Examples — Scenarios and Worked Examples

Everything in this directory is invented. No real project, client, or measurement appears anywhere in it. What is real is the reasoning: each example works through a decision the playbook has an opinion about, and shows what went wrong when that opinion was ignored.

Read them for the shape of the mistake, not for the numbers.

Scenarios (SCENARIO-01…12)#

Twelve fictional projects, each built around one lesson that is easy to agree with in the abstract and easy to miss in practice. Each follows the same arc — Situation · Decision faced · Evidence · What happened · The generic lesson · What would NOT have worked · References — and ends by telling you what to go check in your own work.

The chapters cite these as [SCENARIO: SCENARIO-NN]. A scenario illustrates a rule; it is never the evidence for one. Where a chapter claims something is established, the support is an external source in the ledger, not a scenario. Chapters stay readable if you skip these entirely.

ID Scenario The lesson in one line Illustrates chapters
SCENARIO-01 Consequence-bearing tolerances A threshold with no named consequence is an escape hatch nobody chose 00, 04, 07
SCENARIO-02 Clustered eval, effective N Count your groups, not your rows 03, 04
SCENARIO-03 Learned-router leakage A learned component must beat the group-identity ceiling, or it learned the wrong thing 03, 04, 08
SCENARIO-04 Harness defects Check the ruler before you trust the measurement 02, 03
SCENARIO-05 Hardware purchase discipline Measure the critical path before you spend the money 06, 11
SCENARIO-06 Token-budget confounding An uncalibrated output cap can invent a quality difference that isn't there 04, 05, 07
SCENARIO-07 Deterministic cascade gate Price routing's ceiling first; a few booleans may capture most of it 08, 10
SCENARIO-08 Transport serialization defect "Transparent" is a claim to test, not a property to assume 02, 08, 12
SCENARIO-09 Suite versioning under criteria drift Version the instrument; refuse comparisons across versions of it 03, 13
SCENARIO-10 Pilot optimism collapse Small-sample headlines run optimistic 03, 04
SCENARIO-11 Diagnostic gate A cheap probe, agreed in advance, can settle an expensive question 04, 07
SCENARIO-12 Restart instability, paired controls Determinism is measured, not assumed 02, 04

The synthetic walkthrough#

WALKTHROUGH_rag-document-qa.md — one complete navigator pass (QUICKSTART → profile → tier → archetype → first sprint → frozen contract) on an invented document-QA project. Entirely fictional, and deliberately shaped unlike everything else here — it exists to show the navigator running end to end on an unfamiliar problem.

Synthetic worked examples (SYNTH-01…10)#

Compact invented scenarios — one per project shape — showing the playbook's rules firing on materially different problems. All names and numbers are invented; each exercises at least one gate or tripwire for real.

ID Shape The point it proves
SYNTH-01 API-only greenfield assistant Eval-first bootstrap; the judge-calibration path for open-ended outputs
SYNTH-02 Privacy-mandated local deployment Execution-system identity and reproducibility probes before any number is trusted
SYNTH-03 High-throughput extraction Deterministic grading; operating points; cost-per-successful-task drives model size
SYNTH-04 Tool-calling agent with write actions Tool contracts, permission gating, injection threat model, observe-only graduation
SYNTH-05 Cost reduction via routing Oracle analysis and break-even first; a leaky learned router is refused
SYNTH-06 Training rejected Diagnosis lands on cheaper rungs; the ladder refuses the fine-tune
SYNTH-07 Training justified The evidence bar actually met: rig-first, ablation, forgetting gate, legality check
SYNTH-08 High-stakes audited deployment Tier 3: governance at intake, pass^k, human-gated canary, rehearsed rollback
SYNTH-09 An INCONCLUSIVE result Below-MDE outcomes reported honestly; the decision made on cost, not a fake winner
SYNTH-10 Hardware rent-vs-buy Demand ledger + break-even + pre-committed trigger; the purchase correctly waits