What an FDE delivers after an engagement
Alongside the customer-owned workflow, deliver a short de-identified method or asset with a stated level. Used once: worked in one engagement, with limits recorded. Reusable: independently checked with identifying details removed, rights/confidentiality cleared, prerequisites and verification documented, and someone else able to apply it. Standard: evidence from at least two engagements, portability beyond the original setup, an eval set and baseline, and counted reuse. These are academy evidence labels, not certifications. A capstone simulation alone normally supports “used once (simulation)”.
Before reuse, replace source data with synthetic examples, remove unique combinations that could identify a customer, inspect screenshots/metadata and confirm permission to reuse the method. Pseudonyms or hashes alone are not de-identification.
| Method | Delivered asset and evidence |
|---|---|
| 1. Approval-gate pattern | Complete-envelope schema, state table and tests for authenticated approval, token origin, expiry, replay and missing scope. B1.05/B1.07/B4.04. |
| 2. Action-log pattern | Durable intent and outcome contract, receipt reconciliation, restricted snapshots and retention evidence. B1.06. |
| 3. Connector contract and facade | Versioned actions/auth/errors, private executor boundary and a refusal case for inseparable side effects. B1.02/B3.03/B3.08. |
| 4. Just-enough identity and scopes map | Principal/authority/resource map, declined access, grant owner and revoke evidence. B2.03–B2.05. |
| 5. Emergency and separation-of-duties playbook | Named emergency roles, second check, expiry, reason and review; never bypass approval, scope, freeze or logging. B2.06–B2.07. |
| 6. Idempotent writes and checkpoint-resume | Operation/digest keys, receipt reconciliation and crash/concurrent-retry evidence. B3.07. |
| 7. Knowledge the agent can cite | Owned sources, refresh dates, access filters, supported citations, abstention and stale/conflict tests. B3.09. |
| 8. PDPA minimisation and retention recipe | Purpose/field/reader/transfer map, retention/correction plan, hold review and expiry across copies. B3.06; regional briefing applies. |
| 9. Eval-set template | Synthetic awkward cases, separate owner labels, held-out variants, metric denominators and versioned baseline. B4.01–B4.02. |
| 10. Unhappy-path and red-team suite | Schema/injection/gate/isolation tests, deliberate failing mutations and findings with non-waivable controls. B4.04–B4.05/B4.08. |
| 11. Handover pack template | Five owner answers, current scopes/gates, stop/revoke/recovery steps, exceptions, operator exercise and expired temporary access. B5.06–B5.07. |
| 12. Discovery, redesign and scope brief | Observed workflow, central/tail waits, unknowns, redesign hypothesis, constrained scope and net-value estimate. B6.02/B6.04. |
| Companion: Saying-no and recut playbook | A justified refusal and smaller feasible alternative when gates, evidence, owner or scope are missing. B6.05. |
| Companion: Weekly check-in rhythm | Fifteen-minute agenda, exceptions and approval workload, one decision and a check next week. B6.08. |
Short method template
- Method and level: name; used once/reusable/standard; simulation or engagement evidence; date.
- Problem and boundary: intended outcome, prerequisites, exclusions and conditions where it should be refused.
- Steps and controls: who does what; identity/access; exact approvals; stop/recovery; privacy and retention.
- Evidence: synthetic example, baseline, assertions, failure cases, result and limits; no customer records.
- Reuse: what must change in a new setting, rights/de-identification review, reviewer, known limitations and the evidence required for the next level.
Starter catalogue
| # | Method | What it is | Module that produces it |
|---|---|---|---|
| 1 | Approval-gate pattern | A hard stop for send, pay and delete: bound to the exact payload a named person approved, with a backup, and with no prompt-only rule or timeout that approves itself. | B1.05, reinforced by B1.07, B4.03 and B4.04 |
| 2 | Action-log pattern | An append-only, write-before-execute log that can answer what happened, who approved it, and whether those two things were the same, with personal data minimised. | B1.06 |
| 3 | Connector contract and facade | A versioned contract so a proposal can only execute through policy, with one connector identity per agent, and a clear refusal when the connector cannot honour a gate. | B3.03, introduced in B1.02, with the refusal case in B3.08 |
| 4 | Just-enough identity and scopes map | The agent has its own principal; roles and scopes match the workflow; grants can be reviewed and revoked; there is no tenant-wide convenience access. | B2.03, B2.04 and B2.05 |
| 5 | Break-glass and segregation-of-duties playbook | A logged emergency override that still needs a named person, never skips the log, and does not become the daily path, plus backup approvers and a split between proposer and approver. | B2.06 and B2.07 |
| 6 | Idempotent writes and checkpoint-resume | Every send or pay has an idempotency key; a timeout is reconciled before retry; a long run resumes from its last checkpoint instead of repeating finished work. | B3.07 |
| 7 | Knowledge the agent can cite | Every source has an owner and a refresh date; every answer carries a citation or says there is no evidence; stale or contradicting entries are flagged. | B3.09 |
| 8 | PDPA minimisation and retention recipe | What the agent may keep, for how long, and how personal data is minimised in logs, fixtures and any reused asset, starting from Singapore's PDPA and flagging Thailand's PDPA where people in Thailand are involved. | B3.06 |
| 9 | Eval-set template from real-shaped cases | A de-identified set built from awkward cases (mismatches, missing notes, mixed currency), with a clear measure of "good" and expert scoring on the hard items. | B4.01 and B4.02 |
| 10 | Unhappy-path and red-team suite | Tests for skipped gates, stale payloads, silent writes and log gaps, plus a pre-go-live red-team whose findings are fixed or formally accepted. | B4.04, B4.05 and B4.08 |
| 11 | Handover pack template | A short pack the owner and the next engineer can use: diagram, scopes, gates, log location, revocation, known exceptions, and the five owner questions, with temporary access already expired. | B5.06 and B5.07 |
| 12 | Discovery, redesign and scope brief | How to recover the real workflow from an event log, place steps on the certainty × consequence grid, redesign before automating, and write an honest net-value estimate that counts exceptions, review, change and risk. | B6.02 and B6.04 |
Companion methods
Two close companions, useful once the first twelve exist, but not required in the opening set:
- Saying-no and recut playbook (B6.05) — how to decline or recut a request that has no interceptable gate, no owner, no way to judge "good", or a blast radius you cannot defend.
- Weekly check-in rhythm (B6.08) — a fifteen-minute owner meeting that starts from exceptions, overdue gates and odd spend, and ends with one decision to check the following week.
The capstone (B7.06) is where a learner first submits one de-identified asset and an honest level. That submission is practice for this library, not a substitute for it.