Glossary
The Amdahl-specific words the rest of the docs use, defined once each: the learning model, evaluation, learning data, and the API terms behind them
Amdahl's docs use a few dozen words in a specific sense. Each is defined here once. If a page uses one of these without explaining it, this is the page it means.
The learning model
Learning model: Amdahl's model of what wins with a company's buyers. It
processes and structures the company's data, analyzes it, models what wins, and
optimizes an agent's output against that model. Today you reach it through
messages.optimize, messages.optimize_prompt, Search and Research.
Evaluation: a reading of one output against a fixed set of criteria and,
with evidence: "workspace", against your customers' own conversations.
messages.optimize with mode: "grade" returns one as grade. See
What an evaluation measures.
Evaluation model: the model that scores an output, never the model that
picks a rewrite. grade.judge names it and its version. See
Models and limits.
Criterion: one thing an evaluation measures, named in its dimension field:
relevant_positioning, grounding, verified_specifics, differentiation and
cta_clarity. Four of them read your customers' conversations, so they drop out
of the score when your data is not connected.
Learning data: the record Amdahl keeps of each run with your workspace: the request, the versions it tried and how each scored, and the feedback and send reports you record (a send report keeps hashes, not text). It is stored with your workspace. Training your own models on it is a direction, not something you can do today.
Processed customer data: every customer conversation Amdahl holds for your workspace (call transcripts, meeting notes and emails), processed into one searchable shape, speaker-attributed and joined to the deal. When a page says "your own data" it means this, not the open web.
Dimensions (of a message in your data): before any query runs, classifiers tag every message in your processed customer data across 10 dimensions, so a query reads computed facts instead of raw text. See Search.
Outreach
Optimize call: messages.optimize: what you know about the recipient in,
one outbound message out in the sender's own voice, on the same call. A draft is
optional, and it needs no connected data. See
Optimize a message.
Evaluation without your data: the optimize call's default, with evidence
off. The message is evaluated on how it reads and how clearly it asks, and its
claims are not evaluated against your customers. Every response from one carries
a style_pass_only note.
Evidence (on the optimize call): evidence: "workspace", which also
evaluates the message against the workspace's own conversations and lists the
claims nothing in them backs. evidence.used says whether any conversations
were actually read.
Context: what Amdahl writes the message from: facts (who the recipient
is, why you write now, what you offer and the proof you may cite), rules
(hard constraints, never quoted) and voice examples (messages the sender
wrote, matched and not copied). With recipient, Amdahl adds the recipient's
profile and brief from your own data. See Context.
Draft: on the optimize call, the message field. It is optional. Sent with
facts or rules, it is the starting point of the message: what it states may be
stated, and the message keeps its substance, order and voice. Sent with nothing
else to work from, it is improved, with a draft_improved_no_context note.
Needs input: an optimize result (outcome: "needs_input"). Written from
context, it comes only for a call that sent on_missing_input: "ask": the input
could not support the request as asked, so nothing was written. An ask list
holds the questions only the sender can answer. Without ask, Amdahl writes the
best message the input supports and lists what it left out in omitted.
Omitted: on an optimize result, what the message could not do with the
input (a fact a rule needs, a rule it does not fully follow, a missed limit), as
{what, why} rows. why is written for the sender. [] means nothing was left
out.
Tries: every version an optimize run scored, in order, returned as tries
when you send include_tries: true. Each carries the in-loop score, Human tone,
what it aimed at, and whether it came back. Its scores show the path, not proof:
lift is the only before-and-after to quote. See
Reading tries.
Unsupported claim: a claim in a message that nothing in the workspace's conversations backs and no fact the sender sent states. Listed only when the run read those conversations, and kept in the message unless you ask otherwise.
Asking
Primitive: a call you build on: Search, over your processed customer data, and the optimize call, which writes and evaluates messages and needs no data. See Primitives.
Research: one question in and one cited answer out. A read-only agent works the question server-side over your calls, accounts and deals, and returns a handle you poll. See Research.
Coverage: how much of your data a given answer could actually see. Reported on the response rather than assumed, so a thin answer says it is thin instead of looking like a confident zero.
API terms
These words name fields and modes in the API. The API reference uses them.
Grade: the grade object an evaluation returns, from messages.optimize
with mode: "grade". The evaluation model scores five criteria and the call
writes nothing. The score runs from 1 to 5, and passed and total ride
beside it. See Evaluate a message.
Rubric line: the evaluation model's verdict on one criterion: pass, fail, or
null when it was not applicable, with a sentence of reasoning behind it. The
score comes from the criteria that applied (1 + 4 * passed / total). A
criterion that was not evaluated is out of both sides of the fraction: unknown is
never zero.
Prompt run: one run of messages.optimize_prompt, which evaluates a
reusable prompt on the emails it writes for several recipients, or improves it.
It is async: you poll messages.get_prompt_run. See
Improve a prompt.
Tier: what a quote licenses you to claim. account backs "you told us";
segment backs "teams like yours"; corpus backs "the people we talk to say".
Reaching past a quote's tier is evaluated as a grounding failure, not a style
note.
Lift: on an optimize rewrite, a blind re-evaluation of the draft and the
returned rewrite by a different model, present only when the run read your
conversations and a rewrite reached that re-evaluation. See
Reading lift. On a prompt run,
lift is the revised prompt's mean score minus the original's, on held-out
recipients only. The train scores are never lift.
Corpus: the API's name for your processed customer data, as in the corpus
block a Search answer carries.
Warehouse: the queryable, row-shaped view of your processed customer data. Search runs SQL against it; a conversation becomes rows you can filter, count, and group.
Interaction: the warehouse's unit of conversation: a grouped turn, not a whole call. Several things said in a row by one speaker collapse into one interaction row (they average 2.88 atomic utterances), so an interaction is smaller than a meeting and bigger than a sentence.
Grain: what one row of a result actually counts. A query at deal grain returns one row per deal; at interaction grain, one row per grouped turn. The same question at two grains gives two different, both-correct numbers, which is why a count is only meaningful once you know its grain.
Lane: which of the four search strategies ran: filter (typed predicates,
deterministic, no model), fuzzy (plain language turned into SQL),
semantic (meaning-based similarity) or lexical (exact terms, reached by
quoting a phrase). One endpoint, four lanes; mode picks one or auto routes
for you, and mode_ran says which one answered.
Eval: the separate evals product. On 2026-10-22 evaluation moved into the
optimize call and evals ended: an evaluation (mode: "grade") and a prompt run
replace them.
Wire shapes
Envelope: the data wrapper around every successful JSON response. The
field you want is one level in: .data.results, not .results. Two kinds of
response are deliberately not enveloped: event streams and GET /openapi.json.
Handle: the id an async call returns immediately so you can poll for the answer instead of holding a connection open.
Operation: one callable thing on the API, named resource.verb
(search.query, messages.optimize). The tool catalog
lists every one and names the MCP tool for each. Over MCP each one a client can
use is its own tool. See Connect your agent.
See also
- Quickstart: these words in use, against your own data.
- Tool catalog: every operation, generated from the server.