Glossary
The words on this site, defined.
Every term Eliyce uses for delivering software with AI, in one or two sentences each, as the framework itself uses them. The process first, then the record, then the verdicts, then the names.
What these terms cover
Delivering software with AI inside an enforced process has its own vocabulary: the six moments a change travels, the parts of a delivery record, the three verdict words and the three strengths of evidence. Each definition follows how Paqad, the open-source framework Eliyce builds with, uses the term in its own records.
The process
The moments a change travels, and the words used at each.
Six moments
The six stages every change travels on this site and in Paqad’s record: requested, agreed, built, reviewed, checked and live. The same six for a Programme team and for Eliyce’s own engineers.
Requested
The first moment. The request is written down in the requester’s words and fetched from the tracker, so the work is grounded in what was asked rather than in what someone remembered.
Agreed
The second moment. The request becomes a frozen specification with acceptance criteria, invariants and non-goals, signed off by a person before any code is written.
Built
The third moment. The code is written under the team’s rules, which are loaded on record first, and every file save passes the same checkpoints.
Reviewed
The fourth moment. Someone who did not write the change reads a digest of what the machine already proved and records findings, a rollback plan and a verdict labelled as a judgement.
Checked
The fifth moment. The project’s own format, build and test commands run and their exit codes decide; an independent program then grades the whole record.
Live
The sixth moment. After a person approves, the change goes live with its record attached: the receipt, the inventory of authorship, and every commit linked to the request.
Specification freeze
The point at which the specification is fingerprinted and signed off. From then on it is the contract: if the goal changes, the work stops and asks rather than drifting.
Acceptance criterion with proof method
A numbered statement of what the change must do, paired with how it will be proven: by an automated test or by a person checking. Paqad refuses a criterion without a proof method.
Invariant
A property the change must not break, confirmed by a person at the freeze. In Paqad’s change #557: one canonical composer for the rule text, and only an absent record blocks.
Rules loaded on record
The requirement that the team’s rules applying to the files in play are loaded, and the loading recorded with a fingerprint of the rule text, before the AI’s first source edit. The record attests loading and acknowledgement, not comprehension.
Checkpoint
A small program the coding tool runs at fixed moments: session start, each message, before every file save, at the end of a turn, after a commit or merge. The AI cannot skip, reorder or argue with it.
Decision packet
The written form of a decision the AI may not take alone: the options, the choice a person made and the reason. Paqad’s own repository held 54 resolved packets on 17 September 2026.
Rollback plan
The part of a recorded review that says how to undo the change: which files and entries to revert, and whether any data would need migrating.
Those are the moments. Here is what each one writes down.
The record
What is written down, and by what.
Delivery record
One folder per change, also called the evidence bundle, written by scripts while the work happens: the plan, the frozen specification, the rules loaded, the stage timeline, the review, the checks, the receipt, the inventory and a report. Paqad’s change #557 holds 17 files.
Ledger
The set of files on the developer’s machine in which Paqad records every stage, edit and verdict. It is kept out of version control by default and shared only when the team chooses to.
Checks file
The structured result of the checks stage: each command, its exit code, its duration and, where the runner exposes them, per-test results. The verdict reads this file, not anyone’s description of it.
Receipt
The sealed verdict of a change, an in-toto statement inside a DSSE-shaped envelope, hash-chained to the previous receipt of the same change. Its integrity recomputes from its own bytes: tamper-evident, not a cryptographic signature.
AI bill of materials
An inventory of a change in the CycloneDX 1.6 format naming which AI agent wrote it, which person accepted it, and the graded result. A standard supply-chain document, applied to authorship.
Independent verifier
A separate program, run by the coding tool at the end of a turn, that reads only what is on disk and grades every gate. The AI’s chat is not an input. In change #557 it refused to close the change six times until the review was recorded.
The record ends in a verdict. Six words say how it was reached.
Verdicts and evidence
Three words for the outcome, three for how it was known.
Safe to merge
The verdict when every gate Paqad ran passed. It attests the gates, not that the change is beyond all error, and Eliyce never uses it as its own assurance that software is correct.
Needs your attention
The verdict when a gate is blocking. The receipt names the gate, and the change cannot close until it is resolved.
Inconclusive
The verdict when at least one gate could not reach a confident result. “Could not verify” never rounds up to “verified”. Change #557 reads Inconclusive because three quality measures could not be taken.
Deterministic
Evidence a script proved: a file exists with this fingerprint, a command exited 0, these rules matched these files. Labelled as such in every gate row.
Model-judged
Evidence a language model assessed, such as the specification review or the implementation review. Labelled as judged, so it never borrows the credibility of a deterministic result.
Not measured
A measure that could not be taken this run, with the reason recorded, for example “tool not wired”. Reported as not measured, never as passed.
Meaningful green
A mark in the checks file that the test command actually ran tests. A chained command can no longer report green without running a single test.
Flaky quarantine
What happens to a test that fails when run in parallel but passes alone: it is quarantined and recorded as flaky rather than blocking the change or being retried in silence.
Near-duplicate check
A comparison of new code against existing code for near-copies of eight lines or more at a 0.90 similarity threshold. In warn mode it flags for a person; in strict mode it blocks.
Reuse rate
Calls into existing, untouched code per hundred changed lines. Change #557 recorded 16.4, with 162 reuse calls across 985 meaningful changed lines.
And the names.
Eliyce
The names, and what each one is.
Paqad
The open-source framework Eliyce maintains and builds with, MIT licensed and on npm since March 2026. It enforces the delivery method around an AI coding assistant and writes the record.
Paqad Spine
The shared team view that collects the records Paqad writes on each developer’s machine. Hosted by the team; it comes with the AI Delivery Programme.
AI Delivery Programme
Eliyce’s five-week engagement inside an engineering team’s own codebase until tests, rules and review run on every AI-assisted change, with a record for each.
Build with us
Eliyce’s software development offer: capacity inside a team’s process, or a build a founder can check before accepting it, delivered inside the same six moments.
Key facts
Three numbers behind the words
6
moments every change travels, the same on this site and in the record
Change #557, 12 September 2026
3
verdict words, the same in the chat, the pull request, the report and Spine
Paqad framework, verified 17 September 2026
17
files in the delivery record of one change, each with a defined writer
Change #557, 12 September 2026
Our services
Three ways we make AI work you can check.
Questions
Questions about the words
Because nothing failed. Inconclusive means at least one gate could not be measured this run, and the framework refuses to round a missing measure up to a pass. A failing gate reads Needs your attention.
Deterministic evidence was proved by a script: a fingerprint matched, a command exited 0. Model-judged evidence was assessed by a language model, such as a review. Every gate row carries its label, so one never masquerades as the other.
No. The receipt is hash-chained and tamper-evident: its integrity recomputes from its own bytes and any edit afterwards breaks the chain. It is not a cryptographic signature, and Eliyce does not call it one.
Last reviewed 13 October 2026 by Eliyce.
Hear these words used on your own backlog.
Tell us how your team ships today. You leave with an honest view of whether the programme fits, whether or not you work with us.
We reply within one working day. Or write to hello@eliyce.com.