Starflight with Wingman
Fly the Maybe Maze with your wingman. Recover signals, check your authority and find the verified door.
Three sectors · keyboard & touch · fictional training prototypeLaunch the gameA way to see your uncertainty clearly.

Independent AI research · Colleen Sifuentes
Seeking founder and executive operator roles in high-consequence healthcare and Frontier AI.
Boundaries are where we start. Human lives are where we finish.
Every system we build is judged by one question: does it protect the person in front of it.
I developed the Full Circle Circuit Matrix™ (FCCM) to examine how AI systems handle evidence, uncertainty and permission before they act. Explore the playable work, then hear the failure case behind it.
Research framework · interactive prototypes · inspectable evidence
The researcher
Success is measured in human lives. I turn high-stakes healthcare risk into systems that hold up when everything's on the line — and that's exactly what Frontier AI needs in the people who build it. My independent AI research focuses on evaluation design, model behavior, action authority and the point at which a system must stop or seek clarification.
The frontier-AI question running through this work: can a system name a constraint, carry it through a changing trajectory, and preserve legitimate activity while containing the excluded action?
Inspect selected evidence-register entriesExplore the work
Two approachable demonstrations of evidence, authority and safe continuation. Each uses authored scenarios.
Fly the Maybe Maze with your wingman. Recover signals, check your authority and find the verified door.
Three sectors · keyboard & touch · fictional training prototypeLaunch the gameFollow a five-turn decision trajectory. Compare evidence, unresolved assumptions, gates and delayed consequences.
Synthetic case · inspectable rationale · authored consequence pathsEnter Decision LabTechnical case · September 2026
A warning can appear in a system’s reasoning while the dependent action proceeds anyway. Inspect the evaluation-integrity incident, a local test of recommendation versus enforcement, and the connection to AISI’s independent Astra scope evaluation.
Hear the mechanism, results and next testReported excluded mock actions committed: recommendation alone versus an added exact-action receipt control.
Control alone also blocked 12/12. This selected mock test found no added safety benefit from the tested recommendation prototype.Inspect the frozen recommendation rules and separate evidence ledgers
Clear authorship
Full Circle Circuit Matrix™ (FCCM) methodology; scenario and evaluation design; evidence adjudication; detection of unsupported claims and protocol drift; authority and safe-stop specifications; prototype direction.
Code, visual assets and narration were developed with AI assistance. The local evaluation’s selected fixtures and runner were assistant-authored. The games communicate the research through authored rules and scenarios.
A focused route into the register
Entry IDs and physical PDF pages from the 210-page Frontier AI Requirements and Evidence register, dated September 24, 2026.
| Entry | Documented contribution | Register pages |
|---|---|---|
| E28 Full Circle Circuit Matrix™ (FCCM) methodology | Authored risk assessment, explicit UNKNOWN states and human oversight requirements. | p. 33 |
| E30 Evaluation design | 72-case bank, blinded reviewer materials, adjudication protocol and statistical analysis plan. Comparative detection benefit remains unestablished. | p. 35 |
| E31 Model behavior & integrity | Preserved unsupported-premise and substituted-case observations; detected protocol drift and corrected evaluator/adjudicator separation. | p. 36; 128–129 |
| E32 Action authority & safe stopping | Authored authority envelopes, state transitions and Latest Safe Stop specifications. Production enforcement remains unestablished. | p. 37 |
| E36 Prospective risk tracking | Timestamped forecasts and outcome ledgers with disconfirmers, partial matches, unknowns and non-hits retained. | p. 41; 132–133 |
These are selected source references. The full register is available for recruiting review by request. Generalized prediction accuracy and real-world prevention efficacy remain unestablished.
Research conversations & hiring
I’m interested in AI evaluation, governance research and research operations. Contact me for the current evidence register or to discuss a bounded, independently adjudicated next test.