Software factory glossary
25 terms, one sentence or two each. Every definition links to the page it comes from.
- Agent harness
-
The software layer that surrounds a model with instructions, context assembly, tools, state, permissions, control flow, budgets, verification, observability, and recovery.
Source: darkfactory.dev/glossary/agent-harness ↗ - Autonomy ladder
-
A policy of authority levels per consequence class: A0 explicit approval, A1 a prepared consequence with one judgment, A2 evidence-earned pre-authorization for reversible actions.
Source: contextstack.com/docs/autonomy-ladder ↗ - Bounded run
-
One bounded execution attempt commissioned to advance a piece of work, with a capability envelope, budget, external acceptance and a recorded terminal state.
Source: contextstack.com/docs/bounded-runs ↗ - Consequence authorization
-
Permission to merge, deploy, publish, send or spend, granted separately per lane and action class rather than to a system as blanket trust.
Source: contextstack.com/docs/autonomy-ladder ↗ - Cost per accepted outcome
-
Total cost, including retries and human intervention, divided by outcomes that passed acceptance at equal or better verified quality. The factory's honest unit cost.
Source: contextstack.com/docs/trace-ledger ↗ - Dark software factory Also: dark factory, lights-out software factory
-
A domain-bounded software production system in which humans specify intent, risk, and policy while a model, harness, and execution environment plan, build, verify, ship, observe, and repair software with little routine human intervention.
Source: darkfactory.dev/glossary/dark-software-factory ↗ - Ephemeral cell Also: execution cell
-
A disposable, deny-by-default execution environment created for one attempt at one work order, granted only that order's capabilities, and destroyed afterward.
Source: contextstack.com/docs/bounded-runs ↗ - Gate Also: verification gate
-
A control point outside the attempt being judged that blocks promotion until required evidence has been produced and validated. It fails closed when evidence is missing.
Source: darkfactory.dev/glossary/verification-gate ↗ - Gate wall
-
The full set of acceptance mechanisms for a lane, kept outside the execution that attempted the work. A run can produce evidence; it cannot certify its own outcome.
Source: contextstack.com/docs/gate-wall ↗ - Human on the loop
-
A supervision model in which people monitor an automated system and intervene on exceptions instead of approving each step.
Source: darkfactory.dev/glossary/human-on-the-loop ↗ - Independent verification
-
Checking an outcome with evidence, components, context, or authority meaningfully separated from the system that produced it, to reduce correlated error.
Source: darkfactory.dev/glossary/independent-verification ↗ - Lane
-
A class of work with its own order format, gate wall and authority policy, such as small changes to one application. Factories earn trust one lane at a time.
Source: contextstack.com/docs/gate-wall ↗ - Ledger Also: trace ledger
-
The append-only record of every state transition and execution attempt, from which ownership, routing, authority, evidence and cost can be reconstructed without a transcript.
Source: contextstack.com/docs/trace-ledger ↗ - Promotion
-
The governed decision to move an artifact or change into a more trusted lifecycle state, such as merging, releasing or deploying, after required evidence and policy checks.
Source: darkfactory.dev/glossary/promotion ↗ - Requeue with feedback
-
Returning a red-gated attempt to the order with its failure output attached as context for a fresh attempt in a new cell, within the order's retry budget.
Source: contextstack.com/docs/gate-wall ↗ - Review deck
-
A compact per-order summary for the human decision: the order, the change, the gate verdicts, the cost, and the proposed consequence with its reversal plan.
Source: contextstack.com/docs/autonomy-ladder ↗ - Reward hacking
-
Satisfying a visible proxy for success, such as a test or score, without achieving the intended outcome.
Source: darkfactory.dev/glossary/reward-hacking ↗ - Risk-scoped autonomy
-
Granting an automated system authority in proportion to the risk and reversibility of the specific action, rather than uniformly.
Source: darkfactory.dev/glossary/risk-scoped-autonomy ↗ - Run-autonomy
-
Permission for bounded work to proceed without a person watching each step. It does not include permission to create lasting effects.
Source: contextstack.com/docs/autonomy-ladder ↗ - Run contract
-
A machine-readable, human-auditable agreement defining one agent run's objective, authority, evidence, limits, and recovery path.
Source: darkfactory.dev/glossary/run-contract ↗ - Runner adapter
-
The thin layer that translates a work order into what a specific agent runtime needs and normalizes its output into runtime-neutral events: work order in, events out.
Source: contextstack.com/docs/bounded-runs ↗ - Skill document Also: agent skill
-
A versioned package of instructions, procedures, examples, and sometimes code that teaches an agent how to perform a repeatable class of work.
Source: darkfactory.dev/glossary/agent-skill ↗ - Software factory
-
A repeatable production system that turns software demand into accepted, operated software through reusable processes, tooling, controls, and feedback.
Source: darkfactory.dev/glossary/software-factory ↗ - Trace
-
A captured sequence of model calls, tool calls, events, timings, state changes, and outputs from one execution.
Source: darkfactory.dev/glossary/trace ↗ - Work order Also: order
-
The admitted unit of work in a factory: a human-readable contract stating one objective, its scope, its definition of done, and its limits on time, spend, retries and allowed effects.
Source: contextstack.com/docs/bounded-runs ↗