Every term the Vmax workspace uses, defined once: one word, one meaning. Documentation, hints, and skills link to a term here instead of restating it, so a definition changes in one place.
| Term | Definition |
|---|---|
Action timeout | A setter or solver setting, in seconds: the limit for any single command an agent runs inside its sandbox, so one hung command cannot spend the whole agent timeout.See alsoAgent timeout, Sandbox |
Agent timeout | A setter or solver setting, in seconds: the wall-clock limit for one whole trial. For the setter it bounds authoring a task; for the solver it bounds one attempt, and too short a limit measures speed instead of ability.See alsoTrial, Rollout, Trial configuration |
AppContainer | The shell component every tools page renders: it draws the navigation, the sidebar groups it is given, and the toolbar around the columns a caller passes.See alsoTools page, Form kit |
Atlas | Every task from every campaign an account can read, in one place. The API keeps no cross-campaign index, so the Atlas reads one campaign at a time and merges what comes back, owned campaigns first and the read-only examples after.See alsoTask matrix, Example campaigns |
Campaign | Vmax's self-play task generation product: the Campaign API, the workers that run the loop in sandboxes, and the /campaigns application. It turns proprietary data and evals into new environments and refines agents on new examples of the tasks they are meant to perform. Each run of it is a campaign instance.See alsoCampaign instance, Setter, Solver, Task |
Campaign agent activity | The audit trail of a run: one row per event the campaign agent recorded while it ran the loop, from building the task setting to dispatching solver rollouts. It is what makes every number on a campaign page defensible.See alsoCampaign instance, Reproducibility |
Campaign application | An interactive view of an existing campaign or part of its evidence. Campaign Applications appears between Campaign Summary and the campaign table of contents. The rollout explorers show steps, instructions, branches, and outcomes. The Campaign Rollout File Explorer browses retained trial files across the campaign. Only one app is open at a time and it fills the content column while preserving the underlying campaign. Selecting the open app again returns to the campaign, and selecting a different app replaces it.See alsoCampaign, Rollout explorer, Campaign Rollout File Explorer, Campaign Terminal View, Campaign Conclusion Timeseries, CLI view |
Campaign client error | An answer from the Campaign API that the campaign page did not expect: a value outside the documented contract at a JSON path, stored evidence that failed the API’s own integrity check, a refused request, or no JSON answer at all. The campaign page keeps a ledger of them, with what was expected and what arrived, so the gap between the API the page was built against and the API it is talking to is visible.See alsoStored evidence, Campaign instance |
Campaign Conclusion Timeseries | A Campaign application that compares recorded work progress, cumulative solver solve rate, and the share of measured tasks in the frontier band over elapsed campaign time. One chart shares a cursor and time window across selected curves. Task schedule, start, and conclusion times remain available beside trial registration and conclusion times. Missing timing and unmeasured results stay explicit. Campaign Conclusion Timeseries v2 reads the same timeline and presents it in the layout of the Campaign page.See alsoCampaign application, Frontier band, Task, Trial |
Campaign data | Everything the Campaign API answered for a campaign, as downloadable files: the immutable record, the compact task rows, every worker event with its payload, the retained trial snapshots and log rows, and the inventory of stored evidence.See alsoStored evidence, Reproducibility |
Campaign instance | One self-play run of Campaign. It takes a single INSTRUCTION.md, authors a pack of tasks in that domain over the number of setter iterations requested, and measures a solver against every task. An instance is immutable once created; its record is what every page reads back.See alsoCampaign, INSTRUCTION.md, Iteration, Campaign status |
Campaign Rollout File Explorer | A Campaign application for browsing retained files across campaign preparation and task trials. Task and trial filters narrow the collection, and every capture keeps its source identity. Selecting a file reads its retained contents, and choosing a baseline compares two captures. These artifacts do not establish the filesystem state at every rollout step.See alsoCampaign application, Trial, Rollout explorer |
Campaign secret | A private environment variable supplied at launch and delivered only to its allowed campaign roles. Values are write-only and excluded from saved settings and exported requests. A new campaign requires the values to be entered again.See alsoSecret permission, Setter, Solver |
Campaign settings | The run-level parameters that are also the design of the evaluation: how large the task pack is, how many setter rounds produce it, how many rollouts stand behind each solve rate, the concurrency, and how long the whole run may take. Setter and solver settings bound each side's trials.See alsoIteration, Rollout, Runtime image, Trial configuration |
Campaign status | Where a campaign instance is: building while the loop is authoring and measuring; completed when every iteration ran and the loop ended on its own; failed when the worker exited, timed out, or lost its sandbox; stopped when it was cancelled on request. Failed and stopped campaigns keep every rollout already graded, and neither resumes.See alsoCampaign instance |
Campaign Terminal View | A Campaign application that presents the existing campaign results as the expandable CLI view. It uses the same loaded document and request handlers, with Collapse all, Expand all, and Standard view controls. Selecting it again, or selecting any other Campaign application, turns it off.See alsoCampaign application, CLI view, Rollout explorer |
Campaign trials | The trials of a campaign that no task owns: the certification trial that proved the task pack, a setter trial whose task has not landed yet, and any trial whose parent is unknown. A task's own setter and solver trials are listed with the task instead.See alsoTrial, Certified task pack |
Certified task pack | A task setting whose code, environment, and grading checks have passed certification. The campaign records the pack's identifier and a checksum that identifies the certified contents. You can start a new campaign from an existing certified pack you own. Certification checks the task setting, while solve rates come from the solver rollouts run against its tasks.See alsoTask setting, Smoke trial, Campaign trials |
CLI view | A terminal-style presentation of Campaign results in the content column. It uses one monospace text size, two-character indentation, and bracketed actions. Loaded sections and text start expanded, with every available role of the selected task shown in sequence. Each section opens and closes independently with [+] and [-], forming an indented tree. Collapse all and Expand all control the whole tree without fetching more data. Actions that can fetch additional data use the primary theme colour. Campaign Terminal View in Campaign Applications opens this presentation when results are available. Standard view restores the regular presentation while keeping the same campaign and selected task.See alsoCampaign, Task instruction |
Composer | The page on /campaigns where a campaign is written before it is launched: the prompt field for the INSTRUCTION.md, the Generate, Refine, Import, and Download actions beside it, and the settings panel on the right. Nothing is paid for until Run campaign is pressed.See alsoINSTRUCTION.md, Campaign settings, Campaign instance |
Compute timeout | The total time budget for a campaign attempt, in seconds. The worker reserves time for saving evidence and cleanup inside that budget. Reaching the compute deadline cancels unfinished work while completed evidence is retained.See alsoCampaign settings, Campaign status |
Concept action | A possible future Campaign action shown beside an output in CLI mode, labelled CONCEPT: before its action name. Activating it reveals exploration options or a short explanation of what an agent could discover, compare, or propose using that output. It sends no Campaign request, changes no campaign data, and launches no agent.See alsoCLI view, Campaign |
Concurrent trials | The campaign setting that caps simultaneous trial execution. Generation and its child trials share that capacity, and the API also checks the concurrency requested across an account's active campaigns. More capacity permits overlap when independent work is ready.See alsoSandbox, Trial, Campaign settings |
Context token budget | A limit used by the agent harness for both the growing conversation and cumulative reported token usage across model responses. Usage includes input and output, and carrying earlier conversation into another request contributes again. It is separate from the output cap on one response and the model's context window.See alsoToken, Max output tokens |
Context window | The amount of input and output a model can handle in one request. A long-running agent also has cumulative usage across requests, so a trial token budget and a model context window bound different things.See alsoToken, Context token budget, Max output tokens |
CPUs | The processor cores requested for each setter or solver sandbox. Commands that can use several cores may finish sooner with a larger allocation. This changes sandbox compute resources rather than the model itself.See alsoSandbox, Trial configuration |
Crystal agent | The rigged rabbit character used by the island scenes, walking agents, and crews. Its egg-shaped body has no tail, its round head blends softly into a small muzzle and full cheeks under two plush, backswept rabbit ears, and its legs end in flat-soled feet. Its hands hang open at rest and close into fists around whatever it holds. It has an eyeless face, one opaque body color, and shared procedural poses. The Crystal agent name remains the scene and asset identifier; the help portrait is a separate character.See alsoShowcase, Hint avatar |
CrystalChess | The island scene with white and black Crystal Agents representing two chess armies. Equipment identifies each piece, and each knight consists of a wagon rider and an unarmed puller. The base scene holds the opening formation; a playback mode makes the same pieces play a recorded game.See alsoCrystal agent, Showcase, Preset |
Derrick | A braced lifting machine with a mast, swinging jib, winch, and pulley ropes. The quarry derrick uses articulated pincers to grip a block at the stone face, pull it free, swing it above the output pile, and open to release it.See alsoQuarry, Stone block |
Difficulty mix | The spread of solve rates across a campaign's measured tasks, from attempts that all passed through mixed results to attempts that all failed. You can inspect that spread task by task in the Atlas matrix. The live bar under the generated task difficulty mix caption counts measured, rejected, and unmeasured outcomes. Those outcomes tell you which tasks produced measurements, while the solve rates tell you how the solver fared.See alsoFrontier band, Workload analysis |
Distant islands | The home-page ring of islands out on the open sea, built from the same island presets as the centre and seated on one flat water level over a shared sea floor.See alsoIsland, Preset |
Document trigger | A content rule that selects a post's scene from its body: the image count, whether it has a blockquote, LaTeX, or a code block, and certain words decide the island, its structures, and its mode. Every island doc names the trigger that produces it.See alsoworldType, Island, Post |
Entity | A unit of scene behaviour mounted by the world builder: it publishes what it exposes, subscribes to what it needs, is gated by the manifest, and runs its own per-frame update. Ocean board, terrain, castle, fleet, forest, and volcano are entities.See alsoMount manifest, Island |
Environment | The sandbox a trial runs in: the runtime image, the task files placed on disk, the dependencies installed from them, and the network policy. A task is not only a question; it is a place the agent works, and the environment is that place.See alsoTask files, Runtime image, Trial, Task |
Example campaigns | Campaign instances Vmax ran and certified, listed read-only for every account. Each is a real finished loop with its task pack, trials, and environment files, so the product can be understood from a result before anything is spent.See alsoCampaign instance, Atlas |
Form kit | The Form* components in next-vmax-tools (tables, cards, inputs, disclosures, hints, typography) that every tools page body is composed from, so every page reads as one system.See alsoTools page, Hint |
Frontier band | The measured tasks in the Atlas focus range of twelve through fifty percent solver success, including both boundaries. Their graded rollouts contain both successes and failures, giving concrete evidence to compare near the solver's current limits. Flame cells highlight this band using successful rollouts divided by measured rollouts. The highlight identifies tasks to inspect and does not change how a campaign generates later tasks.See alsoSolve rate, Difficulty mix, Workload analysis |
Glossary | This page: one definition per term the workspace uses, and only one. Docs, hints, and skills link here instead of redefining a term; a new term is added here first, and a word never carries two meanings.See alsoHint |
GlyphBenchFactory | The island scene for the GlyphBench paper: a small round grass island framed close around one three-storey timber building, braced and roofed, at the size of the island's own Crystal Agents. The ground floor typesets a glyph world from a type case and lifts it to the middle floor, where it completes the world and an agent plays it while every move rises as a key; a hoist carries the keys to the top floor, whose press stamps each one into a glyph that falls down a chute back to the ground floor. An agent cranks a geared winch that pulls the hoist, and counters on a balcony outside each floor count tasks, turns, and glyphs.See alsoShowcase, Preset, Document trigger, Crystal agent |
Harness | The agent program that coordinates model calls and tools during a trial. Campaign supports mini-swe and Codex for setter and solver trials, subject to the selected model's supported harnesses.See alsoModel, Trial, Setter, Solver |
Hint | The (?) affordance beside a heading or label that opens a card explaining what the thing is, with a diagram, an optional table, the reading avatar, and links to the glossary terms it uses.See alsoHint avatar, Glossary |
Hint avatar | The portrait beside every hint card and at the top of the glossary. Its READ button sends the text to the speech endpoint and reads it aloud, lighting each word as it is spoken.See alsoHint, Glossary |
Image digest | A content-derived identifier for a particular container image. It identifies fixed image contents, while a tag is a name that can be updated to point to another image.See alsoRuntime image, Reproducibility |
Instruction | The text an agent is handed to act on. A campaign has one, the INSTRUCTION.md that describes the kind of task wanted; every task has one, the task instruction that states one concrete problem. The setter turns the first into the second.See alsoINSTRUCTION.md, Task instruction, Setter |
INSTRUCTION.md | The prompt a campaign is launched with, exactly as submitted in the composer: the domain, what a good task looks like, the deliverable a solver must produce, how it is verified, and what to reject. It is the one input a campaign has; every task descends from it.See alsoCampaign instance, Task instruction, Setter |
Interpreter | The program inside a sandbox that reads and runs code, such as Python. The runtime image decides which one a task's environment has, so every setter and solver trial in a campaign runs the same one.See alsoRuntime image, Sandbox, Environment |
Island | One scene the three-isometric-engine can mount on the centre of the ocean board: a terrain of a worldType plus the content flags and showcase that add structures. Every island is a preset row and has its own folder and doc.See alsoPreset, worldType, Showcase, Distant islands |
Iteration | One numbered batch of requested task samples. Samples are assigned by position and the chosen iteration count, so five tasks over three iterations form batches of two, two, and one. The worker completes each batch before starting the next, with trials inside a batch able to overlap within capacity. Mulberry supplies earlier completed history to later batches.See alsoSetter, Campaign instance, Task matrix |
Learning signal | The contrast a training method can learn from when it compares several rollouts of the same task. When every rollout passes, or every rollout fails, the rewards are identical and the comparison teaches nothing; mixed results carry the signal. Measured tasks in the frontier band are the ones with that contrast, which is why finding them is the point of a campaign.See alsoReward, Frontier band, Rollouts per task, Solve rate |
Long horizon campaign | A campaign left to generate and measure tasks across many iterations within its total attempt budget. Its method determines how generation uses earlier evidence. Naive self play proposes independent candidates, while Mulberry tracks category coverage and gives each new batch a history of completed work. Read the task outcomes and measured solve rates to see what the run produced.See alsoCampaign instance, Iteration, Frontier band |
Masonry | Construction assembled from individual stone blocks in horizontal courses. In the world, each block has a stable part address, a material identity, a physical shape, and a place in the construction sequence.See alsoStone block, Stone curving, Quarry |
Max output tokens | A setter or solver setting: the longest reply the model may write in one step, counted in tokens. It bounds how much an agent can say or write at once, not how much it may read.See alsoToken, Context token budget |
Memory | The working memory requested for each setter or solver sandbox, measured in megabytes. Programs use it for data needed while they run, and insufficient memory can make a process fail. It is separate from the model conversation and its token budget.See alsoSandbox, Trial configuration |
Method | The self-play loop a campaign runs, chosen in the campaign settings before launch. Naive self-play has the setter author tasks and the solver attempt them with no extra machinery between; Mulberry first agrees a taxonomy of categories and then authors tasks against it. Every campaign names one method, and an instance never changes it.See alsoCampaign settings, Mulberry, Setter, Solver |
Model | The language model used by an agent role. Campaign lets you choose and records the setter and solver models independently, while pack authoring, validation, judging, and refinement use their configured models. Compare solve rates with the recorded model and reasoning choices in view.See alsoSetter, Validator, Solver, Reasoning effort, Harness |
Model turn | One response generated by an agent model after reading its current conversation. It can choose commands whose results become input to a later turn. Campaign uses model calls to count agent steps.See alsoModel, Step limit, Token |
Mount manifest | The scene JSON plus the preset flags that decide which entities mount for a scene mode. An entity's manifest gate is the condition under which it is built at all.See alsoEntity, Preset |
Mulberry | The method whose first product is a taxonomy: a taxonomy trial proposes and repairs a category plan for the INSTRUCTION.md until it is accepted, and only then does the setter author tasks, each sampled against a category. It runs on the same campaign routes as naive self-play; the method is a create setting, not a second product.See alsoMethod, Taxonomy, Setter |
Network access | Permission for a trial's sandbox to reach the internet, such as fetching packages or reading pages. Campaign has separate settings for generation and repair setters and for solver attempts, and both default to on. Turning either setting off removes general internet access only for that group. Validators, agentic judges, and pack authoring always run online. Model calls go through the Campaign model proxy and remain available when a sandbox's general network access is off.See alsoSandbox, Campaign settings |
PipeStrike | The SodiumReactor mode that plays one accident. A crewed large ship leaves the boat patrol, sails down the channel between the two islands, and strikes the stone pipe that crosses it. The pipe shatters, both middle batteries snap in half, and the ship explodes. The salt pipes crack and their liquid runs over the land into the sea, the steam halls split and their stacks crumble, both reactors fail, the grass on both islands turns from green to brown, and the trees lose their leaves.See alsoSodiumReactor, Crystal agent, Preset |
Post | A markdown document published to VMAX.ai at /{username}/{slug}: Slate JSON on the server, rendered as a paper with an isometric scene behind it. Research papers are posts curated onto the /research index.See alsoworldType, Document trigger, Scene overlay shader |
Presentation | A slide deck saved on a Vmax account through the internet.dev slides API and edited on /slides. Its source is Markdown, its stored form is a slide document of positioned elements, and every save keeps a revision behind a version counter.See alsoSlide, Publication |
Preset | One row in the island catalog: a worldType, content flags, and an optional showcase. A preset row is the manifest input; the stage instances any preset, and a scene is browsed at /-/vmax-tools-template/islands/{key}.See alsoIsland, Showcase, Mount manifest |
Publication | The public link a presentation gets when its owner publishes a saved revision. The link shows that revision alone, without speaker notes or hidden slides, until the owner publishes again or unpublishes it.See alsoPresentation, Slide |
Quarry | An excavation in an island mountain where a timber derrick grips and extracts stone with articulated pincers, then deposits it in a standard pile. An agent near its hand winch can start a batch. Terraced working levels, a haul entrance, and a finite stone reserve connect the extraction face to the output pile.See alsoDerrick, Masonry, Stone block, Distant islands |
Reasoning effort | A model-specific setting for how much thinking to spend on a response. Supported levels and defaults depend on the selected model. Fixed thinking budgets must leave room inside the output token limit for the answer.See alsoModel, Setter, Solver |
Recursive self-improvement | The broader theme of a model that keeps making itself better, each round of improvement making the next one possible, with no end to the rounds. No system can do that today. The loop it describes is what runs now, one round at a time: find the hard tasks a model solves only some of the time, train on them, then find harder tasks at its new edge, and repeat. Every round needs new verifiable tasks, because a task the model has learned to pass every time no longer carries a learning signal. Campaign supplies the finding through self-play: each task is written with its verifier, solver rollouts measure it, the frontier band marks the tasks worth training on, and a Mulberry campaign leans each later batch harder on purpose. Training runs in your own pipeline on the exported tasks; Campaign does not train the model.See alsoSelf-play, Frontier band, Learning signal, Long horizon campaign, Mulberry |
Refine | The composer action that sends a draft INSTRUCTION.md to the model and returns it rewritten under fixed headings, with anything the draft did not say marked as not specified. It creates no campaign and changes nothing until Accept changes is pressed.See alsoComposer, INSTRUCTION.md, Model |
Reproducibility | The record needed to run a campaign again and expect the same workload: the trial settings the request was launched with, exactly as the worker read them, and the campaign agent's activity log in order.See alsoTrial configuration, Campaign agent activity |
Reward | The final pass or fail that a task's grading gives one rollout: the verifier's checks, plus an agentic judge's grade when the task calls for one. A reward is recorded only when the attempt was actually measured; an attempt that could not be measured gets no reward rather than a guessed one. Rewards can be trusted because the verifier was certified against known answers and the task passed validation before any rollout ran.See alsoVerifier, Rollout, Solve rate, Learning signal |
Rollout | One solver attempt at one task, run in its own sandbox under the solver settings. The task's verifier scores what the attempt produced. You choose how many rollouts to request for each task in the campaign settings, but only attempts with a comparable measurement enter its solve rate.See alsoSolver, Solve rate, Trial |
Rollout explorer | A Campaign application for following recorded agent work on a diagram. The first version follows one trial through numbered entries, and the second connects the campaign instruction to tasks, trial ancestry, and recorded outcomes. Selecting a node opens that node's full retained evidence in an inspection window. The window can compare its retained sources, prepare a separate proposed change, and request a summary where that action is available.See alsoCampaign application, Trial, Task instruction |
Rollout verdict | The worker's record of how one rollout was graded: its final outcome, the result of the task pack's mechanical gate, and the response of every judge that graded it. The API stores the verdicts with the task row and serves them exactly as the worker wrote them, so they show where each reward came from.See alsoReward, Rollout, Verifier |
Rollouts per task | The campaign setting that says how many independent solver attempts are scored on each task. It is the sample size behind every solve rate: four rollouts can only report a rate in quarters.See alsoRollout, Solve rate, Campaign settings |
Runtime image | The container image used as the starting software environment for generated tasks. It supplies system files, an interpreter, and installed tools, while task preparation adds the required workspace and dependencies. Pack authoring has its own environment.See alsoTask files, Campaign settings |
Sandbox | A temporary environment for running the programs of a trial with assigned processors, memory, storage, and a network policy. The sandbox workspace and the model conversation are separate resources. Campaign provides agent model calls through a separate scoped proxy.See alsoTrial, Sandbox backend, Network access |
Sandbox backend | The service that provisions every trial's sandbox; on this deployment it is Modal. A campaign setting that is fixed.See alsoSandbox, Campaign settings |
Scene overlay shader | One of the WebGL overlays drawn over a document from the shared shader context, such as the grain, burn, or water drop effects, rendered into slots the document layer schedules.See alsoPost |
Secret permission | The role allowlist attached to one campaign secret. Pack author, Setter, Validator, Solver, and Judge can be selected independently. A role that receives a secret can read and use its value.See alsoCampaign secret, Setter, Validator, Solver |
Self-play | How Campaign searches for tasks at a model's edge: a setter model writes tasks in your domain, each with its own environment and verifier, a validator rejects any task that cannot be graded fairly, and a solver model attempts every task several times so its solve rate is measured. The tasks that land in the frontier band are the ones worth training on, and each later round looks for harder ones. Vmax uses self-play interchangeably with recursive self-improvement when describing that loop, because self-play is the part of it that runs today. Recursive self-improvement is the broader theme, a model improving itself without end, and no system can do that today.See alsoSetter, Solver, Validator, Verifier, Frontier band, Recursive self-improvement |
Setter | The side of self-play that manufactures the evaluation. A setter trial reads the prompt, writes a task instruction, builds its environment, and ships a verifier, inside its own sandbox under the setter budgets.See alsoSolver, Validator, Trial, Task |
Share link | An owner-created link that lets another signed-in Vmax user read the owner-visible work of one campaign after it has completed, failed, or stopped. The reader uses their own account, and the link gives no permission to change or run the campaign. It includes the campaign brief, settings, named tasks, instructions, activity, trajectories, and certification through authenticated read endpoints, without disclosing the owner's identity or credentials. The full link is shown once when created, and each campaign has at most one active link. Rotating it replaces the old link, while revoking it or reaching its expiry closes access through every copy.See alsoCampaign instance, Stored evidence, Campaign status |
Showcase | A named structure or mode added on top of a terrain, such as the motte-and-bailey castle, the palisade path, or the crystal-agent chessboard. It is the third part of a preset after worldType and content flags.See alsoPreset, Island, Entity |
Slide | One page of a presentation, produced from a level one or level two heading in the Markdown and everything beneath it. Its layout is chosen from what it holds and its type is sized to fit the canvas, so nothing on it is measured by hand.See alsoPresentation, Publication |
Smoke trial | A small rehearsal of a task setting, shown in the console fixtures that provide fixed examples for internal checks. The fixture illustrates a setter authoring a task, a solver attempting it, and a verifier scoring the attempt. Its smoke label identifies the rehearsal so you can distinguish it from campaign measurements. Live campaign task rows are never marked as smoke, and a smoke trial is not a required first stage of a real run.See alsoTask setting, Certified task pack |
SodiumReactor | The island scene with two sodium-cooled fast reactor plants built from stone masonry, one on each of two thin islands that face each other across a channel. In each plant a recessed reactor heats molten salt that boils bright green through a translucent hot vat, a twin-stack hall turns that heat into steam and returns the salt clear and blue through a cool vat, and two long battery halls store the power. A stone pipe crosses the channel from the middle battery of one island to the middle battery of the other, trees grow in the ground the plants leave, and a timber dock stands on the outer shore of each island. A post selects it by containing the sentence AI had its Three Mile Island.See alsoPipeStrike, Masonry, Stone block, Showcase, Preset, Document trigger |
Solve rate | The share of measured solver rollouts that passed a task's grading rules. It is the number of successful rollouts divided by the number measured, rather than the number requested. Read it alongside those counts, because a few attempts give less evidence than many. Zero means measured attempts all failed, while a task with no comparable measurement has no solve rate.See alsoRollout, Verifier, Frontier band |
Solver | The side of self-play being evaluated. Each solver rollout is one sandboxed attempt at a task under the solver budgets, scored by the verifier when it finishes or when its timeout expires.See alsoSetter, Rollout, Solve rate |
Step limit | The maximum number of model turns allowed in an agent trial. The harness counts model calls, so a step is not a count of files changed or individual shell instructions. Time and token limits apply separately.See alsoTrial, Rollout |
Stone block | An individually identified piece of building stone with a length, height, and depth. It can remain straight or be curved for a circular structure while retaining its material identity and mass.See alsoMasonry, Stone curving, Quarry |
Stone curving | Shaping a stone block around a chosen radius. Its length follows the middle of the curved block, making adjoining pieces suitable for round towers, curved ramparts, and arches.See alsoStone block, Masonry |
Storage | The disk resource requested for each setter or solver sandbox, measured in megabytes. It holds workspace files, installed dependencies, and generated output during the trial. Evidence retention follows separate rules.See alsoSandbox, Task files |
Stored evidence | The saved record of finished trials that you can read to understand a campaign's results. It includes the permitted files, logs, and available agent trajectories, with parent and child trials connected into a tree. The API reads this record from storage and checks each saved object's length and checksum before serving it. Private configuration, task grading files, oracle solutions, and secrets are withheld. Missing material and any reported truncation limit what the record can tell you.See alsoTrial, Task, Campaign data |
Task | One evaluation problem with an instruction for the solver, an environment to work in, and a verifier to score the result. A setter authors it, and a validator checks whether it meets the campaign's acceptance rules. A task can be queued or running before it finishes as measured, rejected, or unmeasured. A measured task has passed validation and has at least one comparable solver measurement, while a rejected or unmeasured task has no solve rate.See alsoTask instruction, Task files, Verifier |
Task count | The number of task samples a campaign is asked to produce. Each can finish measured, rejected, or unmeasured, so the requested count is not a guarantee of usable tasks. More samples can broaden the cases covered by the campaign.See alsoTask, Iteration, Campaign settings |
Task files | The files a task puts on disk before the agent starts: the repository checkout, fixtures, and the dependency declarations the environment installs. Together with the runtime image they fix what the agent can reach, and they carry as much weight as the instruction.See alsoEnvironment, Task, Runtime image |
Task instruction | The problem statement the setter writes for one task, the exact text the agent is handed and the most important artefact on a task page. The campaign INSTRUCTION.md described the kind of task wanted; this is the one concrete problem it became.See alsoINSTRUCTION.md, Task, Setter |
Task matrix | The Atlas grid where each row is a campaign and each column a setter iteration. A cell counts the tasks that campaign generated in that iteration and is shaded by their solve rate.See alsoAtlas, Iteration, Solve rate |
Task setting | The code the campaign agent writes from the INSTRUCTION.md before any task exists: a small package that tells the setter how to build a task, what an environment for it contains, and how a verifier decides a reward. The setter runs inside it on every iteration.See alsoCertified task pack, Setter, Verifier, Smoke trial |
Taxonomy | The fixed category plan a Mulberry campaign uses to guide task generation. Each category has a key, a label, and a description of the work it covers. The plan is accepted before task generation and saved in the taxonomy trial's stored evidence. Mulberry then chooses the least-covered category for each new sample and lowers its target solve rate as the run progresses. The plan helps you see how the campaign interpreted your subject and where it is trying to add coverage.See alsoMulberry, Task, Stored evidence |
Temperature | A model sampling setting that adjusts how concentrated the choices of next token are. Lower values concentrate sampling on more likely choices, while higher values spread it across more alternatives. Lower temperature can reduce variation without guaranteeing identical results.See alsoToken, Solve rate |
TheImmortalGame | The CrystalChess mode that plays the recorded Immortal Game once, from 1. e4 e5 through 23. Be7 checkmate. Pieces walk between squares, captures throw opposing agents just beyond the island shoreline, and the surviving black army kneels when the game ends.See alsoCrystalChess, Crystal agent, Preset |
three-isometric-engine | The Three.js engine under every island and portrait on this site: an isometric world with physics, hand-written shaders, and the entities that build terrain, castles, forests, fleets, and agents from a manifest. It renders the scene above a paper, the islands catalog, and the characters who read these cards.See alsoIsland, Preset, Entity, Mount manifest |
Token | The unit a language model reads and writes in: a word or a piece of one, roughly four characters of English. Budgets in a campaign are counted in tokens because that is what the model is charged and limited by.See alsoModel, Context token budget, Max output tokens |
Tools page | A page on the shared Vmax tools shell: the top navigation, a sidebar, the fixed bottom toolbar, and one or more body columns composed from the Form kit. /-/vmax-tools-template is the reference every tools page clones.See alsoAppContainer, Form kit |
Trebuchet | A counterweight siege machine that throws stone blocks. A built trebuchet accepts a target zone within its firing range; an agent fires it, resets its arm, and carries a fresh block from a nearby stone pile before the next shot.See alsoStone block, Masonry |
Trial | One agent run inside a sandbox with a particular job and a set of limits. A setter trial authors or repairs a task, a validator trial checks it, and a solver trial attempts it. Other trials can prepare a taxonomy, author or certify a task pack, or judge an attempt when the task calls for an agentic judge. Trials can launch child trials, so one task's evidence can contain the work of several agents.See alsoSetter, Solver, Validator, Rollout, Campaign trials |
Trial configuration | The request a campaign was launched with, read back from the campaign record rather than the settings panel, so it is what the worker actually ran.See alsoCampaign settings, Reproducibility |
Validator | The trial that checks an authored task against the campaign's acceptance rules before it becomes an evaluation item. A task the validator rejects is never measured and is not counted.See alsoSetter, Verifier, Task |
Verifier | The program shipped with a task that decides its reward: it reads what a solver rollout produced and returns pass or fail. A task can call an agentic judge as part of its grading procedure.See alsoTask, Solver, Solve rate |
Verifier timeout | The time allowance for grading work attached to an agent trial. It is separate from the agent execution timeout. Campaign derives relevant task checking and judging policies from the role settings, while pack certification uses its own checking limits.See alsoVerifier, Trial configuration |
Vmax Campaigns Overview | A reference diagram of how the /campaigns application talks to the Campaign API, titled with the source version it was built from, such as Vmax Campaigns V3.7.6 Overview. Each node is a real piece of the source: the browser controller, the client, the routes, the Postgres tables, the external worker that runs each trial, and the components that render the answers. Lines show which call spins off the next. A simulated campaign instance plays the calls in order, from the INSTRUCTION.md to its stored evidence and share link, with made-up identifiers, while every request and response keeps the shape the real client decodes.See alsoCampaign, Campaign instance, Trial, Stored evidence, Share link |
Workload analysis | The view of how a campaign's requested tasks turned out. Its live bar counts finished task outcomes as measured, rejected, or unmeasured and updates as work completes. Running tasks remain visible in the task list without entering those outcome totals. Measured tasks passed validation and have comparable solver measurements, rejected tasks failed validation after the allowed repairs, and unmeasured tasks have no comparable result. Read a task's reason and stored evidence to understand its outcome, and its solve rate to understand how the solver performed.See alsoDifficulty mix, Iteration |
worldType | The terrain family a scene is built on: GRASS, ICE, DESERT, or their variants. It is derived from a post's content, and each type has its own terrain recipe, palette, and the structures it may carry.See alsoDocument trigger, Island, Preset |
Zoom | Camera framing that gives a scene more of the available canvas. Mode:Zoom in the island catalog identifies a closer presentation of the listed content.See alsoIsland, Preset |