Glossary

The concepts used in this book, grouped by the book's six parts. Each entry lists the English term, its Chinese original, a one-line definition, and the chapter where it first appears.

---

Part One · What an FDE Actually Is

Term	中文	One-line definition	First appears
Forward Deployed Engineer (FDE)	前线部署工程师	An engineering role stationed at the customer's site and accountable for the system producing results in real business operations — delivering many capabilities for a single customer, rather than a single capability for many customers. In Chinese it is rendered 前线部署工程师, also commonly seen as 前沿部署工程师 in hiring contexts; abbreviated FDE.	Chapter 1
the Four Responsibilities: Discovery / Production / Adoption / Feedback	责任四分法	A four-cell coordinate system for locating delivery roles: discovery, production engineering, adoption, and feedback into the product. It describes position, not merit — covering only one or two cells is the norm, not a scorecard.	Chapter 1
Echo / Delta	Echo / Delta	A two-person split originating in Palantir practice: Echo decides what should be done (discovery and relationships), Delta builds it (delivery and engineering). It separates responsibility, not seniority or tech stack.	Chapter 1
AI FDE	AI FDE	The productization of part of the FDE's work: natural-language instructions turned into platform actions, executed within the user's existing permissions, fully logged, pausing on anomalies for clarification. It is Palantir's product name — a capability claim from their product documentation, not an independently verified result.	Chapter 4
the three tests: where the code goes, where the experience goes, how it is priced	三个评判标准	Three tests for whether a role is really an FDE: where the code goes, where the experience flows, and how the engagement is priced and held accountable. They replace the unreliable habit of matching a role against a job-description checklist.	Chapter 2

Part Two · How the Problem Gets Solved

Term	中文	One-line definition	First appears
the five filters	五道关	Five sieves in the discovery phase that eliminate candidate problems one by one until a single entry point remains; the first asks "does this problem matter?", and the remaining four narrow further.	Chapter 6
the discovery deliverables	发现三件套	The three things discovery should hand over: a problem card, a data-accessibility inventory, and a prototype interaction you can actually touch. Not a slide deck — the test is whether work can start the very next week.	Chapter 6
Minimum Viable Deployment (MVD)	最小可行部署	The smallest delivery that runs end-to-end in a real business and can be measured — not another demo. It differs from an MVP in the "deployment": it must have real users, real data, and a real workflow.	Chapter 7
the three "real" conditions: real users, real data, real workflow	三个「真」	The three necessary conditions of an MVD: real users, real data, real workflow. Missing any one of them, what you built is still a demo.	Chapter 7
outcome acceptance sheet	效果验收单	The pre-launch document that defines what counts as success; omit any of its five elements and it falls apart. It divides labor with the operational readiness checklist — this one governs outcomes, that one governs operations.	Chapter 8
operational readiness checklist	运行就绪清单	The pre-launch checklist that settles who runs the system: who watches the alerts, how on-call is staffed, where the runbook lives. A runbook that exists only on the vendor's laptop does not exist.	Chapter 8
responsibility handoff table	责任交接表	A table that names every mode of failure and assigns each one an owner. Its design principle: every way of dying has a name.	Chapter 8

Part Three · Engineering Foundations

Term	中文	One-line definition	First appears
the eight walls	八堵墙	The eight categories of capability gap between PoC and production: data and permissions, integration, tool boundaries, controllable operation, provable outcomes, trustworthy results, cost and stability, and transferable accountability. Stronger models only lower the first few walls; the rest are set by organization and process.	Chapter 8
harness	harness	The layer outside the model that turns it into a usable system: tools, integrations, filesystem, orchestration, supervision. (The Chinese edition keeps this term untranslated.) Together with "model capability" and "interfaces" it forms the book's three-layer division.	Chapter 9
Data Contract	数据契约	A written agreement, settled before touching data, that pins down four things: access scope, visibility of results, retention and destruction, and audit requirements. Three parties use it — the customer's compliance review, the FDE's own discipline, and post-hoc accountability.	Chapter 5
Tool Contract / ToolSpec	工具契约	The spec defining which actions an agent may call, how parameters are validated, and how out-of-bounds calls are refused. It and the data contract govern the two sides: the data contract governs what you read, the tool contract governs what you do.	Chapter 9
Run State Machine	运行状态机	A decomposition of one agent run into state transitions that can be observed, paused, and resumed — the underlying structure that makes checkpoint & resume and approval tickets possible.	Chapter 9
approval ticket (HITL)	审批工单	A pending ticket generated before an irreversible action executes, continuing only after a human approves. Not a popup — popups are instant and easily clicked away; tickets can be held accountable and can wait overnight.	Chapter 11
Checkpoint & Resume	检查点与恢复	Continuing from the latest state after an interrupted run instead of starting over. The direct reason it exists: an approval that sat overnight should not force a full rerun.	Chapter 9
tiered authorization	分级放权	Deciding whether human involvement is needed by the risk level of each action, rather than routing every action through approval. Pairs with the approval ticket: the ticket is the mechanism, tiering is the policy.	Chapter 11
Golden Set	基准样本	A standard set grown out of real cases, defining "what counts as correct" — the first step of an evaluation system. Distinct from the regression set: the golden set sets the standard, the regression set prevents regressions.	Chapter 12
Regression Set	回归集	A mistake notebook that only grows: every failure is added, and every subsequent change must rerun it. Its enforcement is the release gate — no change deploys without passing the regression set.	Chapter 12
LLM-as-Judge	模型裁判	When open-ended tasks have no standard answer, a model scores outputs against a written rubric. The rubric itself is a "prompt change" — changing it goes through the same gate as changing the system.	Chapter 12
release gate	门禁	The mandatory checkpoint where a change must pass the regression set before it may deploy. It is the enforcement side of the regression set — mechanism and enforcement.	Chapter 9
eval set	评测集	The umbrella term for the golden set, the regression set, and the judging rules — one of the assets an FDE can take with them. What travels is task types, difficulty tiers, judging rules, and failure modes — never the customer's case data.	Chapter 7
the three layers of trust	可信度三层	The three-layer structure for proving yourself when the customer challenges you: every answer carries its sources, every decision step can be replayed, and failures are classified. Lose any one layer and the proof chain breaks at exactly that layer.	Chapter 13
EvidenceRef	证据引用	Citing the underlying source alongside every conclusion — the concrete form of the first layer of trust.	Chapter 13
Trace / Replay	回放	Reproducing every decision step of a run exactly as it happened, for after-the-fact verification — the second layer of trust. It differs from logging: a log is a record; a replay can be re-enacted.	Chapter 13
Service Level Objective	SLO	Service-level commitments written into the contract, with availability, latency, and quality stated separately. Their cost curves differ; collapsing them into one phrase like "the system runs stably" commits to none of the three. The quality bar comes straight from the eval set.	Chapter 14
cost attribution	成本归因	Breaking the bill down to every run and every stage instead of staring at the total. Without attribution there is nothing to optimize — it is the precondition of cost governance.	Chapter 9
model routing	模型路由	Dispatching requests to differently-tiered models by task difficulty, targeting the disease of model mismatch — using a flagship model on a simple task is delivering takeout in a sports car.	Chapter 14
budget circuit breaker	预算熔断	Automatically stopping or downshifting when spend hits a threshold, before the bill runs away. It differs from graceful degradation: the breaker trips on cost, degradation on failure.	Chapter 14
the four resilience primitives: timeout, retry, idempotency, graceful degradation	韧性四件套	Four mechanisms against four ways of dying: timeout, retry (bounded, with backoff), idempotency, and graceful degradation. Degradation's last rung is falling back to humans and spreadsheets — only a plan that dares to write down its fallback dares to go live.	Chapter 13

Part Four · The Real Work Begins After Go-Live

Term	中文	One-line definition	First appears
activation	激活	The moment the system is genuinely used by the front line and enters daily behavior. Strictly distinct from go-live: go-live is a process, activation is behavior. The quiet death after a "successful launch" is the pattern to watch for.	Chapter 15
executive sponsorship	一把手工程	Cross-department priorities, budget ownership, and permission rulings must be endorsed directly by the decision-making level, or nothing moves. Not a universal slogan — there is a boundary question of "when can it stop being an executive-sponsorship project".	Chapter 6
the three exit conditions	撤场三件事	The three things that must be complete before handoff: business integration, knowledge governance, and system interoperability. Its design intent: delivery is walking down the stairs, not jumping off a cliff.	Chapter 17
who defines the last mile	定义权	Who gets to decide where the "last mile" ends — which directly determines whether the FDE does engineering work or negotiation work. It explains much of the China–US difference: in the Silicon Valley narrative the FDE defines it; in Chinese field practice, often not.	Chapter 12

Part Five · How Does This Business Actually Work?

Term	中文	One-line definition	First appears
the big ledger vs the small ledger	大账与小账	The small ledger counts hours and API fees; the big ledger counts the cost of what gets replaced and the price of errors. Cost's first-principles question is "compared to what" — compare against the replaced human decision cost, not against zero.	Chapter 19
the six metrics	六个指标	The six numbers that put the big ledger on paper, including implementation leverage and customization decline rate. A personally derived framework: the book borrows its vocabulary but presents no industry benchmark.	Chapter 19
Result as a Service (RaaS)	结果即服务	The vendor tying part of its fees to business outcomes, not just to delivering a system. Behind "no pay unless it works" lie at least four questions to settle first: where the effect is measured, who it is attributed to, how long to wait, and who fronts the cost.	Chapter 20
declining customization	定制递减	The amount of customization each similar requirement needs must fall with every subsequent customer, or the business does not stand. The critique is not customization itself but customization that cannot decline. The dividing line: configuration changes parameters, customization changes code.	Chapter 19
the six reusable assets	六类资产	The six kinds of things that can flow back from the field: business ontology, connectors, tool and approval templates, workflow templates, evaluation structures and failure-mode libraries, and demand signals. What flows back is not "experience" — experience scattered in individual heads is not accumulation.	Chapter 22
three steps and two gates	三步与两道闸	The process that gets assets genuinely back to the platform: record the context, discuss it with the product team, have other FDEs validate it. One gate at the entrance classifies demand; one at the exit assigns every asset a clear destination. A review that only collects and never dispositions becomes a junk drawer.	Chapter 22

Part Six · People and Organization

Term	中文	One-line definition	First appears
the three pillars	三大支柱	The three facets of FDE capability; what matters is the intersection, not excellence in any single one. The weighting shifts with market structure — in Silicon Valley samples engineering dominates; in Chinese samples consulting and sales come first.	Chapter 24
discoverer and deliverer	发现者与落地者	Two kinds of people within one role: those good at finding what should be done, and those good at getting it done. Isomorphic to Echo/Delta; useful for hiring discussions and self-assessment.	Chapter 24
the trio	三人组	The minimal unit common in the Chinese market: a strategist, an industry consultant, and an AI full-stack engineer — one person, two, or a whole team. Read it as the Chinese variant of Echo/Delta; headcount is not the point.	Chapter 1
Forward Deployed Product Manager (FDPM)	前线部署产品经理	The product role paired with the FDE; one FDPM typically carries two to three FDEs (Cresta's ratio). Another modern variant of the Echo/Delta pattern.	Chapter 26
Tu FDE (lit. "grassroots FDE"; community coinage)	土 FDE	A self-given name used in the Chinese community for FDEs operating as independent vendors — running several engagements in parallel and building their own case libraries and reusable modules. An informal term; the English edition keeps the pinyin rather than translating it literally.	Chapter 27
fractional	分数型	Serving multiple clients as an FDE part-time and in fixed proportion. The difference from moonlighting lies in the structure of responsibility, not just the ratio of hours.	Chapter 27
the market segmentation quadrant	市场分层四象限	Segmenting customers by value (high/low) × decision speed (fast/slow) and targeting the "high-value × fast-decision" cell. (Framework contributed by an interview guest; its figures are all self-reported.)	Chapter 27

---

50 entries in all.