10. Day 84 — Mon 2026-04-06, São Paulo

    The evidence pack, exported ahead of the fixed date

    Evidence pack exported, ahead of the fixed date:

    The comparison the compliance office draws for the board: manual SOC 2-style evidence preparation costs organizations 100–500 engineering hours per audit cycle [Workstreet, 2025; Atlant, 2026; Drata, 2026]. This pack was not a project; it is the exhaust of a governed pipeline — nothing in it was produced for the audit.

    EVIDENCE PACK ep-2026Q1-integration (export summary)
      audit span     : 2026-01-12 -> 2026-04-06, continuous
      providers      : provider-1, provider-2 (single span, no fork)
      scopes         : LGPD (BR data subjects) + LFPDPPP (MX data subjects)
      contents       : gate decision log (denials + approvals)
                       routing manifest revisions 1..96
                       model-call audit trail, both providers
                       cutover checkpoints + rollback drills
                       PII access map, claims domain
      integrity      : sha256 manifest of contents (illustrative hash)
    

11. Day 90 — Sun 2026-04-12

    The date held

    The date was fixed. The date held. Directive 01/2026, closed on its own terms: integrated rails live since stage 3, spend flat, headcount untouched.

12. Day 92 — Tue 2026-04-14

    Demobilize — what stays

    Squads scale to a monitoring watch (illustrative staffing detail). What does not demobilize — all of it in Banco Meridian's own cloud account (BYOC), versioned, one audit span across two providers and two regulators: the routing manifest (rev 96, illustrative), the eval suite that gates every model change, the write-scope decision history, the evidence pack. The framing the committee should carry into the next acquisition: the task length agents can complete doubles roughly every seven months, replicated across nine domains [METR, 2025]. Labor does not compound; a versioned decision history does. The next acquired core starts from trained routing, tested gates, and ninety days of decisions already on file — not from a memo and a bet.

The engagement ends. The intelligence doesn't.

Start with evidence — the AI Governance Assessment