Open the third-party AI Assistant (opens in a new tab)The Assistant is a third-party ChatGPT experience operated by OpenAI. Do not submit sensitive information, confidential records, or material that is restricted from third-party processing. This third-party ChatGPT experience should not receive sensitive information.
Know what happened. Know when it happened. Know how sure we are.
Human Performance OS creates a common language for observations, context, digital-twin estimates, recommendations, experiments, outcomes, and scientific evidence. Every conclusion retains its timing, origin, quality, uncertainty, applicability, privacy classification, and permitted purpose.
Architecture in review · Implementation not started
One observationMeaning + trust move together
01Observation
02Context
03Estimate
04Proposal
05Outcome
TimeSourceQualityUncertaintyPurpose
T TimeS SourceQ QualityU UncertaintyP Purpose
01Follow one piece of information
A report becomes useful only by passing through clear boundaries.
Jordan reports lower-than-usual sleep after travel. The system may organize the report, add context, and propose a reversible adjustment. At no point does the report become a diagnosis or an automatic decision.
◇The system proposes.The person decides.
Step 1 of 8
Observed record
Observation
Jordan reports lower-than-usual sleep after travel.
Metadata that must stay attached
Observed timeJordanSelf-reportWellness purpose
✓ Allowed to claim
Record what Jordan reported and when.
× Not allowed to claim
Claim that the report explains how Jordan will perform.
02Three rules
Change how you read human-performance data.
These three distinctions prevent confident-looking information from outrunning what the evidence can support.
Rule 01
≠
selected map
The twin is not the person
A digital twin is a purpose-bounded, probabilistic representation.
It can help organize estimates for a defined question. It is not a complete identity, permanent score, diagnosis, or ground truth.
Show why
A snapshot selects particular observations, contexts, and assumptions for a stated purpose. Change the purpose or evidence, and the representation may need to change.
Rule 02
QualityUncertainty →
High qualityLower uncertainty
High qualityHigher uncertainty
Lower qualityLower apparent uncertainty
Lower qualityHigher uncertainty
Quality is not certainty
Good data can still support an uncertain conclusion.
Quality asks whether the inputs are trustworthy. Uncertainty asks how sure we should be about the conclusion.
Compare examples
Complete, recent input. The conclusion can still be uncertain because travel context is complex.
Show why
Uncertainty may come from the model, the context, natural variation, or an outcome that has not happened yet—even when the source data is excellent.
Rule 03
08:00Event A
· · · · · →
11:00Event B
Next ≠ caused by
Sequence is not causality
What happened next was not necessarily caused by what happened first.
A temporal sequence can support investigation. It does not prove that one event caused another.
Show why
Other influences, missing context, coincidence, or selection effects may explain the sequence. A causal claim needs a stronger design and evidence standard.
03Interactive trust check
Can I trust this recommendation?
Inspect the conditions behind a proposal. Change the example states to see when the responsible answer becomes “more evidence” or “abstain.”
Illustrative proposal
Jordan may consider reducing cognitive load for the next work block and testing one small recovery adjustment.
A fictional, non-medical example. Not medical or professional advice.
04Versioned knowledge
Late data never rewrites history.
New evidence creates a new version. It never silently changes what the system knew before.
1Observation received
2Twin snapshot v1 created
3A late correction arrives
4Twin snapshot v2 created
Compare what was known
Snapshot v1
Travel report: one short night
Available information
Jordan’s initial report and known travel context
Conclusion
Consider a lighter first work block
Original record · retained
+ late correction →
Snapshot v2
Correction: report referred to local time
Available information
Original report, correction, and revised timing
Conclusion
Timing signal is less unusual; reassess the proposal
New record · linked to v1
At 9:00 AM, only the original report and travel context were available. Snapshot v1 preserves the conclusion produced from that evidence.
05Purpose propagation
Privacy travels with the information.
A useful inference is not automatically an allowed inference. Derived records inherit restrictions from their sources.
Source APrivate wellness data
Permitted: coaching · personal review
Source BRestricted contextual data
Permitted: personal review
Derived recordRestricted · personal review only
Most restrictive class + shared purpose
Evidence→Estimate→Proposal│Human decision
Authority boundary
People remain responsible for human decisions.
The system may organize evidence, estimate state, and propose an action. A person or properly authorized decision-maker retains final authority.
06Refusal rules
What Human Performance OS refuses to do.
A trustworthy architecture must define what happens when a claim or use crosses the boundary.
01×
Present an estimate as a diagnosis
Estimate ≠ diagnosis
02×
Hide material uncertainty
Show the range
03×
Treat sequence as proof of causality
Next ≠ caused by
04×
Use information beyond its permitted purpose
Purpose required
05×
Silently rewrite earlier conclusions
Version, don’t erase
06×
Turn a recommendation proposal into an automatic human decision
Human authority
Trust is not created by making the system sound certain. Trust is created by showing what it knows, what it does not know, and what it is permitted to do.
07Architecture transparency
Separate what is accepted, proposed, and not started.
Status is part of the trust model. Proposed work never appears as approved or operational.
✓
Accepted architecture
Current design authority
Thin semantic kernel with operational projections
Narrow semantic and contract authority
Class-based conformance
Nine bounded ontology modules
Canonical entity and relationship boundaries
Immutable observations and twin snapshots
Contextual Cognitive Capacity Reserve
Scientific and causal guardrails
◇
Proposed architecture
Under review · not accepted
Mandatory operational envelope
First-class temporal and provenance graph
Multidimensional quality and missingness model
Typed uncertainty and calibration model
Evidence-certainty and applicability model
Privacy-classification and purpose-propagation rules