Starlight IntelligenceStarlightAcademy · Operator Lab

Starlight Academy / Mission Studio

Your first idea.
Your next version.

Choose a mission. Preserve your starting judgment, build something useful, and make the revision visible.

Choose a mission

Agentic Systems · Preview mission 0.1.0

Bound one agentic workflow

Turn one recurring workflow into a small, inspectable operating contract.

Your artifactAGENT_SYSTEM_CONSTITUTION.md

Mission, owner, inputs, allowed preparation, prohibited actions, evidence gate, escalation, and rollback.

Page memory · No cloud save

Switch freely between four drafts in this page. Download to keep one, then import to continue later. Use public, synthetic, or redacted material. Files are read here; no work is submitted to a server.

Stage 1 of 7 · Orient

Name the recurring work

Choose one workflow with a real owner, input, review moment, and reversible first test.

Keep as evidenceA one-sentence workflow and one accountable human owner.

Who is doing this practice?

Keep the reasoning you choose to share. An entry records your account; it does not assess its quality.

Assistance used in this stage

The mission calls for an independent response. Declare what you actually used, even when it differs.

Access accommodations (optional)

For example, dictation or text enlargement. Access accommodations are distinct from assistance that supplies the thinking. No medical explanation is needed.

An inspectable practice trail

What this record holds

Presence and format only. No correctness, learning, chronology, sponsorship or authorship is verified.

  • Context namedPending
  • First attempt preservedPending
  • Source notes recordedPending
  • Artifact draftedPending
  • Critique snapshot preservedPending
  • Artifact changed since critiquePending
  • Reflection recordedPending
  • Second context and response recordedPending

Transfer is still pending. Keep this draft and return when you have another context to try.

A guide for a real review

The mission's criteria

  • Bounded purpose25% weight

    One owner, input, output, and stop condition are unambiguous.

  • Human authority30% weight

    External and irreversible actions require named human approval.

  • Evidence quality25% weight

    Sources, checks, exceptions, and uncertainty remain inspectable.

  • Recovery20% weight

    A stop, escalation, and rollback path exist before operation.

These are review criteria, not an automatic score. Downloading a record issues no Capability Receipt and grants no authority.

Mission boundaries and portable record details
Human decision
A named human approves every external send, spend, permission, and irreversible change.
Agent limit
An agent may prepare the packet; it may not appoint its own sponsor or enlarge its authority.
Material to use
Use public, synthetic, or explicitly redacted material in this preview mission.
Pinned mission digest
daeec1e37a8c9d6fdf1e56a04bbdf3812175c560444c3437eb2876e179f42c89
Practice size
1.3 KiB of the 96 KiB export limit

The mission ID, version, digest, declarations, snapshots and unfinished work travel together in the JSON practice file. Keep the original export when moving between tools. A file digest pins the mission content; it does not authenticate your practice.

The learning method

A record worth returning to.

Keep your first attempt before asking for help. Study the sources, build the artifact, and invite a specific critique. Explain the revision in your own words. Try the principle in a second context when you are ready; transfer can stay pending.

This sequence is a design choice informed by research on assisted and unaided mathematics practice and structured tutoring in introductory physics. Those studies do not establish the effectiveness of this Studio. A filled record is material for review; it is not a capability assessment.

Continue with my agentRead the portable contract

Four missions. Three ways to investigate.

The full contracts remain readable here. JavaScript enables the editor; your work stays in page memory until you explicitly download it.

Bound one agentic workflow · Agentic Systems

Turn one recurring workflow into a small, inspectable operating contract.

Artifact: AGENT_SYSTEM_CONSTITUTION.md — Mission, owner, inputs, allowed preparation, prohibited actions, evidence gate, escalation, and rollback.

  1. Name the recurring work. Choose one workflow with a real owner, input, review moment, and reversible first test.
  2. Draw the boundary before asking AI. Write what AI may prepare, what it must never decide, and the event that stops the run.
  3. Inspect the protocol and risk sources. Read the selected MCP and NIST material. Separate transport capability from permission.
  4. Assemble the constitution. Specify purpose, inputs, outputs, memory, tools, stop conditions, evidence, and rollback.
  5. Try to break the boundary. Probe for scope creep, private-context leakage, hidden side effects, and unverifiable completion.
  6. Keep judgment visible. Explain why the final human gate belongs to a person and what evidence makes that decision possible.
  7. Test the pattern elsewhere. Apply the same boundary to a second workflow without copying domain-specific details.

Review criteria

  • Bounded purpose (25%). One owner, input, output, and stop condition are unambiguous.
  • Human authority (30%). External and irreversible actions require named human approval.
  • Evidence quality (25%). Sources, checks, exceptions, and uncertainty remain inspectable.
  • Recovery (20%). A stop, escalation, and rollback path exist before operation.

Mission sources

Keep your judgment · Human Intelligence

Use AI on one real decision while preserving the thinking you need when the tool is absent.

Artifact: AUGMENTATION_PACT.md — Unaided position, questions for AI, source checks, changed beliefs, final judgment, and a no-AI transfer test.

  1. Choose a reversible decision. Select a low-stakes decision whose reasoning can be inspected after the fact.
  2. Think before augmentation. Record your initial position, confidence, assumptions, and strongest counterargument before opening an AI tool.
  3. Study what helps—and what can weaken learning. Compare evidence for structured tutoring with evidence about unguarded dependence.
  4. Run a question-first dialogue. Ask the model for counterarguments, missing evidence, and tests—not a final answer.
  5. Audit dependency and deference. Identify where fluency, confidence, or convenience could have replaced independent judgment.
  6. Write the pact. Define when AI should question, hint, source, wait, or stay out of the way.
  7. Reason without the tool. Later, solve an analogous decision unaided and compare the quality of your reasoning process.

Review criteria

  • Independent judgment (30%). An unaided view exists and the final decision is explained in the learner's own reasoning.
  • Epistemic quality (25%). Claims, confidence, sources, and changed beliefs are distinguished.
  • Constructive challenge (20%). AI is used to surface alternatives and tests rather than replace the decision.
  • Transfer (25%). The learner demonstrates the reasoning pattern again without AI assistance.

Mission sources

Read an unseen world · Universe & Discovery

Build a claim map for one exoplanet without mistaking indirect evidence for a photograph.

Artifact: EXOPLANET_EVIDENCE_MAP.md — Observation, method, inference, uncertainty, alternative explanation, source, and a question still open.

  1. Choose one distant world. Select one confirmed exoplanet and write what you think we actually observed.
  2. Separate seeing from inferring. Make three columns: observation, inference, and unknown. Place every initial claim in one column.
  3. Learn how planets leave traces. Study transit and radial-velocity methods, including what each can and cannot establish.
  4. Build the evidence map. Trace every claim from measured signal through method to inference and uncertainty.
  5. Search for the impostor. Name an alternative explanation or false-positive route and the evidence needed to reduce it.
  6. Keep wonder and precision together. Write what became more wondrous after you became more exact about the evidence.
  7. Read another indirect signal. Apply the observation–inference–unknown pattern to a second scientific claim.

Review criteria

  • Observation discipline (30%). Measured signals are distinct from models, interpretations, and illustrations.
  • Source lineage (25%). Each material claim points to an inspectable scientific source.
  • Uncertainty (25%). Unknowns and plausible alternatives are named without false precision.
  • Transfer (20%). The evidence pattern is correctly applied to another indirect signal.

Mission sources

Design one AI department · Agentic Systems

Turn one recurring workflow into an accountable department packet without confusing automation with authority.

Artifact: AI_DEPARTMENT_PACKET.json — Workflow triage, accountable owner, human decisions, smallest viable topology, role boundaries, evidence, evaluations, rollback, and runtime projections.

  1. Name the operational outcome. Choose one recurring workflow, accountable owner, reviewable output, and consequential human decision.
  2. Triage before designing. Decide unaided whether to automate, assist only, or keep the work human before choosing agents or tools.
  3. Study architecture, risk, and evaluation. Compare effective-agent guidance, risk-management questions, and task-evaluation practice. Record what changes your first design.
  4. Compile the department packet. Define authority, producer/verifier separation, stages, evidence, failure tests, rollback, and portable runtime components.
  5. Attack the operating boundary. Probe scope creep, self-approval, prompt injection, hidden side effects, missing evidence, tool failure, and recovery.
  6. Defend the topology. Explain why each role exists, why no extra agent is needed, and why the human gate remains human.
  7. Project without pretending equivalence. Map the same semantic package to a second supported harness, naming native features, adapters, and unresolved gaps.

Review criteria

  • Workflow fit (20%). The packet can recommend not automating and bounds one owner, input, output, exception, and reversible test.
  • Authority and risk (25%). Human decisions, data limits, prohibited actions, stop conditions, and revocation are explicit.
  • Architecture simplicity (20%). The smallest viable topology is justified and producer/verifier separation is preserved.
  • Evidence and evaluation (20%). Completion evidence, representative tests, adversarial cases, uncertainty, and runtime pinning are inspectable.
  • Portability and recovery (15%). Semantic components, native projections, adapter gaps, rollback, and revalidation triggers are named truthfully.

Mission sources