Skip to content

Glossary

The Evals harness's vocabulary, alphabetical. The engine's terms live in its own glossary. Most entries name the file or ledger entry (evidence/2026-09-15-release-one-decision-ledger.md, "the ledger") that holds the meaning.

admitted

A cell with refusals 0 and reaches 0 (confinement.Finding.admitted). Only admitted cells enter a deciding denominator; a hidden basename reached inside the cell's own worktree is reported and does not un-admit (C3 D5).

allowlist (class)

A passing state exists but a file outside source_paths voids the patch (census design §7). Never read from a replayed patch in new work (ledger 2026-10-04, ruling 1).

ambiguity (class)

The prompt admits a reading the hidden suite rejects (census design §7). A task defect: the task is fixed or leaves the set, and nothing is claimed against it.

arm

One measured configuration: the attempt command, the model, and the tool surface. Baseline and Engine are the two arms. An Engine arm file also pins a satyrn-engine commit and seven file digests, and launch refuses a checkout whose HEAD differs from the pin.

attempt command

The executable the harness runs in a cell to do the work: Pi with no product extensions plus the eval's confinement extension for Baseline, the Engine's /implement for the Engine arm (unisolated-harness design §3, C1).

backstop

The wall-clock limit on a cell's command, 4,800 s in the C3 records. A cell the backstop ends is a named wall-clock cut; it stands and is not re-run (ledger 2026-10-03, "C3 nights A and B ran").

budget

The token and turn limits that stop a cell, and the wall-clock backstop. A cell that runs out of budget is graded, not discarded. The run budget (48,000 output tokens / 72 turns) and the declared line (32,000 / 48) are different numbers; see line.

budget (class)

A pass state is reached late and the 32k line falls before it (census design §7).

budget-shaped

A task with at least 3 admitted cells of which at least half are not a pass at the line (C4 pre-registration §5). Every other deciding task is floor.

C0–C4 (Phase C)

The steps that re-derive isolated evidence on the confinement harness: C0 marks it unconfirmed, C1 re-qualifies, C2 re-reads the census process classes, C3 re-runs the census, C4 re-derives the finishing counterfactual (ROADMAP.md, "Next"). All five were done by 2026-10-03.

capability (class)

The cell has the facts, spends its budget and never reaches a pass state (census design §7).

cell

One attempt: a task run once, in its own attempt worktree under the confinement extension, under one arm (ledger 2026-10-02, "C0").

ceiling set

The tasks a comparison may claim on. Ruled at the R0 sitting on 2026-10-04: run-record-gate and preflight-quiet.

census

The classified set of Baseline cells, eight class columns by turn, read before any Engine remedy is built, to see where an arm actually fails. The current one is C3: 36 cells under confinement, all admitted (evidence/2026-10-03-c3-census/).

class column

One of eight per-cell flags (information, ambiguity, capability, budget, finishing, runaway, hunting, allowlist), signed by the maintainer. The primary class is the one whose removal would have changed the verdict at 32k, cited by turn; "pass" means no primary (census design §7).

controller

The attended agent session that dispatches and reviews agents, launches records at the maintainer's word, and writes the ledger. Pushes are the maintainer's.

D-numbered decision

A numbered open decision in a plan (D1, D2, ...) with lettered options and a recommendation. The maintainer rules it, and the ledger records the ruling with the evidence it rests on.

deciding

A record or cell whose result can support a claim: purposes admission, route-proof and campaign (run_record.DECIDING_PURPOSES), with the rule pre-registered.

development

A record purpose with no task outcome in its decision rule. Development cells are read beside, never pooled, and never support a claim (ledger 2026-10-04, "Engine pin moved").

EB0–EB3 (Phase EB)

The engine-budget phase: EB0 development cells on floor tasks, EB1 an offline read of them, EB2 remedies each with an offline estimate, EB3 a pre-registered floor read (engine-budget design §3, branch worktree-engine-budget).

Engine terms

Guards, the finish nudge, the scope guard, the red-stop gate, edit parity and the Engine's confinement root are Engine vocabulary, defined in the engine glossary.

fidelity note

A fidelity: reason on a classifier row whose replay skipped a bash writer, so the replayed tree may differ from the cell's. C4's rule withheld such rows as unmeasured (C3 census page, "Replay artifact on allowlist").

finishing (class)

A pass state is reached within 32k and the cell keeps working past the line (census design §7).

flagged

A cell with a refusal or a reach. It is reported, is not a pass, and stays outside the denominator (unisolated-harness design §3, C3).

floor

A deciding task that is not budget-shaped; it is in scope for harm (C4 pre-registration §5).

frozen

Fixed before any cell runs: a run record is committed first, and a sitting runs from a named commit recorded in the ledger as "C3 frozen <sha>".

go, verify, not-the-lever

C4's three verdicts. Go needs at least 2 qualifying tasks, floor harm under 2 and no floor task insufficient; verify needs at least one qualifying task; not-the-lever means no budget-shaped task has net 1 or more (C4 pre-registration §5). C4 read verify.

grade-line

satyrn-evals grade-line (src/satyrn_evals/line_grade.py), the at-line reader: it grades each finished cell's line.diff at the crossing on both arms and uses the harness verdict for a cell that never crossed. It reports an unavailable cell apart from the pass denominator.

hunting (class)

A root search or a bounded command precedes the failure (census design §7). Mechanical under C2's rule, no override.

information (class)

The prompt omits a fact the hidden suite requires, and no cell finds it (census design §7). A task defect, never claimed against.

instrument-only piece

A piece of work that changes how things are measured and runs no measurement. Two consecutive instrument-only pieces stop the loop (AGENTS.md).

insufficient

A task with more than 1 unmeasured admitted cell, or fewer than 3 admitted cells (C4 pre-registration §5).

isolation (retired)

The two-uid satyrn-cell and bwrap profiles. Retired at 35c298d for confinement; the user and its sudoers rule were removed on 2026-10-03.

launcher

The satyrn-evals launch command that runs a frozen record's cells, arms interleaved, stopping on infrastructure failure and resuming a stopped night. On a capped exit (4) the sitting stops and the next sitting resumes the record.

line

The declared read point, 32,000 output tokens / 48 turns, inside a 48,000 / 72 run budget (line_token_budget, line_turn_budget). A record must declare it for its cells to be read at the line.

line harvest

The patch the launcher writes as line.diff when a cell crosses the declared line, without disturbing the cell.

net, rescue, harm

Rescue: actual not-pass, counterfactual pass. Harm: actual pass, counterfactual not-pass. Net per task is rescues minus harms, on admitted cells (C4 pre-registration §4).

night

One unattended run of a set of frozen records through the launcher, named A, B, and so on; the launcher writes one ledger per night.

oMLX

The local model server the arms call; preflight checks its settings.

Ornith

The model, Ornith 1.5 9B, served by oMLX as Ornith-1.5-9B-MLX-8bit.

outside

Computed and reported, never deciding; speed-probe sits outside C4's decision for its prompt ambiguity.

pass at the line

The cell's own tree at the line: its line.diff when it crossed the declared line, the harness verdict when it ended inside it; one reader and one grader for both arms (ledger 2026-10-04, ruling 1). The reader is grade-line.

patch

The cumulative diff from the workspace's base commit, harvested after the attempt — untracked files included, so a model git commit hides nothing.

Pi

The agent CLI both arms run. The Engine runs its own inner Pi inside /implement.

pre-registration

The decision rule, denominators and verdict table, approved and committed before any cell it reads exists. A judgment made after the read is what it forbids (ledger 2026-10-03, C4 D9).

R0 sitting

The attended sitting, at most 60 minutes, in which the maintainer alone fixes the claim row, ceiling set, budget, win rule and power (census design §8). Ruled 2026-10-04.

reach

A file-tool path or bash argument that resolves to the corpus root, a task directory, a hidden basename or a fixture, found by the post-hoc reach audit (unisolated-harness design §3, C3). A hidden basename inside the cell's own worktree is an in-worktree reach: reported, never counted for admission.

re-cut

Rebuilding a task's base from its source. C1's re-cut 7fc679f removed the nested selfhost task trees from three bases and changed their task tree digests.

refused

A cell with a confinement_refused entry: the extension blocked a file-tool path outside the worktree or a bash command naming a protected root. Not admitted.

replay

The classifier's reconstruction of a cell's tree at a turn. Diagnostic only (class columns, pass turn), never a deciding "pass at the line" (ledger 2026-10-04, ruling 1).

root search

A bash search whose path argument lies outside the cell's worktree, such as / or the temp root (cell_evidence.root_search, which sees through timeout, env and similar prefixes since c9ce8e5).

run 1, run 2

The classifier's two counterfactual readings: run 1 applies the 2026-09-15 §3–§5 rules with amendments 7.1–7.4; run 2 grades the trigger-turn tree by replay. Run 2 decided C4, with the fidelity amendment; run 1 is printed beside.

run record

The frozen JSON that names the task, arms, model, budgets, schedule, and the decision rule, committed before any cell runs. Its purpose is admission, route-proof, campaign or development.

runaway (class)

A length-stop turn precedes the failure (census design §7).

self-stop

A cell that ended by itself, before tripping the budget or a cut (self_stop_tokens, self_stop_turn in cells.json).

sitting

Two senses. A launcher sitting is one start of a record's slots, listed in the result's sittings. A decision sitting is an attended session in which the maintainer decides, such as the R0 sitting.

slot

One of a record's n cells per arm. A finished slot never runs again; a slot lost to infrastructure is replaced (launch.py).

task

One self-contained piece of work: a base commit, a contract, a hidden oracle test suite, and a public feedback surface. Captured by satyrn-evals capture.

transcript

The model's stream, preserved byte-for-byte, from which tool calls and usage are read.

trigger turn

own_green_turn: the turn of the first green test run after the first source edit, counted only inside the line (finishing-counterfactual spec §3). The counterfactual stops the cell there.

UNCONFIRMED, SUPERSEDED, KEPT

Ledger marks. UNCONFIRMED (<sha>): evidence from a harness later found defective, kept as evidence but built on by nothing until re-derived. SUPERSEDED: replaced by a re-derivation. KEPT: the mark stands after a re-derivation that did not cover it (ledger 2026-10-02, "C0"; C4 D9).

unmeasured

A cell the read cannot score, listed with its reason and never dropped. In confinement, a cell with no finding; in C4, a fidelity: or raised: row, counted as no change; at the line, a crossed cell with a harvest error or no line patch, counted as Baseline-favouring (ledger 2026-10-04, ruling 1; grade-line reports it apart, and the pre-registration picks one).

verdict

The outcome of offline grading by the oracle test hook — pass or fail, never an exit code or stdout.