Glossary

The Amdahl-specific words the rest of the docs use, defined once each

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.

Rewriting

Optimizer — the Message Optimizer, messages.optimize: one outbound draft in, a stronger version out in the sender's own voice, on the same call. It needs no connected data. See Message Optimizer.

Style pass — the Optimizer's default. The draft is judged on how it reads and how clearly it asks, and its claims are not checked. Every response from one carries a style_pass_only note.

Evidence (on the Optimizer) — evidence: "workspace", which also grades the draft against the workspace's own conversations and lists the claims nothing in them backs. evidence.used says whether any conversations were actually read.

Kept draft — an Optimizer result where nothing it tried beat the draft, so the draft came back unchanged (unchanged: true). A real result, not a failure; kept_suggestions lists the edits the sender can make instead. A draft too weak to rewrite is kept too, with an ask list of questions only the sender can answer.

Exact details — names, numbers, dates, links and merge fields such as {{first_name}}. The Optimizer returns them as written and discards any rewrite that changes one.

Context — what the sender gives the Optimizer to respect: rules (hard constraints, never quoted), voice examples (messages they wrote, matched and not copied) and facts (what they know first-hand, which no recording holds). See Keep your voice, rules and facts.

Tries: every version an Optimizer run scored, the draft first, returned as tries when you send include_tries: true. Each carries the Optimizer's own 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 draft 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.

The data

Corpus — every customer conversation Amdahl holds for your workspace: call transcripts, meeting notes, and emails, normalized into one searchable shape. When a page says "your own corpus" it means this, not the open web.

Warehouse — the queryable, row-shaped view of that corpus. 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.

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.

Asking

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.

Primitive — one of the two calls over your corpus that the orchestration layer is built on: Search and Eval. The Optimizer is the third call, and the one that needs no corpus. See Primitives.

Chat — the orchestration layer. You hand over a question that needs real investigation, and Amdahl runs an agent server-side that calls Search, Eval and (when allowed) outside market search, and returns handles you poll.

Master — the general-purpose agent that runs a Chat turn. It has access to the warehouse, customer-voice themes, and — when enabled — outside market search. A named agent from the library is a Master turn running that agent's prompt.

Agent — a saved, reusable prompt your workspace can dispatch by name. Not a separate engine: the same Master turn, specialized.

Routine — a cron that fires a fresh Chat on a schedule. A standing weekly sweep, a daily pipeline pulse.

Grading

Eval — a configured grading pipeline: the inputs it accepts, the cases it grades, and the graders that score each one. prompt-and-message-eval is the built-in one.

Run — one execution of an eval against one submission. Runs are immutable once finished, so a run id is a stable thing to cite.

Dimension — one axis a submission is graded on (relevance, specificity, evidence, and so on).

Rubric line — the judge's verdict on one dimension: a binary pass or fail with a sentence of reasoning behind it. One per dimension, which is why the score is a count of dimensions passed rather than an average of opinions — checks_passed / checks_total is that fraction.

Gate run — a run that grades only what you submitted and stops: no rewrite, no improved version, no suggestions. Cheaper and faster than the full loop, and the right mode when you only need a pass/fail before sending.

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 graded as a grounding failure, not a style note.

Evidence pool — the quotes retrieved for a run, before the judge saw them. Bigger than the handful the report cites; the counts strip on a report opens it.

Lift — on an eval, the improved side's score minus the submitted side's, where both halves come from the same judge call (before_paired). It is not the gap between the two dials on a report: those are graded under different conditions, which is why the server computes lift rather than leaving you to subtract. Read the caveat on Evals before building anything on it — lift mostly measures how low the submitted side scored, so the largest lifts come from the weakest drafts. The Optimizer's lift is a different measurement: a blind re-grade 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-grade. See Reading lift.

Wire shapes

Envelope — the data wrapper around every successful JSON response. The field you want is one level in: .data.results, not .results. Four responses are deliberately not enveloped — the .csv and .jsonl file downloads, event streams, and GET /openapi.json.

Handle — the id an async call returns immediately so you can poll or stream for the answer instead of holding a connection open.