Glossary

Appendices

An engagement can fail over one word: a threshold nobody wrote down, or a pilot with no end date. Each term the book relies on is defined once below, with the Chinese term the Chinese edition uses and a link to the section that argues it. Where a definition and its chapter seem to disagree, the chapter is right.

E.1

Using this glossary

Entries are alphabetical, in groups of letters. Each row gives the term, the Chinese term used in the Chinese edition, and the book's meaning in a sentence or three, ending with the section where it is defined or argued. A definition here is a summary. The evidence, the caveats and the tier of each source are in the linked section.

Many entries are this book's own proposals: the rescue-or-walk rule, the skill map's levels and the five load-bearing conditions among them. Nobody has tested them against outcomes, and the chapters say so. They are proposals here too. The same goes for the running example: terms marked as part of the composite describe an invented insurer, not a reported case.

If you read only a handful, read the four load-bearing commitments (threshold, data path, sponsor, exit) and the entries that make the first of them concrete: threshold owner, held-out set and baseline. Judgment Under Pressure puts the reason in one line: the concession that sinks an engagement "rarely looks fatal: it is the threshold, the data path, the sponsor or the exit, given up to keep a meeting friendly."

Two terms come in families. The evidence tiers are filed under T, with the rule that separates T2 from T4. The four skill domains are each filed under their own name, and their levels under L.

E.2

A–D

Acquire, Expand, Scale

获客（Acquire）/ 扩展（Expand）/ 规模化（Scale）

The three customer phases in Palantir's S-1. In Acquire it runs pilots "at no or low cost" and at a loss; in Expand it keeps losing money on purpose; in Scale customers become "more self-sufficient" and the margin turns. In 2019 the top quarter of Scale customers ran at an 87% contribution margin (T2), the most quoted number in the table and one with four limits the filing itself states. Origins and Economics

Adoption

采用

Whether the system changes anyone's work, defined by what users do: suggestions shown, accepted and overridden, and the share of eligible work that goes through the new path. It is a deliverable, so it needs what the threshold needed: a definition, a baseline, an owner and a date. Its three failure shapes are ignored, overridden, and accepted without looking; the last is the dangerous one. Production and Handoff

Audit log

审计日志

The record of what the agent actually did. For a propose-only agent, the minimum per suggestion is the input references, the output and any reason given, the model, prompt and code versions, what the human did with it, a timestamp, and the identity the agent ran as. The human's decision is the field teams forget and model risk cares about most. Security, Governance and Model Risk

Baseline

基线

How the current process performs before AI, measured in week 1: a sample of recent cases, timestamps the systems already hold, and two of the customer's experts labelling the same cases independently. No baseline means no win. The labelled sample becomes the first held-out set. Scoping and the First Win

Business lever

业务杠杆

A number the business already watches and someone senior is paid to move: cycle time, cost per case, loss ratio, time to hire. The sponsor names the lever before anyone names a technique. A request that cannot be tied to one is a technology looking for a use. Discovery

Ceiling

上限

How often the customer's experts agree with each other on the same cases. No agent can be shown to do better than that, so the ceiling goes into the eval plan next to the threshold. In the composite the senior adjusters agree in the low-to-mid 90s. Evals: Proving It Works

Composite

合成案例

The invented engagement the book follows, one step per chapter: a regional property-and-casualty insurer, a mid-level FDE, a part-time lead and an account executive. Three rules hold it to the evidence: it always sits in a box marked Composite; it uses ranges, never point figures, and never names or hints at a real company; and every step ends by citing the sourced pattern it dramatises. How to Use This Book

Consultant and contractor

顾问 / 外包工程师

The two neighbours in Tryolabs' test (T4). The consultant runs discovery, hands over a strategy and leaves "before a line of code is written": the front without the back. The contractor builds what is on the ticket but won't say the ticket is wrong or stay to see it used: the build without the front. The FDE has to own both ends. What an FDE Is, and Isn't

Contract test

契约测试

A scheduled check on every data source that the fields, types and codes the translation layer depends on are still there and still mean what they did. When it fails, the agent stops suggesting rather than suggesting from bad inputs. Data Access and Integration

Credit on conversion

转化时抵扣

The pilot fee is credited towards the annual contract if the customer signs within a set window. It takes the sting out of paying for a pilot and rewards a fast decision. One of six terms the practitioner advice converges on (T4). Before the Engagement

Customer domain

客户（领域）

One of the skill map's four domains: getting the right problem and an agreed number. It means finding the observed problem behind the stated one, negotiating a threshold with the person who will live with it, and delivering bad news so that the engagement survives it. The Skill Map

Data residency

数据驻留

Which region each data store and processing step sits in, and whether that is a contract term or a configuration someone could change. It is often decided by a setting, sometimes one fixed when a project is created, so check it on every service before the build starts. Security, Governance and Model Risk

Decomposition round

拆解轮

An interview round that hands you an open problem and watches you turn it into something buildable. The book treats it as discovery on a clock: the lever, the workflow and its data, the first win and the "not now" list, the number, and who runs it after you. Getting Hired

Deployment company

部署公司

A business that sells forward-deployed hours: delivered outcomes are the product and utilisation is the cost line. In 2026 OpenAI and Anthropic each moved deployment capacity into one (T3; the details sit in a Verified box). The charitable reading is that it protects the labs' own FDEs; the sceptical one is that the services trap now has its own company. Origins and Economics; Where FDEs Work

Deployment Strategist

部署策略师

The non-coding partner role. Palantir lists Deployment Strategist posts beside its Forward Deployed Software Engineer ones: the strategist owns the problem and the adoption, beside the engineers rather than as one. Salesforce describes pods of one deployment strategist and two FDEs (T4). Some employers also use the title for FDE work itself. What an FDE Is, and Isn't; Leading FDEs

Development set

开发集

The set you work against every day: the first five expert cases plus the failures from the shadow-run. Its score should start low; one you pass easily is not finding anything. A held-out set whose failures you study to decide what to fix has become a development set, and you need a new held-out set. Evals: Proving It Works

Discovery

需求挖掘

The phase that treats the customer's stated problem as a hypothesis and finds the lever leadership cares about and the workflow where it lives, on site, before any model is chosen. It is finished when one page exists with the stated problem, the observed problem, the evidence and the lever, and the sponsor signs the observed version. Discovery

E.3

E–H

Echo

回声数字

One self-reported figure that several outlets repeat until it looks independently confirmed. The number of outlets adds no confirmation: it is one source, counted many times. The book names every echo it finds and traces it to its source. Known echoes include the Indeed posting-growth figures, Morgan Stanley's "98%", Tomoro's "~150 FDEs", and Klarna's "two-thirds" versus "75%". How to Use This Book

Echo and Delta teams

Echo 团队 / Delta 团队

Palantir's two team types as described in a write-up of Bob McGrew's podcast appearance: Echo teams of domain experts and Delta teams of rapid prototypers (T3). Unrelated to the echo convention above. Origins and Economics

Engagement

交付（一次交付）

One customer, from first contact to handoff. Part II follows the composite's engagement phase by phase. What an FDE Is, and Isn't

Engagement brief and pilot terms

交付简报与试点条款

The one-page brief the FDE, the lead and the account executive fill in before signature: sponsor, decision-maker, budget owner, the one KPI and its baseline source, duration, price and credit on conversion, data access by day 5, exit criteria, and what happens on success. A field that cannot be filled is the finding. Before the Engagement

Eval set

评测集

Cases, a grader for each, and a number that comes out the same whoever runs it. The book's claim is that the eval set is the contract: build it with the customer's experts before the agent, so that a pass is the customer's finding rather than the vendor's claim. It comes as three sets: held-out, development and regression. Evals: Proving It Works

Exit

退出机制

The agreed way out: the gate that decides go-live, the stop rule, and the handoff. One of the four load-bearing commitments. Without it the vendor becomes the operator. Judgment Under Pressure

Field log

现场日志

The FDE's own log, kept from week one and separate from the customer's tracker. Four sources feed it: eval failure categories, workarounds, override patterns and second builds. Each one-line entry has six fields: date, source, what happened (no customer identifiers), cost, destination, and whether any of it belongs to the customer. The destinations are product, research and the playbook. Feeding the Product

First win

首个成果

The first scoped result: one that matters to the business but fits in weeks. High stakes, small scope: cut along the business (one line of business, one step, one region or share of traffic, propose rather than act), never away from it. A first win on a side workflow proves nothing, and one without an agreed number cannot be won. Scoping and the First Win

Forward Deployed Engineer (FDE)

前线部署工程师（FDE）

The book's definition: an FDE owns the customer's outcome across discovery, scoping, build, production and handoff, and writes the code as the means of getting there. Take away the front and you have someone who builds what they are told; take away the back, someone who leaves before it works; take away the code, an advisor. For the Chinese term and its aliases, see E.8. What an FDE Is, and Isn't

Grader

评分器

The check that scores each eval case. When the number looks wrong, suspect the grader before the agent; in an engagement the commonest broken grader is a wrong label. Relabels on the held-out set go through the threshold owner's experts and are logged with a reason and date. Evals: Proving It Works

Guardrail

护栏

A class of error that must never happen, paired with the main metric: in the composite, a litigation-flagged claim reaching the fast queue. It is usually what the customer's risk function asks about first, and it is enforced in code outside the model. Scoping and the First Win

Handoff

交接

The point at which the customer runs the system without the vendor. It needs three customer-side owners by name (the service, its behaviour, its adoption), a written support window with an end date, and change control they can run. The FDE plans to leave from the first week. Production and Handoff

Handoff drill

交接演练

An acceptance test, not a training session: the customer's operators deploy, break and recover the system from the runbook alone, from their own accounts, while the FDE stays silent. Prepare at least three failures (a dependency down, an input it cannot handle, a bad release to roll back), fix the gaps, and run it again weeks before you leave. Production and Handoff

Held-out set

留出集（留出验收集）

The set the threshold is measured on. The customer's experts label it, the customer owns it, and it is used only on agreed dates; the builders never tune against it. Seed it from the week-1 baseline sample, so the same documents measure the old process and the new one. Scoping and the First Win; Evals: Proving It Works

E.4

I–L

Indirect prompt injection

间接提示词注入

Instructions planted in content the model reads but nobody on the team wrote: claimant emails, text in photos, scanned letters. OWASP ranks prompt injection first (LLM01:2025) and says retrieval and fine-tuning do not fully mitigate it. You cannot promise it will be prevented; you can promise what a successful injection is able to do, and make that small. Security, Governance and Model Risk

Job post (as evidence)

招聘启事

Two uses, kept apart. As evidence of how the work is done, a job post is T4: it is what a company hopes the job is. As evidence of what employers ask for, the same post is primary. The book says which use it is making. How to Use This Book

Judgment domain

判断（领域）

One of the skill map's four domains: knowing what is load-bearing. It means spotting scope creep and data risk early, saying no to a sponsor and making it stick, and calling a failing engagement while there is still time to act. The Skill Map

Kill switch

紧急停止开关

The switch that takes the system out of the workflow. Off must mean "back to the old way", not "broken": for a propose-only agent, suggestions stop and work flows through the old queue. Write into the runbook what in-flight work does and who must be told. Production and Handoff

L1–L4

L1–L4（级别）

The skill map's four levels in each domain, written as behaviour an observer could see. Every L2 involves the customer's real conditions; every L3 involves someone with power who could have said no; every L4 changes something beyond your own engagement. Score a level only when done unaided, twice or kept up across an engagement, with an artefact and a witness. Levels are not titles. The Skill Map

Load-bearing commitments

承重的承诺

The four promises the rest of an engagement depends on: the threshold, the data path, the sponsor and the exit. The test for a request: if you agree to it, which other promise stops being true? Almost everything else (demo dates, model choice, the queue view) is negotiable and should be given away readily. Judgment Under Pressure

Load-bearing conditions

承重条件

The five conditions the rescue-or-walk rule checks, one per recurring failure: a problem the sponsor's business still cares about; a working path to the data; an owner for the result on the customer side; a sponsor who owns the number and still has time for it; a threshold that is agreed, owned and reachable. When Engagements Fail

E.5

M–P

"Not now" list

“暂不做”清单

Where scope requests go instead of being refused. Each item carries a reason and a trigger ("revisit once auto meets threshold"). It tells the asker they were heard, turns "no" into "after the threshold is met", and becomes the expansion plan. An item with no trigger is a promise you have not decided to keep. Scoping and the First Win

Paid pilot

付费试点

A short pilot the customer pays for, with one KPI the sponsor owns, a fixed duration (thirty days is the consensus target), the annual contract agreed in advance, credit on conversion, and milestones with owners on both sides. The book adds a data-access date and a day-1 security-review start. The case for it is the mechanism, not a measured conversion rate. Before the Engagement

pass^k

Anthropic's measure of consistency: "the probability that all k trials succeed", as against pass@k, at least one success in k attempts. Recommended "for agents where consistency is essential" (T4). Run each held-out case several times and report the share that pass every time. Evals: Proving It Works

Post-mortem (engagement)

交付复盘

The template for a failing or finished engagement. Section 1 is copied from the signed documents, including the decision rule, so the rule is fixed before anyone writes what happened. Run it when two failure signals appear, or after the engagement ends. When Engagements Fail

Product domain

产品（领域）

One of the skill map's four domains: making the next engagement cheaper. It means turning what you built for one customer into evidence the product team can act on, and eventually into components others reuse. The book argues this loop is what separates an FDE from a contractor. The Skill Map

Product-feedback brief

产品反馈简报

One finding per brief, written from the field log, checked by the lead for what belongs to the customer, and sent to a named person who owns a roadmap, with a decision asked for by a date. It carries failure categories and costs rather than anecdotes, and says what is not being asked for. Feeding the Product; getting it adopted

Proof of concept (PoC)

概念验证（PoC）

The free, open-ended trial sales often promise. The book's argument is that it is designed not to reach production: nobody on the customer side owns it, no number is agreed, and it has no end date. Palantir's phrase for the alternative is "proof, not proofs of concept" (T2). Before the Engagement

Propose-only

只提建议（不执行）

The agent suggests and a person decides: in the composite, a suggested queue in the adjuster's existing view, with no tools that act. It shrinks the risk without shrinking the stakes, and caps what a successful injection can do at a wrong suggestion. It gives up some of the time saved; authority is raised later, on evidence. Security, Governance and Model Risk; Scoping and the First Win

E.6

Q–S

Read-only identity

只读身份

The credential the agent runs as, allowed to read and nothing else. With propose-only and rules in code, it is one of the controls that is not the model. An instruction in the system prompt is not a control. Security, Governance and Model Risk

Readout

汇报会

The meeting where the number is reported to the sponsor; in the composite, the week-4 readout. When the agent is short, the book's order (practice, not a study) is: the number against the threshold in their words, what passed, the plan by failure category with a re-measure date, what does not move, and a decision for the sponsor. Evals: Proving It Works

Regression suite

回归套件

Every failure you have fixed, kept so that it stays fixed, including seeded injection packets. It runs on every change (a prompt edit, a retrieval change, a model upgrade), and after handoff it becomes the gate on every release the customer ships. Evals: Proving It Works

Rescue-or-walk rule

“挽救还是撤出”的规则

Written into the terms on day one, against the five load-bearing conditions. Continue if all five hold. Reset if each missing one has a named person on the customer side, with authority over it, who commits to restore it by a close date. Walk if any has nobody. Walk from the scope, not the customer. This book's recommendation, not a tested method. When Engagements Fail

Rules outside the model

在模型之外由代码强制执行的规则

A rule the customer calls "never" cannot be met by a percentage, so it sits in code that runs on every request: before the model (known facts decide the route), after it (output checked against a fixed schema and an allowed list), or around the action (the agent proposes; a person acts). Building in Their Environment

Runbook

运行手册

The document the customer's operators run the system from after the FDE leaves. The handoff drill tests it: every time the FDE has to explain something, that sentence is missing from the runbook. Production and Handoff

Sanitised fixtures

脱敏测试样本

Synthetic cases built while data access is pending, from what you can already see. A good set has the real shape, the real mess, and a label saying it is synthetic: fixtures are for building, never for scoring the threshold. They hold only the cases you thought of. Data Access and Integration

Say-no note

说不简报

One request per note, sent within a day of the meeting to the person who asked and copied to the owner of the commitment, kept to a page. It hands the decision to the people who own the approval, so it reads as a statement of fact, not a personal refusal. Judgment Under Pressure

Services collapse

服务式塌缩

The role shrinking to the customer domain alone: engagement work with no technical depth and no product loop. Its mirror image is the technical domain alone, which is AI engineering under the FDE label. Both are respectable jobs; neither is this one. Services revenue is easy to sell, which is why the first collapse keeps happening. The Skill Map

Shadow-run

影子运行

The agent runs on live work, but its suggestions reach no one yet; what people actually did is recorded beside what the agent would have done. In the composite, model risk approved a read-only shadow-run in week 3. It is also the first stage of a staged rollout. Security, Governance and Model Risk; Production and Handoff

Shape check

形态检查

Section 4 of the skill self-assessment: your strongest and weakest domains, whether they are two or more levels apart, and which collapse you drift toward if nothing changes (services or AI engineering). Compare shapes, not sums; never add the four levels into one number. The Skill Map

Side table

旁路表

A store outside the system of record that the customer's team owns, where the agent writes its suggestions and nowhere else. The existing queue view reads from it, so nothing writes into the claims system. It is how write-back is deferred without losing the workflow. Data Access and Integration

Situation index

处境索引

The book's entry by symptom ("accuracy is stuck around 80%", "the sponsor changed jobs"), each row pointing to the chapter and section that deals with it. How to Use This Book

Sponsor

发起人

The customer-side executive who owns the number and the budget, and signs that the metric matters. One of the four load-bearing commitments: an FDE cannot stop a sponsor's priorities moving, but can refuse any arrangement that stops the sponsor hearing the number. Judgment Under Pressure

Staged rollout

分阶段上线

Go-live as a sequence of stages, each small enough to fail cheaply: shadow mode, one team or a small share of traffic, then wider slices, with the old process behind a switch. Each stage's exit rule is written before it starts. A release is everything that changes behaviour (code, prompts, pinned model version, rules), deployed together so it can be rolled back. Production and Handoff

Stated and observed problem

客户说出的问题 / 观察到的问题

What the customer asks for, and what you find when you watch the work. They are written side by side, with the evidence and the lever, and the sponsor signs the observed version. In the composite, "a customer chatbot" was a proxy for "claims take too long". Discovery

Statement of work (SOW)

工作说明书（SOW）

The contract document that defines the work. FDE job posts ask for help informing it; how often FDEs are actually consulted before signature, no public source measures. Before the Engagement

Success metric

成功指标

Written before the first line of agent code, in four parts: a metric the business recognises, a threshold, a held-out set, and a baseline. Pair it with a speed or cost metric and at least one guardrail. Scoping and the First Win

E.7

T–Z

T1（证据等级）

Measured, primary and independent of the vendor. Example: the reporters who tested New York City's MyCity chatbot themselves. Of the ten strongest cases in the book, the only T1-measured outcome is that failure. How to Use This Book

T2（证据等级）

Measured and primary, but reported by the vendor or the customer itself. Example: Palantir's S-1. The T2/T4 line is measurement: a vendor's or customer's own measured result is T2, while its advice, doctrine, stated targets or opinion are T4, even from the same vendor or speaker. How to Use This Book

T3（证据等级）

Secondary reporting: a journalist or analyst relaying what someone else measured or said. Example: The Register on NHS trusts using the Federated Data Platform. How to Use This Book

T4（证据等级）

Anecdote, opinion, advice, doctrine, a stated target, or a job post. Example: OpenAI's FDE posting. A job post is T4 for how the work is done but primary for what employers ask for. How to Use This Book

Technical domain

技术（领域）

One of the skill map's four domains: making it work in their world. It means shipping on the customer's real data, inside their infrastructure, against an eval they accept. A prototype on clean data is the start of this domain, not its centre. The Skill Map

Threshold

阈值

The bar the metric must clear, written as a number or rule with no room for interpretation, and also in the sponsor's words; if they ever disagree, the plain words are what you are held to. Its value is negotiable; its existence is not. Lowering it is the threshold owner's decision, in writing, and never the FDE's suggestion. Scoping and the First Win; Judgment Under Pressure

Threshold owner

阈值负责人

The person whose team absorbs the errors (the claims director, the head of the contact centre, the clinical lead), who signs that the threshold is good enough to use. Not the FDE, and often not the sponsor. If nobody will own the threshold, you have found a workflow with no owner, which is a reason to change the target. Scoping and the First Win

Translation layer

翻译层

The logic layer between the raw data and the model that injects the business meaning access alone does not give you. Colin Jarvis of OpenAI says building it takes most of his FDEs' time getting to production, and calls it "an old-school concept" made useful again because a model now reads it (T4). Data Access and Integration

Trap check

陷阱检查

Section 4 of the before-you-sign question sheet: count, out of five, the answers that point to services. They are: the threshold belongs to someone else; no field-to-product change can be named; the team is measured on utilisation or bookings; variable pay rides on bookings; the only path up is management. Where FDEs Work

Verified box

核实框

A box stamped with the month it was checked, holding facts with a shelf life: pay, headcounts, employer moves, model or product names and policy dates. The argument around it never depends on it: delete the box and the chapter still stands. How to Use This Book

Write-back

回写

The agent writing its answer into the customer's system of record. Harder than reading, and its absence costs efficiency, but a pilot is the wrong place to pay for it. It is earned when four things are true: the threshold is met on real data, the system owner has approved a change, every write can be traced and reversed, and users trust the suggestions. Until then the agent proposes and people act. Data Access and Integration

Zero data retention (ZDR)

零数据保留（ZDR）

A model provider keeping no copy of prompts or outputs. It answers the review's first question ("will our data train your model?") only in part: ZDR at the provider does not cover your own logs, traces, eval sets or backups, and reviewers know it. Answer with terms, never adjectives. Security, Governance and Model Risk

E.8

The Chinese term

The Chinese edition calls the role 前线部署工程师 throughout. How to Use This Book explains the choice once, and this glossary repeats that note as written there:

There is no settled Chinese name for the role. The Chinese edition uses 前线部署工程师, the most common rendering in the press and the title Tencent uses in its job ads. A Beijing municipal policy document of July 2026, 京发改〔2026〕1185号, uses 前沿部署工程师 instead (Beijing Municipal Government, T2), and other outlets use 前置 or, in Taiwan, 前線. One rendering, 前端部署工程师, is misleading: 前端 means front-end, and the role has nothing to do with front-end development. The book says this once, here and in the glossary, and then uses one term throughout.

The source is A note on the term. The other Chinese terms in this appendix are the ones the Chinese chapters use; where a chapter uses two forms, such as 留出集 and 留出验收集, both are given.

Agree the words before the work. Put the threshold, the threshold owner, the held-out set and the "not now" list into the one-page scope in the customer's own words, and point the customer's team to these definitions when the readout starts. An argument about what a word meant is cheaper in week 1 than at the conversion meeting.