7. TARS SI
The chat: bubble, tab, three layers, modes, skills, history, and its limits.
What TARS SI is
TARS SI stands for Traceable Applicability Reasoning System. It is the chat of PYRION: a bubble at the bottom right of every page, and a full page under the TARS SI tab.
You can ask it about a condition, about NASA documents on combustion and spacecraft fire safety, or about this site. It works like a colleague who knows the subject and shows its sources.
What makes it different from a general chat is that its answer is split into layers, and code checks each layer before you read it.
The three layers of an answer
🔒 Engine. Facts the engine or a calculator computed: verdicts, counts, partial pressures. Any number here must come from a tool result.
📄 NASA said it. A paraphrase with an exact quote, the document title, the page and a link. The quote is checked against the cited page by code; if it is not found there, the whole block is withheld.
🧠 TARS own reasoning. TARS's own explanation of the physics. It is labelled "not verified against a source". A number in it that is not from a tool or a verified quote is replaced by "[number removed: not from a source]".
When a block is withheld or a number removed, a small line says so.
Easy, Advanced and Auto
Easy uses plain wording and folds the steps and the receipts. Advanced shows the figures, the receipts, the model that answered and the steps open. Auto chooses Advanced for a question about a condition and Easy for a how-to question.
The facts are the same in every mode. Only the wording and the amount of detail change.
The skill buttons
Map the unknown takes the condition on the Envelope page and answers in three parts: what is known (with receipts), what is unknown and why, and what would close the gap.
What would close this gap shows the engine's plan of experiments for the same condition.
Both need a current condition. If there is none, TARS asks you for one.
What if we ran this experiment?
Ask TARS what would change if an experiment were run, for example a PMMA rod at 34 % oxygen and 8.2 psia in lunar gravity. TARS adds one hypothetical test to the records in memory and shows how the verdict counts and the map would change.
Every such answer carries a banner: Hypothetical: not a measurement. The simulated test is never saved to the data, and it exists only in that answer.
History and follow-ups
TARS remembers the conversation. A follow-up such as "and at 25 %?" keeps the previous condition and changes only the oxygen.
The History button opens your conversations. You can start a new chat, open an old one, rename it, delete one, or delete all. Conversations are kept in this browser only; nothing is stored on a server. An anonymous device id, a random number in a cookie, is used only for fair rate limits.
What TARS does and does not do
It does compute with the engine and calculators, search a library of NASA documents, quote them exactly, explain physics in its own words, and point at parts of the page.
It does not decide a verdict by itself, do arithmetic by itself, invent a quotation, or answer unrelated or harmful questions. It cannot browse the web. A model answers its open questions, and the page says which model and whether a fallback was used.
If no language model is available, TARS still answers questions with numbers by code, and questions about the site from the help topics.
TARS SI page overview
What it is. The full-width TARS chat.
What it shows. A conversation list, mode switch, skills, the log and the input.
How to use it. Ask a question; read the three layers.
Example. Ask about oxygen partial pressure at 34 % and 8.2 psia.
What it does not do. TARS answers verdicts only through the engine.
Pick from the page (reference)
What it is. The cursor button (an arrow) next to Send in the TARS panel.
What it shows. While it is on, anything on the page is outlined as you move over it; click it to attach its visible text to your next question as a reference (up to 280 characters).
How to use it. Click the cursor button, click a verdict card, a table or a quote, then type your question and Send. The × removes the reference; Esc cancels picking.
Example. Pick the BASS-II card on the Envelope and ask "why is this only an extrapolation?".
What it does not do. It sends only the visible text of what you picked, from this site; it reads nothing outside the page and changes no verdict.
TARS bubble
What it is. The round button at the bottom right of every page.
What it shows. It opens the TARS chat panel.
How to use it. Click it or tab to it; Esc closes the panel.
Example. Open it on any page.
What it does not do. The bubble is hidden on the TARS page itself.
Close TARS
What it is. The cross in the panel header.
What it shows. Nothing.
How to use it. Click it or press Esc.
Example. Close after asking.
What it does not do. It does not erase the conversation.
TARS chat panel
What it is. The floating chat window.
What it shows. The same chat as the TARS page.
How to use it. Type in the input; use the mode and skills.
Example. Ask what a panel on the page is.
What it does not do. It is not a human; it is the Traceable Applicability Reasoning System.
Honesty line
What it is. A line under the controls.
What it shows. "Honesty is enforced by code: the verdicts and every number come from the engine."
How to use it. Read it.
Example. Compare with an answer.
What it does not do. It describes rules enforced by checks, not a promise of perfection.
Mode switch (Easy, Advanced, Auto)
What it is. Three buttons.
What it shows. Easy: plain wording, steps folded. Advanced: figures, receipts, steps open. Auto: Advanced for a condition, Easy for a how-to question.
How to use it. Click one.
Example. Switch to Easy to fold the steps.
What it does not do. The facts are identical in all modes.
Map the unknown
What it is. A skill button.
What it shows. Builds the three-part answer (Known, Unknown, Next) for the current condition on the Envelope page.
How to use it. Click it after picking a scenario.
Example. Click it on NASA's Moon Base.
What it does not do. It needs a current condition; otherwise it asks you for one.
What would close this gap (skill)
What it is. A skill button.
What it shows. The engine's plan of experiments for the current condition.
How to use it. Click it.
Example. Use it on the Moon Base.
What it does not do. It is a PYRION proposal.
History button
What it is. A chip.
What it shows. Opens the conversation list.
How to use it. Click it.
Example. Reopen yesterday's chat.
What it does not do. Conversations are kept only in this browser.
Conversation list
What it is. The left list.
What it shows. Your conversations, newest first.
How to use it. Click one to open it.
Example. Open a chat.
What it does not do. Nothing is stored on a server.
New chat
What it is. A button.
What it shows. Starts an empty conversation.
How to use it. Click it.
Example. Start a new topic.
What it does not do. It keeps the old chat in the list.
Conversation entry
What it is. A list item.
What it shows. The chat title.
How to use it. Click it to open.
Example. Open a chat.
What it does not do. Titles are the first thing you said.
Rename conversation
What it is. A pencil button.
What it shows. Opens a name box.
How to use it. Click, type, press Enter.
Example. Rename "Moon base".
What it does not do. It renames only in this browser.
Rename box
What it is. A text field.
What it shows. The new name.
How to use it. Type and press Enter.
Example. "Lunar atmosphere".
What it does not do. It is not saved on a server.
Delete conversation
What it is. A cross button.
What it shows. Removes one conversation.
How to use it. Click it.
Example. Delete a test chat.
What it does not do. It cannot be undone.
Delete all
What it is. A link.
What it shows. Removes every conversation after you confirm.
How to use it. Click it and confirm.
Example. Clear the browser's chats.
What it does not do. It cannot be undone.
Chat log
What it is. The scrolling area.
What it shows. Your messages and TARS's answers.
How to use it. Scroll it.
Example. Read an earlier answer.
What it does not do. It is not exported anywhere.
Working status
What it is. A line with a pulsing dot.
What it shows. What TARS is doing: Checking the engine, Searching NASA documents, Calculating.
How to use it. Wait for the answer.
Example. See "Calculating…".
What it does not do. It is not a progress bar.
Message box
What it is. The text field.
What it shows. What you type, up to 300 characters.
How to use it. Type and press Send.
Example. "What is 34 % at 8.2 psia?".
What it does not do. Do not paste secrets: it is sent to the server to answer.
Send button
What it is. A button.
What it shows. It sends your message.
How to use it. Click it or press Enter.
Example. Send a question.
What it does not do. It is disabled while TARS is working.
KNOWN block
What it is. The first part of a three-part answer.
What it shows. What the extracted PSI records support at the condition, with receipts.
How to use it. Click a receipt to open its card.
Example. Read which investigations apply.
What it does not do. If nothing applies it says so.
UNKNOWN block
What it is. The second part.
What it shows. Where the evidence ends and why: a regime barrier or a gap.
How to use it. Read it.
Example. "No PSI test in this gravity regime".
What it does not do. It is about PSI, not about the world.
NEXT block
What it is. The third part.
What it shows. The nearest routes and what would close the gap, as a PYRION proposal.
How to use it. Read it.
Example. Follow a route.
What it does not do. It is not a NASA plan.
Engine fingerprint
What it is. A hex string.
What it shows. The SHA-256 of the engine's evaluation.
How to use it. Compare it across runs.
Example. The same question gives the same fingerprint.
What it does not do. It proves reproducibility of the computation, not correctness of the data.
How TARS got this
What it is. A collapsible list.
What it shows. The steps that ran: the reading, the engine, routes, the plan.
How to use it. Click to open it.
Example. Check that the engine ran.
What it does not do. It lists tools, not hidden reasoning.
Show on the envelope
What it is. A button in an answer.
What it shows. Loads the answer's condition into the dashboard.
How to use it. Click it.
Example. See the chart for the condition TARS used.
What it does not do. It replaces your current controls.
Show me (highlight)
What it is. A button in an answer.
What it shows. It scrolls to and outlines a part of the page.
How to use it. Click it.
Example. Show me "Where the evidence is".
What it does not do. It only points; it changes nothing.
Receipt link in an answer
What it is. A small link.
What it shows. An investigation id.
How to use it. Click it to open its card.
Example. Open PSI-69.
What it does not do. It opens the Envelope page card.
DOI link in an answer
What it is. A link.
What it shows. The investigation's DOI.
How to use it. Click it.
Example. Open the dataset.
What it does not do. It is shown in Advanced.
Open the document
What it is. A link under a NASA quote.
What it shows. The NASA document that contains the quote.
How to use it. Click it.
Example. Read the page around the quote.
What it does not do. The quote was checked against the cited page by code.
Engine layer
What it is. The 🔒 layer of an expert answer.
What it shows. Facts restated from engine tools or calculators.
How to use it. Read it as computed by code.
Example. "19.22 kPa".
What it does not do. Numbers there come only from tool results.
NASA layer
What it is. The 📄 layer.
What it shows. A paraphrase with an exact quote, document title and page.
How to use it. Click the document link.
Example. A quote about FM2.
What it does not do. If the quote is not found on the page, the whole block is withheld.
Own-reasoning layer
What it is. The 🧠 layer.
What it shows. TARS's own reasoning, labelled "not verified against a source".
How to use it. Treat it as an explanation, not a source.
Example. "Why a flame is round in microgravity".
What it does not do. Numbers without a source are replaced by a removed-number marker.
Help topic suggestion
What it is. A link in a Support answer.
What it shows. The title of a help topic.
How to use it. Click it to open the topic.
Example. Click "Add a source (as built today)".
What it does not do. It asks Support again.
Take me there (help)
What it is. A button in a help answer.
What it shows. Goes to the page and highlights the control.
How to use it. Click it.
Example. Go to the Add a source button.
What it does not do. It moves you; it does not click the control.
Brief me
What it is. A panel that turns this conversation into a document.
What it shows. A preview of the brief with a bibliography and the engine fingerprints.
How to use it. Press Build the brief, then download it or print it.
Example. Brief a conversation about the ISS cabin air.
What it does not do. Only checked statements are included; the model is not called.
Brief me (button)
What it is. A skill button in the TARS panel.
What it shows. Opens the Brief me panel.
How to use it. Click it after you have asked a few questions.
Example. Ask two questions, then click it.
What it does not do. It needs at least one answered question.
Brief preview
What it is. The brief as it will print.
What it shows. The questions, checked findings, quotes and the bibliography.
How to use it. Scroll it, or print it.
Example. Find the engine fingerprint at the end of a question.
What it does not do. It is a record of this conversation, not a verdict.
Build the brief
What it is. The button that builds the document.
What it shows. Nothing until pressed.
How to use it. Press it after asking questions.
Example. Press it, then read the preview.
What it does not do. It recomputes the engine part and re-verifies every NASA quote.
Check a claim
What it is. A panel where you paste a text and each claim in it is classified.
What it shows. A table of claims, each with one class: supported in PSI, not supported in PSI, contradicted or overclaim.
How to use it. Paste the text, press Check the claims, read the table.
Example. A claim naming PSI-69 and a verdict at 30% oxygen is checked against the engine.
What it does not do. It never calls a claim false because PSI is silent: that is shown as not supported in PSI.
Check a claim (button)
What it is. A skill button in the TARS panel.
What it shows. Opens the Check a claim panel.
How to use it. Click it, paste a text, press Check the claims.
Example. Paste a paragraph that says a flame is safe at a condition.
What it does not do. It does not decide anything by itself: classes come from code.
Check the claims
What it is. The button that runs the check.
What it shows. A model may help split the text, but every class is decided by code.
How to use it. Press it once the text has at least a sentence.
Example. Paste a paragraph, press it, wait a few seconds.
What it does not do. It counts against the same limits as a question.
Claims table
What it is. The result of a check.
What it shows. Each claim, its class badge, the reason in one fixed sentence, the engine reference with its fingerprint, and NASA quotes that code found on the cited page.
How to use it. Hover a quote to see its page; hover a badge to see what the class means.
Example. A contradicted claim shows the verdict the engine gave.
What it does not do. A discarded proposal is counted, never shown.
Close the skill panel
What it is. A link that closes the open skill panel.
What it shows. Nothing.
How to use it. Click Close.
Example. Close the claims table.
What it does not do. It does not delete your conversation.
Compare
What it is. The button that runs the comparison.
What it shows. Nothing until pressed.
How to use it. Press it with at least two conditions chosen.
Example. Pick two scenarios and press it.
What it does not do. It is computed by code and does not use a model.
Compare atmospheres
What it is. A panel that puts several conditions side by side.
What it shows. A table with oxygen, total pressure, oxygen partial pressure, how many investigations are comparable, solid-fuel status and the nearest evidence, and a ranking rule.
How to use it. Pick scenarios or write your own, then press Compare.
Example. ISS cabin against a 34% oxygen, 8.2 psia cabin.
What it does not do. The ranking says how much measured evidence PSI holds near a condition, not which is better.
Compare atmospheres (button)
What it is. A skill button in the TARS panel.
What it shows. Opens the Compare atmospheres panel.
How to use it. Click it, pick two or more scenarios, press Compare.
Example. Compare the ISS cabin and NASA's Moon Base.
What it does not do. It does not say which atmosphere is safer.
Comparison table
What it is. The result of a comparison.
What it shows. One row per condition, and the ranking rule under the table.
How to use it. Read across a row; read the rule under the table.
Example. See which condition has more comparable investigations.
What it does not do. It never estimates a number PSI does not hold.
Download .md
What it is. Saves the brief as a Markdown file.
What it shows. Nothing; it saves a file.
How to use it. Press it after building.
Example. Keep the file with your notes.
What it does not do. The file is made in your browser from the built brief.
Print / PDF
What it is. Opens the print dialog for the brief.
What it shows. Nothing; it prints the preview.
How to use it. Press it and choose Save as PDF.
Example. Hand the PDF to someone who can check it.
What it does not do. Printing is done by your browser.
Scenarios to compare
What it is. A list of the site's scenarios.
What it shows. One checkbox per scenario.
How to use it. Tick two or more.
Example. Tick ISS cabin air and NASA's Moon Base.
What it does not do. It lists only the scenarios the site already has.
Show these points on the map
What it is. A button that draws the compared points on the Envelope map.
What it shows. Numbered diamonds on the map, one per condition.
How to use it. Press it; the Envelope page opens if you are elsewhere.
Example. Numbers on the map match the numbers in the table.
What it does not do. It draws points only; verdicts stay on the cards.
Text to check
What it is. A box for the text.
What it shows. Nothing yet.
How to use it. Paste up to 3000 characters.
Example. An abstract or a paragraph from a report.
What it does not do. Nothing you paste is stored on the server.
Your own condition
What it is. Three boxes: name, oxygen in percent, pressure with unit.
What it shows. Nothing until you compare.
How to use it. Write the oxygen as a number and the pressure with its unit, such as 8.2 psia.
Example. 34 and 8.2 psia.
What it does not do. It refuses a unit it cannot read instead of guessing.
Replay
What it is. A button under an answer that has a condition.
What it shows. Runs only the engine part again and says same fingerprint ✓ when the bytes match.
How to use it. Click it on any answer.
Example. Replay an answer from earlier today.
What it does not do. It does not ask a model anything, and it does not redo the model's words.
Explanation layer unavailable
What it is. A line above an answer built by code alone.
What it shows. That no model could answer right now, and that everything shown comes from code and NASA documents.
How to use it. Read it; ask again later for a model's explanation.
Example. Ask about a condition while the models are busy and still get the engine's verdicts.
What it does not do. It does not explain or reason; it only retrieves.
Help topic (code answer)
What it is. A help topic TARS found by code when no model was available.
What it shows. The topic text from the site's own help.
How to use it. Read it; ask again later for a model's explanation.
Example. Ask about a condition while the models are busy and still get the engine's verdicts.
What it does not do. It does not explain or reason; it only retrieves.
Privacy line
What it is. The small line under the TARS message box.
What it shows. That there is no account, that conversations stay in this browser, and that an anonymous device id (a random number in a cookie) is used only for fair rate limits.
How to use it. Read it; there is nothing to click.
Example. Several people on the same Wi-Fi each have their own allowance because each browser has its own random id.
What it does not do. The id is not linked to a name, an email or an account, and it does not carry your conversations. Clearing cookies gives a new id, but the limit of the whole network still applies.
Citation preview
What it is. A NASA quote in an answer. It is focusable and reacts to hover.
What it shows. On hover or focus, the real text of the cited page from the NASA library, with the quoted sentence highlighted, and the document link.
How to use it. Hover it, or tab to it, to read the quote in context. Press Esc to close the preview.
Example. Hover a quote about FM2 to see the whole page it comes from.
What it does not do. It does not re-check the quote (code already did that before the answer was shown), and it shows one page, not the whole document.
Suggested next questions
What it is. Up to three small buttons under an answer.
What it shows. Follow-ups that fit the answer, for example what would close the gap, a comparison with ISS air, or showing the condition on the envelope.
How to use it. Click one to ask it, or to load the condition.
Example. After an answer about 34 % oxygen at 8.2 psia, click "Compare with ISS air".
What it does not do. The suggestions are fixed templates chosen by code from what the answer used; they are not written by the model and they are not predictions.
Ask 3 models
What it is. A button under an answer in Advanced mode.
What it shows. The same question asked to GLM-5.3, Qwen and Nemotron, side by side, with the engine fingerprint of each run. If the fingerprints are identical the page says "Same verdict, different words": the models may phrase things differently, but the engine computed the same thing.
How to use it. Click it once; it takes a while because three models run.
Example. Compare how three models explain the physics of the same condition.
What it does not do. It does not run on its own (it uses credit and counts against your limits), it never substitutes one model for another, and a model that is not available says so. The three answers are each checked by code; their expert layers are still unverified explanations.
Second opinion
What it is. A strip under an answer in Advanced and Auto modes.
What it shows. A review of the checked answer by a model of a different family than the one that wrote it: it agrees, objects (with the reason, and a NASA quote that code verified), or is unavailable. The reviewing model is named.
How to use it. Read it after the answer; hover an objection's quote to see its page.
Example. "Second opinion: agrees · Nemotron 3 Ultra".
What it does not do. It never changes the answer. A reviewer's wording is scrubbed of numbers that are not in the answer, and an objection's quote is dropped if code cannot find it on the cited page. It is an opinion, not a verdict.