Everything the harness does.
One question
at a time.

The guided tour

On your first visit, the guided tour points out the project, cards, connections, model access and notebook. You can close it and restart it from Help (?). The working canvas stays available while you learn.

Your canvas and the agent's map

The investigation opens on one working canvas. Add your question and evidence there. Once the agent is watching, its developing hypotheses, directions, tests and results also become visible on that canvas, with their agent origin and tentative status shown on each card. The agent keeps its own map in the background; you decide what to explore, ignore, edit or accept as your own judgment. A computed verdict does not set your verdict.

The Agent button opens the rail for discoveries, activity, decisions, automation and the notebook. The presence control above the canvas reads Following this investigation while the watch is active. The assistant at the bottom has separate Ask and Steer modes.

Seven words to know

• Investigation, also called a project: one question, one working canvas, and its evidence, agent activity and history.
• Node: a card on the canvas. A paper, a dataset, a question, an idea, a parameter, an experiment.
• Relationship: a typed edge between two nodes. The interface calls it a hypothesis, because it is one.
• Agent watch: the standing instruction to follow this investigation within the automation level and daily budget you set.
• Agent idea: a hypothesis, direction, test or candidate developed by the agent and marked tentative until evidence and your review settle it.
• Publication: a frozen, reviewed copy of My canvas, optionally with notebook entries, visible to signed-in members.
• Build package: one bounded product definition composed from an investigation, kept in the idea library and sent to a build tool. Chapter 14.

Chapter 17 has the full glossary.

Getting in

The harness lives at synatrail.com/app/. Registration is open: anyone who completes the form gets an account at once, with no approval step. Your investigations are private until you choose what to share, and the Terms of Use and Privacy Notice say exactly what is stored and why.

The header, left to right

• Current project shows the open project's title. Press it to open its details.
• Save (⌘ or Ctrl + S), New and Projects, which opens your investigation library.
• The save indicator: ✓ Saved, or ○ Saving, Unsaved, Conflict, Needs attention or Offline. Chapter 12 explains each state.
• On the right: Build package (chapter 14), Public investigations, Share, the theme toggle and your avatar.

Project details

Current project opens the panel Project details. It holds the two things a project needs: Project name and The question you are following. Press Save project to store them. Changing an existing question substantively shows Why the reframe? (optional); saving then records a Reframe entry in the notebook with the former and new wording, and the notebook entry opens project details. Canvas changes save on their own, and the strip at the top says so: Canvas changes save automatically. Under FROM INVESTIGATION TO BUILD sit Create a build package and Idea library, both explained in chapter 14. The More in this project disclosure links to All investigations, folders & Trash, Branches & comparisons, Sources & experiments, Community and Open an illustrative example.

The investigation library

Projects in the header opens a full-screen library titled Investigations. Its sidebar has All investigations, Unfiled, Idea library, a FOLDERS list with a New folder button, and Trash, each with a count. Above the list: New investigation, a search field, Search names or questions, a sort, Recently updated or Name A–Z, and a refresh button. Each card shows a drawn preview of the canvas, the title, a ✓ Current badge on the open project, its question, its folder, the node count and the last change, with Open, Rename, Move (to a folder or back to Unfiled) and Delete. Deleting asks Move to Trash? and warns that active research will stop and the public snapshot will be withdrawn; Restore in Trash brings the investigation back, still private and with research stopped. Folders offer Rename folder and Delete folder, which moves their investigations to Unfiled. Esc or Back to canvas closes the library.

Switching, naming, exporting

New projects are named Untitled project, then Untitled project 2, and so on, until you rename them. You cannot switch projects while edits are waiting to be saved. The message reads Save or resolve this investigation's pending edits before switching.

Export canvas in the status bar downloads the investigation canvas and your notes as a JSON file. Export workspace, in Settings under Account, exports everything you own. These files are records for you. There is no importer for them. Evidence enters the harness through uploads, links and data connections.

Four ways to add

• Press the + tool in the floating rail, whose full name is Add source or node (+), or press + on the keyboard.
• Double-click an empty spot on the canvas.
• Drag from a node's round output on the right and release on empty canvas. A new node appears already connected.
• Drop files anywhere, or paste. A pasted web address becomes a source node, pasted text becomes a context note, and pasted files upload.

The first three open the Start with a piece of something. menu:

On ordinary cards, Explore opens the assistant for that card. The separate arrow button, Choose next step, opens the next-step choices.

The menu ends with a link to Source catalog & saved work, the panel called Bring in what you already know., where the field Find saved evidence searches sources, notes and tests already in this project and marks those that are on canvas.

Every kind of node

The icon, outline and label identify a card’s role: folded documents for sources, a cylinder for datasets, a compact control for parameters, a flask for experiments and a chart for results. Some kinds are created by you, some by the agent, some by experiments.

Any kind of card can begin a flow. A card with no incoming connections, including one standing on its own, is shown in amber with a Start badge. It keeps its round output and any named inputs underneath, but has no left input at rest. Start is a position in the graph, separate from the Starting point kind. Once a connection enters any input, the badge and amber styling disappear and the normal left input remains visible; removing the last incoming connection makes it a start again.

Shape a node

Press Enter on a focused card, double-click it, or use its Open button to open the node panel, titled A piece of the investigation. Its fields:

• Title, up to 200 characters. On a question card this field is called Starting point, and the panel is titled Where the investigation begins.
• A body. For questions it is called Context: the place for what you already know, what puzzles or frustrates you, where you feel lost, and what you hope to find out. For everything else, What should the agent understand? Up to 18,000 characters.
• The optional Capture context disclosure on starting points, notes and source-like cards: What prompted this? keeps the moment or stimulus and Why this caught my attention keeps your personal reason. Both stay separate from source text, summaries and extracted facts.
• For parameters, a grid: Value, Unit (the placeholder suggests °C, EUR, kg or 1), Minimum, Maximum and Step.
• Source link, a web address.
• The Attribution & assessment disclosure: Tags, Who, where and when / source on capture cards (or Source / author elsewhere), Rights notes and Evidence status.

Save on canvas stores the card, with the toast Node saved. On a starting point, Capture what prompted this adds a context card and a raises question connection in one undoable edit. On an existing capture card, This prompted my question adds that relationship only when you choose it; ordinary paste, drop and add gestures leave the card unconnected. Explore this sends the node to the assistant with the question What is worth investigating about this node? Propose the sub-questions, the evidence to gather and the next steps as cards. The trash button, Delete node, removes it. Undo brings it back. A locked card first asks you to unlock it.

Parameter cards show their value and unit under the word Assumption, with a slider you can drag on the card itself. Source cards add Open original source, View uploaded original and the extraction tools of chapter 05.

Three ways to connect

• Drag. Every card has a round output on its right (Connect from …). A card with an incoming connection also shows a square input on its left (Connect to …). Drag from an output to another card’s input. While you connect, temporary left inputs appear on starts so you can join freestanding cards. Claims, answers, experiments and results keep their named diamond inputs underneath.
• Keyboard. Focus a port and press Enter or Space. The toast says Choose another node's handle to connect. Escape cancels. Then activate the other card's port.
• The Connect tool. Press C or the rail button Connect nodes (C), then choose an output and an input. With exactly two cards selected, the selection toolbar also offers Connect.

Cards with incoming connections usually have two ports. A Claim has four, adding Supports and Challenges; an Answer adds Evidence and Context; a Computed result adds Settings and Context. An Experiment with an incoming connection has six: the two side ports and four diamonds underneath, Evidence, Data, Settings and Context. Named inputs accept matching kinds: Settings takes parameters; Data takes datasets, data connections, observations or computed results. Connecting a named input sets the relationship, such as supports, contradicts or informs. A filled port shows that it has a connection. A start hides only the left input, so it shows one port on a basic card, three on a claim, answer or result, and five on an experiment. All named inputs stay available; the left input reappears temporarily while you connect.

Connecting two cards that are already connected opens their existing edge instead, with the note These nodes are already connected. Refine their relationship here.

The edge panel

Its overline reads A RELATIONSHIP IS A HYPOTHESIS and its title What connects these? Eight quick chips sit at the top: related, informs, supports, contradicts, analogy, depends on, feeds input and joins on. The Relationship dropdown holds all twenty-three types.

Use Text on the connection to add up to 240 characters below a relationship’s type on the canvas. Save it with Save relationship. Clear the field to remove the annotation; the relationship type and its rationale stay separate.

Below the type, the field Why do you think they connect? takes up to 2,000 characters of rationale. Assessment is one of unassessed, tentative, supported or refuted. Reverse direction swaps the arrow. Save relationship stores it.

Two buttons hand the edge to the assistant, once a model is connected in chapter 07: Interpret asks what the connection could mean and Challenge asks What evidence would contradict it?. A third, Make testable, is yours alone: it opens a test plan on its own Experiment card, carrying this relationship and the revision it has right now. The footer states who made the edge: Human interpretation or Agent proposal. Edges the agent proposed can carry a Supporting evidence disclosure.

One connection is guarded. Connecting a parameter to an experiment whose input expects a different unit fails with This experiment expects …; the parameter uses …. Set compatible units first.

Drop a file

Drag files anywhere over the canvas. The overlay reads Drop it into the investigation. with the limit UP TO 10 FILES · 25 MB EACH. The same limits apply to the file picker and to pasted files. Each file becomes a card whose body reads Original received. Preparing extraction…, then extracting…, then understanding… if a model is connected. If something fails the card says Needs attention · open to retry.

Anything else is preserved but not read: This original is preserved, but its format requires a text or supported document export.

Read what was extracted

Open a source card. The disclosure Source text & reading coverage starts with a coverage line such as 12 of 30 pages extracted · 3 read with OCR · Review coverage, or 48,120 characters extracted · Complete extraction. Passages are labelled Page 4, Paragraph 12 or Passage 3 so a citation can point back to them.

OCR results come with honest warnings: reading order is approximate, numbers and tables need review before computation, and a confidence score is a recognition score rather than a probability of being right. Pages with low confidence are listed by number.

To fix an extraction, open Correct the extracted text, edit Corrected source text and press Save a correction. The toast reads Correction saved as a new extraction revision. Nothing is overwritten. Extraction versions (N) lists every revision, and human corrections are marked as such. If the source has been re-extracted since you started editing, the panel asks you to compare before saving.

Two more buttons: Retry extraction runs the reader again, and Choose the original again replaces the file. Understand this source asks your connected model to summarise it into Extracted facts; the toast reads Source understanding queued.

Bring a link

Paste a web address on the canvas, or use Bring a link. The card is titled A source worth exploring and reads Reading this source… while the server fetches it. Public reads are bounded to 2 MB and 25 seconds. A server without web reading says so: Web reading is unavailable on this server. Add an excerpt or upload an export.

Connect an API

A data connection is a bounded GET request that uses the source's own credentials and turns each response into a dataset version. Its panel is called Connect it where you use it. and has four sections.

Save connection stores the settings. Connect & test performs one real read and reports Read completed in 1 call(s). A versioned sample is ready., or explains why it stopped early: call budget, row limit or size limit. The Retrieved sample shows the first twelve rows. Disconnect removes the credentials; earlier samples remain.

The zoom control offers Zoom out, Fit, Zoom in and Toggle minimap. Zoom runs from 15 to 200 percent; below 40 percent, cards switch to a low-detail overview so a large canvas stays legible. F fits everything into view.

The selection toolbar

Select one or more cards and a toolbar named Actions for the selection appears. It counts the selection, then offers Next step, Find connections, Ask, more actions and Clear selection. Under more actions are Group, Trace, Compare alternatives, Connect and other actions that fit the selection. Drag its grip to move it; arrow keys nudge it and Home resets it.

The … button also leads to All actions for the selection:

• Next step: open the guided next action for the selection. Chapter 09 explains the agent rail.
• Branch and compare: keep this interpretation as a baseline. Chapter 11.
• Keep in notebook: a notebook entry pointing back to the selection.
• Lock / unlock evidence: pinned cards cannot be moved or removed by accident, and their status reads Pinned evidence.
• Exclude from exploration: the agent will not use these cards. Their status reads Excluded. An optional Dead-end reason records why in the notebook. When the selection holds excluded cards, Include in exploration brings them back; the saved research direction follows the change, affected research pauses until you resume it, and stale evidence stays stale.
• Group selected evidence and Share selected evidence.

Read a busy canvas

Relationship labels stay quiet until you hover or select. Selecting a card fades unrelated cards and wires; click empty paper to clear the focus. Each card initially shows only its strongest few wires, with +N more for the rest. Open Layers to choose stages, focus a direction or Show only a group. Under Relationships & details, Show relationship labels and Show all wires reveal the full graph. These settings change the view, not the saved investigation.

Open a card and press Trace the reasoning to follow its numbered chain and the explanation for each step. Group connectors count bundled relationships; opening one shows the underlying links. A research direction places its hypothesis, test and result in readable rows, with supporting evidence and settings wired into the test.

Research groups

A group is a frame with a name and a purpose. The assistant receives the group's meaning together with the evidence inside it, so asking about a group is different from asking about its members one by one. Press G and a guide appears: Group connected evidence, Drag around nodes, or tap nodes to select them., and a live count such as 3 selected. Press Create group.

The group's own panel, Give this group a purpose., holds Group name and What connects this group?, with the placeholder A shared question, hypothesis, theme, or reason these sources belong together…. Save group stores it. Ask this group sends the group, its nested groups and its sources to the assistant. The In this group list lets you remove members, Add selected (N) adds the current selection, and Choose nodes on canvas returns to the canvas to pick more. Ungroup removes only the frame: Keeps every node and its connections. Undo restores the group.

Undo and redo

The floating undo control and ⌘ / Ctrl + Z step back through canvas edits: moves, additions, deletions, connections, group changes. Undo is aware of revisions, so it never discards an update that arrived later from a source reader or the computation worker. What it cannot undo, the help panel says plainly: Completed research actions and provider charges remain in history.

The status bar at the bottom reads 5 nodes · 4 connections, then a hint that follows the active tool, then Export canvas, then the visibility of this project: Private investigation, Private working copy or Public snapshot.

Model access

Press your avatar, Profile and settings, and open Model access. Choose an intelligence source and give processing consent. BYOM offers your personal provider connections; managed mode lists only qualified routes and asks for no personal API key or USD provider allowance. If managed routes are unavailable, the panel says so and does not send work through BYOM instead. BYOM offers these providers:

Connected providers show Connected. A personal Codex route also appears, marked unavailable until its isolated integration is verified.

With an API key

Choose a provider and fill in API key, Model and Default research allowance (USD). The allowance is a ceiling for autonomous work, between 0.10 and 100 dollars, ten by default. Then tick the consent line, which is required: Allow this model to understand sources I add to this workspace, within my research allowance. Press Connect & activate. The form validates the key without starting research, and the toast reads Model access saved. Your canvas remains available. Keys are encrypted on the server and never shown again. Remove key deletes one.

The capability check

Autonomous research needs two things a model may or may not offer: structured tool calling and a cited public search. The Research capabilities disclosure runs both, for real and billed, within Maximum allowance for the capability check (USD), five dollars by default. Press Check tools & search. The toast reads Research capability check completed. Until you run it, the panel says Run the capability check before activating autonomous research.

With a Claude subscription

Choose Claude Code · Pro / Max. The panel explains that your Claude plan will handle ideas, source understanding and research, then states the consent: By choosing Connect Claude, you select Claude for this workspace and allow it to process sources you add, generate ideas and research your approved public queries. Press Connect Claude, then Sign in on Claude ↗ and approve access with your own account. If Claude gives you a code instead, open Claude gave me a one-time code, paste it into Code from Claude and press Finish connecting. The status strip moves from Waiting for Claude sign-in to Claude connected. Your sign-in stays in a private hosted session for your next visit. Disconnect Claude ends it.

On a server where hosted sign-in is not enabled, the panel says Browser sign-in has not been enabled on this server yet.

Routes for answers and background work

For an explicit question, Answer with chooses the connection that answers; unconnected models say · Connect first. For the standing agent watch, a paired Claude Code connection can be verified yet offline. If it goes offline and you also have a verified API connection, background passes can use the lowest-cost verified API model until Claude Code returns; the rail names that fallback and charges it against the saved daily budget.

Tell the model about you

The Research context tab has two long fields, What are you trying to understand? and How should a model work with you?, up to 12,000 characters each. They accompany every question and mandate. Save research context stores them.

The dock

Its first line tells you the scope: Thinking with your canvas, Thinking with "A paper on urban heat", Thinking with 2 selected nodes or Thinking with the selected relationship. Next to it, Ask sends a scoped question, Steer tells the agent what to focus on within its saved limits, and Answer with chooses the model for a direct answer.

The text field reads Talk to your canvas. Enter sends, Shift + Enter starts a new line. The book button on the first line opens the conversation history.

Without a model, sending shows Connect a model and enable source processing in Model access to start research. and opens the settings for you. With one, the toast reads Question saved. The assistant is working with this scope.

The conversation panel

Answers arrive in Think with the evidence. Messages are marked YOU and RESEARCH ASSISTANT. While the model works, a progress block titled Working with your evidence carries a LIVE badge, the latest status, and a disclosure Activity and model summaries that lists each tool call and summary. Text streams in under ANSWER IN PROGRESS.

A finished answer ends with three things. A disclosure counting its evidence references, each pointing to a card and revision. A disclosure called Limits and open questions. And a provenance line such as Selected model: OpenAI · gpt-5-mini. A Claude subscription run that cannot confirm which model actually replied is marked Actual model unverified.

Continue with Continue with this evidence and Ask a follow-up. If a run stops early the panel says so, The answer is paused. or This run finished without an answer., and offers Try question again, which retries with the original model and evidence, and View this run.

Keep an answer

An answer also proposes the evidence and sub-questions it calls for, as dashed cards on the canvas you asked from, marked Proposed · approve or reject. Their panel explains: The research assistant proposed this card alongside an answer. Approve it to make it yours, reject it to remove it, or reconnect it first. A proposed source, paper, study or dataset must carry a public address; the server checks the link before saving, and a card whose link could not be verified says so. For those cards the footer button on My canvas reads Approve & read: approving also fetches the address as an ordinary source, with its extraction and coverage, so a proposal becomes evidence only once it has actually been read.

Every answer offers Add answer to canvas and Keep in notebook. The first creates an Answer card titled Model answer · review the evidence on the canvas you asked from, beside the cards you selected and connected to the evidence it cites. Open it and you see where it is saved, its Limitations and Evidence used, and the button Ask a follow-up; an answer asked on Agent canvas also offers Keep on My canvas. If the cited evidence changes later, the card warns: Evidence has changed since this answer. Review its original sources before relying on it.

The edge panel's Interpret and Challenge buttons, the card footer's Explore, the group panel's Ask this group and the selection toolbar's Ask and Ideas all lead here. Make testable does not: it opens a test plan instead, in chapter 11.

Start the watch

Open Agent in the left command bar. The rail shows Discoveries, Activity, Decisions, Agent and Notebook. Under Agent, choose an automation level and daily budget, save them, then press Start the agent. The presence control reports Following this investigation and the analyses remaining today. Pause, Analyse now and Stop are available in the same rail; Stop asks you to confirm and leaves the map intact.

Manual acts when asked. Suggest keeps thinking and proposes next steps. Research can investigate and run private tests. Full auto can run supported, configured actions within the daily budget. The rail shows which abilities are available at the saved level and why an ability is unavailable.

Approve public research

External research is off until you turn it on for this investigation. The rail shows the short, general terms that may be disclosed to public search. Review those terms and the search budget before saving the policy. Queries are composed from that approved policy; private notes and raw project text are not sent to public search. If a test needs web evidence and no policy is approved, its card says Needs you: Approve a search policy. You can approve one and let the test continue.

Read, steer and decide

Use Discoveries to read a candidate, then Explore, Why or Ignore. Why shows the source trail and any relevant notes the agent consulted. Activity shows recent passes and costs; Decisions records what you chose. In the assistant dock, choose Steer and write what the agent should focus on next. Steering uses the connected model and its saved limits.

The agent can remember hypotheses, dead ends, discoveries and retrieved sources across your own projects. Its notebook entries remain live references; they provide background for a new pass rather than silently deciding what is true on this canvas.

Open Notebook in the Agent rail, or choose it under More in the command bar. The drawer's overline reads COLLECT THE USEFUL BITS. Three filters: All notes, Agent findings and Yours.

Every entry is labelled Your note or Agent finding, with its kind, and carries a Return to canvas button for each card or edge it references. Agent findings offer Keep in my notes, which promotes them into your own notes so they survive a change of mandate.

Write your own with the field Something worth remembering… and Add a note. Other flows write here too: Keep in notebook from the selection toolbar and from answers, and Keep comparison in notebook from chapter 11.

Share notebook publishes the notebook alone. The canvas stays private. Chapter 13 explains the review step.

Follow an agent test from proposal to result

On a research direction, the agent proposes a test and compiles a plan with a metric, threshold, disproof condition, sample rule, cost cap and method. The test card shows its state and Step 1 of 8 through Step 8 of 8: compiling, choosing data, reading sources, wiring, running, presenting, deriving and finished. A refusal or failed compile names the reason. At Research level a prepared test can run within the saved limits; at Suggest level you choose Run; at Manual level you can make it your own test.

The agent binds an existing dataset, extracts a cited public table, or asks for a value only you know. It makes separate evidence, context, setting and data cards and wires them into the test's four named bottom ports. Public evidence quotes must match a stored page word for word; the card labels the agent's reading as supporting or challenging, never as proof. The test panel has Plan, Data, Wired, Context, Receipts, Result, Agent reading, Derived and Your verdict sections.

A verified result appears beside the test with a measured number, unit, source and limits. It connects back to the hypothesis with an evaluates relationship. The hypothesis shows Verdict: supported by computation, Verdict: weakened by computation or Verdict: inconclusive. The agent then derives a candidate, alternative claim or sharper test from the result. You still record your own verdict separately.

Write your own test plan

Every Experiment card can hold a test plan: what the test asserts, what observation would show it is wrong, and how it is carried out. Make testable on a relationship opens one, already pointing at that relationship and the revision it has right now, with Open the relationship to go back to it. Choose A test you carry out yourself and the card is a plan and nothing else — nothing is queued and no calculation runs. Choose A calculation on your evidence and the calculation appears below; it still waits for Run experiment, and it will not run until the plan is saved, so the result names the exact plan revision it answers.

Changing the claim, the falsifying observation, the method or the target saves a new revision, and the heading counts it: TEST PLAN · REVISION 2. Earlier results keep the revision they answered. Writing down different conditions is not a new revision. If you later delete or rewrite the relationship the plan points at, the plan survives and says so.

Record what you observed

Once a manual plan is saved, Record what you observed appears under it. Write the date, what actually happened, the conditions as you carried them out, who did it and what the result cannot settle. Two questions are kept apart. Did the test run? is about the work: it ran and gave a clear observation, it ran but settled nothing, or it could not be carried out. Your reading of it is yours: not read yet, tentative, supports the claim, or refutes it — and a reading needs a reason.

The result is saved on its own Observation card, drawn to what it tested with an evaluates line, and written into your notebook, all in one save. Nothing is calculated and no receipt is invented. A result that refutes your claim still counts as a test you carried out, and it never changes the relationship's assessment on its own — you record that verdict on the relationship itself, with your reason. Saving twice never makes two cards. Correct this result saves the next revision of the same record and leaves the earlier entries in your notebook.

What a finished calculation tells you

When the worker finishes, the result appears as its own card in the investigation, labelled Computed result or Simulation result. That label comes from what actually ran, not from what anyone wrote, so a plan can never look like a finished output and a simulation always says it is conditional on its assumptions. Open the card to inspect its receipt; nothing you type can create or change one.

Open the card and you can read the exact versions it was computed from: the output version, the software and method versions, each frozen dataset with its hash and the moment it was saved, the bound parameters and any random seed. From there, Open the test plan takes you to the exact plan revision it answers, and the plan's Open the relationship takes you to the claim itself. The card says whether the run started from that plan or was linked to it after the run, so the order of events is never blurred.

Connect to what it tested draws an evaluates line from the result to the card it tested. It is drawn unassessed on purpose: accepting a result never makes a relationship supported. A run that failed or was cancelled publishes no card at all, and an unrelated result answers nothing — the plan tells you how many results answer its falsifying test as it states it now.

Open an experiment

Use Create an experiment in the add menu, Adjust on a parameter, Configure test on an experiment, Open result on a saved result, or Branch and compare from a selection. The panel is called Make the connection testable. Give it an Experiment title, choose a Calculation, and pick an Input dataset from the versioned datasets on this canvas: uploaded CSV and JSON files, and data connection samples. Without any, the panel says Drop a CSV on the canvas to make a versioned dataset available here.

Units are checked first

Every unit field offers the supported symbols: 1, %, m, km, cm, mm, s, ms, min, h, day, kg, g, mg, K, °C, A, mol, Hz, N, Pa, J, W, kW, kWh, EUR, USD, GBP and count. Symbols combine with * and / and integer powers, so EUR/h and kWh/day work. Formula tools state it under the form: Formula units are checked before execution. Incompatible units stop a run before it starts, and a parameter with the wrong unit cannot be bound.

Bind a parameter

Every scalar field has a second dropdown, Bind … to a parameter, with Typed value and every parameter card on the canvas; incompatible ones are marked. Connecting a parameter’s round output to an experiment’s Settings diamond, or dropping the parameter onto the experiment, opens Put this input to work. Choose the Calculation input and press Bind input, or Bind in an alternative to leave a finished result intact and start a comparison branch instead. The Bound parameter values disclosure lists each binding with the note canvas value pinned at run.

Run, read, trace

Press Run experiment. The run passes through queued, running and completed, and a Computed result card appears in the investigation. In Saved experiments each result shows metric tiles, a chart, a table of up to twelve rows, and for simulations the disclosure Which assumptions move the result?

The Inputs & calculation receipt disclosure is the provenance: each input dataset with its version timestamp and content hash, each step with its method version and duration, the software version, the seed, and the Output version. Locate source jumps to the dataset card. Rerunning with unchanged inputs reuses the output and says so. Export result downloads everything.

Compare a change

Compare a change opens What changes? Choose a Saved baseline, name the alternative (the placeholder suggests What if the demand increases?) and press Branch & change assumptions. The note explains what happens: The alternative starts with the baseline's pinned datasets, parameters and calculation steps. Your canvases and baseline result stay intact. Change what you want in the branch and press Run alternative.

Then, under Compare saved outcomes, pick the Alternative result and press Compare changes. The verdict is one line, The output changed or The output is unchanged, followed by tiles for Inputs, Parameters, Recomputed steps and Unaffected steps, and a table of Measure, Baseline, Alternative, Change. Keep comparison in notebook records it. Preserved branches lists every branch with an Export branch button.

What it is made of

Only saved events: a source that arrived or was read, a correction, a finding or relationship the agent proposed, an answer, a calculation that finished and which plan revision it answers, your decisions and notes, and how a research run ended and why. Nothing is asked of a model to write it. Every line names the thing it came from and takes you there. If that card has since been deleted, the line still reads and takes you to Activity instead, which keeps everything.

Got it, and what it means

Got it marks what you just read as seen. That memory lives in this browser only — sign in somewhere else and the summary starts fresh. If more happened than fits on one page, the banner says so (Oldest 40 of 55) and Got it marks only the page you were shown; the rest waits for you next. Nothing marks itself seen on its own: leave the tab open in the background all week and everything that arrived is still there when you look.

Still open

Underneath, Still open lists what nobody has settled: what a stopped run said it could not answer, the evidence a finding says it is missing, and any test plan with no result yet. Answer drops the question into the ask box as a draft. Use as next goal puts it into the research mandate. Neither sends anything or spends anything — you still choose the model, the scope and the budget, and press Start investigation yourself.

Closing the tab with unsaved work triggers the browser's own leave-page warning. If the browser refuses to keep an offline copy, a bar warns This browser could not keep an offline copy. Export your changes before closing. with an Export pending changes button.

When two versions disagree

A conflict shows a banner: Edits conflict with another saved revision. and 2 fields need a decision. Both versions remain available. Press Review changes to open Keep your work. It lists each disputed field, marks entities that were removed in one version, and shows Your pending version and Current saved version side by side. Choose Keep my conflicting edits or Use the saved version. Everything that did not conflict is merged either way.

Undo works across all of this. It steps back through your own edits while preserving updates that arrived from a source reader, the computation worker or the agent in the meantime. The help panel is explicit about the boundary: Completed research actions and provider charges remain in history.

Publish

Press Share in the header, or Share notebook inside the notebook. The panel is called Share your investigation. and starts with your status: Your investigation is private. or A public copy is available.

• What would you like to share?: Investigation · My canvas or Notebook only.
• A short introduction, required: What are you investigating, and why does it matter?
• What would help?: A source, a counterexample, a way to test an idea…
• Topics · separated by commas.
• Include my notebook entries.

Press Review public copy. The review is the exact copy, marked REVIEW BEFORE PUBLISHING, with every card, relationship and included note, and one disclosure: This exact copy will be visible to signed-in community members. Source text shown below is included. Original files, connection settings, credentials, and your Agent canvas stay private. Press Publish reviewed copy. Later edits stay private until you return here and press Publish this update.

Make private asks for confirmation, Make this investigation private?, and warns that working copies others have already made remain theirs. Withdrawing blocks new reads, votes, comments and submissions.

The public board

Public investigations in the header opens the board, titled Public investigations. Filters: All investigations, Yours, Contributed to. A search field, Find an investigation…, and a sort, Most upvoted or Newest. Each row shows the vote control, the author, the title and introduction, topics, counts of comments, contributions, nodes and notes, the author's LOOKING FOR line, and Open investigation or Open notebook. Every member has one vote per publication, up or down, changeable any time.

Inside a publication

Four tabs: Canvas, a read-only copy you can pan and zoom, with every card and relationship listed below it; Notebook; Discussion, where Your comment and Add to the discussion start a thread; and Contributions. Copy link copies an address that returns to this publication after sign-in. On your own publication the buttons read Open your original and Manage publication.

Contribute

On someone else's publication, press Make a working copy. You get a private investigation of your own with every canvas tool, and a banner: Your working copy of Mara's investigation, with Discussion and Submit changes. The header's Share button now reads Contribute. Add a source, a relationship, a note or a counterexample, then press Submit changes.

The panel Start with a useful addition. shows only the actual differences, counted as new nodes, edited nodes, removed nodes, new relationships, edited relationships, removed relationships, new notes and changed question. Explain them in What does this addition help explain? and press Submit reviewed changes.

Review what others found

As the author, open your publication's Contributions tab. Each submission has Review submitted changes, a field-by-field Before and Proposed view, and two buttons: Accept contribution and Decline. Acceptance merges into your private original and preserves your unrelated edits and private source bindings; conflicting edits are flagged for review. The toast reminds you of the last step: Contribution accepted into your original. Publish an update to share it.

Threads, people, connections

An earlier community panel, Think together., remains available with four tabs. Shared threads holds evidence snapshots published from a selection, with Copy to canvas, Report and, for your own, Withdraw. People lists members who opted into a profile, with Request a connection where they accept it. Connections lists requests to Accept, Decline or Close connection. Share selection, reached from the selection toolbar, publishes a reviewed snapshot of selected cards, optionally with source excerpts, after Review selected material. Administrators see a Reports tab to withdraw reported contributions.

Open the composer

Press Build package in the header, open Current project and press Create a build package under FROM INVESTIGATION TO BUILD, or select the exact cards you want and choose Prepare a draft package from the selection menu. The selection route keeps those cards pre-marked. The panel is titled From investigation to build. and explains the sequence: mark, review, research the pitch, then build.

Mark what belongs

Three lists with All and None buttons: CARDS · 8/8, NOTEBOOK · 3/3 and EARLIER ANSWERS · 2/2. Cards you had selected on My canvas are pre-marked; excluded cards, pending proposals and group frames are not offered. Earlier answers show the question they answered. At most 30 cards may be marked; the form says Mark at most 30 cards. Group or unmark some before composing. Then choose What kind of product: Let the evidence decide, Software, Hardware or Hardware and software; write Steering (optional), whose placeholder suggests Who it is for, constraints, what must be left out, preferred stack or parts…; and pick the Answering model. Three buttons close the form: Plan the pitch research, Check the scope with the agent and Skip the pitch, build the package. Without a connected model, any of them takes you to Model access.

Check the scope with the agent

One bounded turn over the whole investigation. The review names the unmarked cards, notebook entries and answers the package would need, as RECOMMENDED · 2 UNMARKED ITEMS · 1 ESSENTIAL, each with a reason and a checkbox, and lists GAPS THE INVESTIGATION DOES NOT COVER with the warning Gaps become open questions in the package unless you add evidence first. Then: Add 2 and plan the research, Plan with my marks only, Skip the pitch, build the package or Back to marking.

Research the pitch

Planning is one more bounded turn that proposes up to eight generalized public search queries, each with what it verifies. The review says These generalized queries go to public search. Edit or drop any that could reveal something private, then approve. Edit any line, untick any, or Add a query. Set the limits: Searches (1 to 8), Sources read (1 to 12), Minutes (3 to 30) and, for API providers, Max USD. Press Start the research · 5 queries. The panel then reads Researching the pitch with public sources… with live counts of searches, sources read, turns and USD, and a log of every SEARCH and READ. You can close it; the pitch appears in the idea library when it is ready.

Under the pitch: Build the full package, Revise the pitch, which returns to marking, Download, which saves the pitch as Markdown, and Remove. A pitch cannot be sent to a build tool; only a package can.

The package

Composing the package is one large bounded turn. The panel reads Composing the build package from your marked evidence… and warns that it can take a few minutes. When it is ready the panel is titled A product, defined. Its head states HYBRID · 3 modules · 7 stories · 2 parts · 1 drawings, the product name and one line. Four tabs follow.

• Overview: Problem space (who, pain, context), Solution idea with its differentiators and NOT IN SCOPE, Architecture, Open questions, the separately frozen Recorded investigation basis with exact selected material, named concepts, next test, assumptions, unresolved risks and selected human decision notes, and Evidence, where each cited card is a button that jumps to it on the canvas. Provider prose cannot replace the recorded basis, and the viewer does not infer that a nearby decision chose a concept.
• Build plan: Modules with dependencies, Waves marked parallel or sequential, PRDs as one disclosure per module with numbered stories and their acceptance criteria, Bill of materials for hardware, Agent concurrency, Controls and tests and Dependencies.
• Drawings: the physical parts of a hardware or hybrid package. Chapter 15.
• Full spec: the whole definition as Markdown, the same text that ships as PACKAGE.md.

Every list in a package has a hard bound, so it is a definition, never an open-ended document: at most twelve modules, six waves, eight stories per module, thirty parts, twelve physical parts, eight risks, eight open questions.

Send it to a build tool

Under SEND TO A BUILD TOOL: VS Code downloads the archive; unzip it, open the folder and start your coding agent, and AGENTS.md tells it where to begin. Lovable, Base44 and Replit Agent copy the build prompt to your clipboard and open their new-project page; a disclosure Build prompt for Lovable · copied keeps the prompt in view. The prompt embeds the spec up to 24,000 characters and points to the archive for the rest. Download archive and Remove sit at the bottom.

The idea library

Press Projects and choose Idea library in the sidebar. Every pitch and package you composed is a card: PITCH with its verdict and confidence, or the package kind with 3 modules · 7 stories · 2 parts. Each offers Open, which switches to the package's investigation, Download, a Send to… selector for packages, and Remove. Before anything is composed the page reads No pitches or packages yet.

The parts in the package

Section 10 of the spec, Geometry, lists up to twelve parts, each with an id, a kind (mechanical or electronic), an envelope in millimetres, a material, a mass, the module and bill-of-materials line it belongs to, and a preferred tool: build123d for machined, moulded and printed parts, openscad for simple primitives, kicad for a schematic. Software packages have no parts and no Drawings tab.

The Drawings tab

The tab opens with a line naming what the runner can do, RUNNER · BUILD123D · OPENSCAD, and one entry per part: its name, MECHANICAL · BUILD123D, its stated 100 × 60 × 25 mm · PLA · 40 g · module firmware and description, and a status chip. Press Generate drawings. The chip moves through Writing the source…, Queued for the runner and Building on the runner to Drawings ready when every check passes, or Checks failed or Failed with the reason. Generation needs a connected model and the package's own investigation open; otherwise the tab says Open this package's investigation to generate geometry. A deployment without a runner says No geometry runner is connected to this deployment. and keeps the button disabled.

What comes back

Four views, FRONT, TOP, SIDE and ISO, each with a title block reading 1:1 · mm · third angle and overall dimension lines; hidden edges are dashed. A table of checks with PASS or FAIL in the last column. A line such as build123d · attempt 1 · 0.11.1 · 5.1 s · about 74.4 g. The files: source.py, source.scad or source.kicad_sch, part.step, part.stl, part.glb, top.dxf and the four SVGs, each a download with its size. Two disclosures, Engineer’s notes and Assumptions, hold what the model said about its own work.

Revisions

When a check fails, or the source does not build at all, the harness starts exactly one revision turn by itself. The entry reads A revision turn is fixing the failed checks…, and the next turn sees the previous source, the failed checks, the error and the runner's log. The second version appears above the first, and the entry counts 2 versions. After that the button reads Revise with the failures for a part that still fails, up to a third attempt, or Regenerate for one that passed.

In the archive

Once a part has completed geometry, the package archive gains cad/README.md, with each part's tool, verdict, measured envelope and mass and every check, and cad/<part>/ with the source and exports of the latest completed version, preferring one that passed. AGENTS.md gains a CAD section telling the build agent to work from that source instead of redrawing, and the build prompt for hosted tools mentions the folder.

Profile

Display name, Bio, Skills, Interests and Your current question describe you to other members. Two checkboxes decide visibility: Show my profile in People and Accept connection requests. Both are off until you turn them on. Save profile stores them. Your avatar shows the initials of your display name.

Account

Export workspace downloads everything you own as JSON. Below it, Current password, New password and Change password & sign out. Sign out is refused while edits are pending: Resolve or export your pending edits before signing out.

Credits · available now

The Credits tab shows your available balance, the packs you can buy, the cost of each commissioned job before you start it, and a history of every reservation, charge and refund. With BYOM, your model provider bills through your own key or plan; Synatrail-managed model usage is charged separately in platform credits. The pricing page explains what is never metered and what credits buy.

Appearance

The header's Change color theme button switches between the sand-paper light theme and the deep blue dark theme. The choice is remembered in this browser, and the sign-in page and the project pages follow it.

Administration

Administrators see one more tab. Invite an email address (optional) and Create invitation produce an invitation code, and the member list shows each email, name, organisation, role and status with Suspend and Activate. Registration mode is set by the operator. On synatrail.com it is open: the sign-up form admits and signs in anyone who completes it, and invitation codes are only needed on closed deployments. The details you gave at sign-up, your consent and its version are kept with your account, as the Terms of Use and Privacy Notice describe.

Addresses you can bookmark

Not in this release

• Simultaneous editing of one canvas by several people. Contributions go through working copies and review instead.
• Multi-factor or federated sign-in. Accounts are email and password, managed by the administrator.
• Arbitrary executable connectors. Data connections are bounded GET reads and experiments are the nine tools; the only generated code that runs is geometry source, in the isolated runner of chapter 15.
• Photoreal renders of hardware parts. Drawings, not renders, for now.
• Schematics on synatrail.com. The runner there has build123d and OpenSCAD; electronic parts need a runner built with KiCad.
• Importing the JSON exports back into a project.
• Personal Codex as a research route.

One last thing the harness says about itself, and this documentation repeats: implemented features are not claims of verified scientific findings. The example in these pages is a way to learn the tools, not a result.