Mission Copilot · Research preview

Answers that cite their receipts.

Ask a mission question in plain language. Mission Copilot answers only from mission evidence, cites the signed receipts behind every claim, and tells you when the evidence is not there. The receipts check offline; the AI never becomes the source of truth.

StatusResearch preview
Answers on this pageScripted examples
EvidenceHybrid-signed receipts
00 The problem

A fluent answer is not evidence.

Mission teams need to know why the rover stopped, not a confident paragraph about it.

General-purpose assistants write well and cite loosely. In mission operations that is dangerous: a plausible answer about a thermal limit or a missed pass can send a team the wrong way for days, and nobody can later show which data the answer came from.

Mission Copilot is designed the other way round. It works over the mission's own receipt chain: commands, intents, decisions, health records, pass plans and imagery provenance. Each sentence in an answer names the receipts it relies on. If no receipt supports a claim, the copilot says so instead of guessing.

Because every R-Series receipt carries Ed25519 + ML-DSA-65 signatures and a parent link, a reviewer can check the cited records themselves, offline, without trusting the copilot or COSRYX.

01 Scripted preview

Pick a question. Watch the citations light up.

These answers are pre-written to show the citation pattern. They are not generated by a model, and the receipts are simulated.

Mission CopilotSIM-MARS-07 · rover receipts sol 1,150–1,190Scripted · Simulated
The rover entered a safe hold on sol 1,184: measured slope reached 21°, above the 18° bound in its signed intentR1R2. The intent allowed a reroute without asking Earth, and the reroute ran two sols laterR3.
Navigation packet A7 first, then anomaly report B17. Both are marked mission-critical in the current planR4, and the next orbiter pass has budget for both before imagingR5.
Slowly, yes. Current draw on wheel motor 4 is up about 11% across 212 health receipts since sol 1,150R6. No limit has been crossed. No receipt explains the cause yet; the copilot will not guess one.
During 97 sols without contact the rover sealed 4,118 records on board, in one unbroken sequenceR7. Every autonomous decision in that window traces to the signed intent in forceR2.
Cited receipts Simulated
R1Hazard decision · sol 1,184slope 21° > bound 18° · safe hold · sha256 9c1e…02fa
R2Signed intent bundle · sol 1,180bounds: slope 18°, energy 40%, 3 sols · Ed25519 + ML-DSA-65
R3Reroute decision · sol 1,186parent R1 · within bounds · seq 41,140
R4Mission plan · priorities v12A7, B17 marked critical · signed by ops lead
R5Pass plan · orbiter Bwindow 11 min · budget 38 MB · DSRP timestamp
R6Mobility trend · 212 health receiptsdraw +11% · seq 41,050–41,262
R7Blackout chain · seq 37,022–41,1394,118 sealed records · no gaps
Scripted research preview: answers are pre-written and receipts are simulated, to show how citations work. Nothing here is generated from spacecraft data.
02 One loop per answer

Retrieve, check, answer, cite.

Question→Retrieve receipts→Check both signature legs + chain→Compose answer→Cite each claim→Human decides→Decision receipt
01 · RETRIEVE

Only mission records

The copilot searches the mission's receipt chain and attached documents, not the open web. The retrieval set is itself recorded.

02 · CHECK

Unverified evidence is flagged

A record with a broken chain or a missing signature leg is marked as such and is never cited as fact.

03 · CITE

Every claim, a receipt

Each sentence carries the identifiers it depends on. Where nothing supports a claim, the answer says "no evidence" in plain words.

03 Inside an answer

From a question to a decision receipt.

Copilot01 / 06
01 · ASK

A question in plain language.

"Why did the rover stop short of waypoint 4?" The question is recorded with who asked it and when, so the answer can be replayed later.

02 · RETRIEVE

Find the records that matter.

The copilot pulls the receipts near the event: the hazard decision, the intent that was in force, the reroute that followed.

03 · CHECK

Verify before quoting.

Each retrieved receipt is checked: canonical bytes, SHA-256 digest, the Ed25519 leg and the ML-DSA-65 leg, and its place in the chain. A record that fails is not used as fact.

04 · ANSWER

Short, specific, bounded.

The answer states what the evidence shows, and only that. Confidence comes from the evidence, not from fluent wording.

05 · CITE

Every claim points home.

Citations link each sentence to its receipts, so an engineer or a reviewer can open the exact records and check them offline.

06 · DECIDE

A person approves, and that is a receipt too.

If the answer leads to an action, the operator's approval is recorded as a decision receipt that cites the answer it relied on.

04 Questions it is designed for

The questions mission teams actually ask.

WHY

Why did it do that?

Trace an autonomous decision back to the intent and bounds in force, and the sensor readings it saw.

WHAT FIRST

What should go down first?

Rank the downlink queue against the plan's priorities and the next pass budget.

TREND

Is something wearing out?

Summarise a subsystem's health receipts over time and show the records behind the trend.

GAP

What happened while we were dark?

Walk a blackout backlog, confirm the sequence is unbroken, and list the decisions taken.

WHO

Who approved this?

Find the human approval or signed policy behind any command or high-impact action.

PROVE

Can we show a third party?

Export the cited receipts as a bundle a reviewer checks offline, without COSRYX in the loop.

05 Guardrails

The copilot advises. People and signed policy decide.

Mission Copilot has no command authority. It cannot uplink, change a plan or widen a rover's bounds. High-impact actions go through a human approval, and the approval is receipted.

When the copilot uses tools, such as querying a telemetry store, the design follows COSRYX-HYBRID-03: an air-gapped MCP connector that produces local, chained receipts for each tool call without storing raw tool arguments or outputs (28/28 tests; Built · definition-of-done sign-off pending).

Copilot actionDefaultWhat the record showsStatus
Answer a questionAllowedQuestion, retrieval set, cited receiptsResearch preview
Call a mission toolAllowed, read-onlyLocal chained tool receipt, no raw payload keptBuilt · definition-of-done sign-off pending
Recommend an actionAllowedRecommendation with its citationsResearch preview
Uplink a commandNot possible from the copilotOnly a human or a signed intent can commandNot a copilot capability
06 What it reads

Evidence from built COSRYX products.

The copilot is new; the records it reads are not. They come from products that are built and tested on the ground.

SourceRecordsStatus
COSRYX-EDGEOnboard autonomous decisions, sealed and chained (31 tests)Built · Simulation-validated
COSRYX-MCCMission-control commands and approvals (28 tests)Built · Simulation-validated
DSRPRelay pass plans and light-cone timestamps (78 tests, 100/100 scenarios)Simulation-validated
COSRYX-NEXUSCustody receipts across delayed, interrupted links (65 tests)Built · Simulation-validated
COSRYX-LENSImagery provenance for each derived product (29 tests)Built · Simulation-validated
Health heartbeatSigned subsystem health every 5–10 minutesIn development

If the evidence is not there, the answer says so.

07 Questions

About the research preview.

Which model does the copilot use?

The research preview is model-agnostic by design: the citation and verification layer sits around the model. Model choice, hosting and keys are agreed per pilot. Bring-your-own-key is a design target.

Can it be wrong?

Yes. That is why every claim cites receipts and why it has no command authority. A reviewer checks the receipts, not the prose.

Does mission data leave our boundary?

Pilot data boundaries are agreed in writing before any data is shared. Sovereign, private-cloud and on-premises deployment are a future direction, not something we offer today.

Is the scripted preview the product?

No. It shows the citation pattern with pre-written answers and simulated receipts. A pilot runs on your own or simulated mission records.

08 Pilot

Ask your own mission records.

A copilot pilot starts with one question set over simulated or historical records, and ends with an evidence bundle your reviewers can check.

Request pilot accessSecurity model