Somebody on the team asks the bot a question in the channel where they work: why don't we allow the integration the client keeps asking for? (uncommitted) Every company has a bot that will answer that now. (uncommitted) Most of them will paraphrase a help page with total confidence, and the confidence is the problem, because a wrong answer and a right answer sound identical coming out of a language model. (uncommitted) Worse than useless, actually: an answer with a stale citation attached inherits credibility it has not earned.1 The person reading it has no way to tell, so they either trust the vendor's machine or they go interrupt the person who knows. (uncommitted)
A perspective-grounded agent answers with receipts.2 The claim arrives carrying the person who decided, the date, the conversation it happened in, and the diverging view if one was recorded.2 The teammate reading it sees that the integration was ruled out by the chief architect in March, in these words, over this concern, and that a member of the team disagreed and why. (uncommitted) The agent never asks to be trusted.2 It shows where the answer came from and lets the reader decide whether they trust the person it cites, which is a decision humans already know how to make.2
Just as important is what the agent does at the edge of the record.3 It knows what the organization has settled and what it has not, and when a question reaches past the perspective it stops, escalates to a human in the same channel, and names what it was missing.3 The escalation is principled, on defined triggers: a situation the record does not cover, a contradiction with a settled principle, tension between claims, or input from the other side that the perspective cannot ground.4 That discipline is the trust contract that makes delegation real.5 The expert is reachable exactly when the agent is out of its depth, and never interrupted for what the record already answers.5
The loop closes in observation. Everything the agent says in your name is reviewable with the same citations its recipients saw, so checking on your delegate takes minutes, and what the agent could not answer becomes the roadmap: every escalation names a conversation worth having, and the next extraction turns it into something the agent can answer tomorrow.6
During a March 19, 2026 session on bounded-domain fidelity, Scarlet Dame warned of a critical safety inversion in organizational memory systems:
"stale graph + citations = worse than having no citations"
An inaccurate graph, when cited by an AI delegate, lends unearned credibility to incorrect information. Providing stale references is actively more damaging than delivering generic AI responses with no citations at all, which is why freshness and provenance are inseparable in the product's design.
In a June 4, 2026 design call, Scarlet Dame established the core mechanism for grounding agent responses in verifiable human history:
"Design goal: LLM must cite every claim as it works, with footnotes attributing statements to named actors in named conversations."
This design enforces anti-hallucination by requiring the agent to present its receipts directly. It shifts the trust model away from blind machine authority and lets the reader evaluate the original human context and decisions instead.
During a June 30, 2026 session, Scarlet Dame pushed agent design past task execution into self-knowledge at the boundary:
"can we define not only the knowledge but the bounds of task, and also understanding and intent and goal and interaction, the world limits of an agent? So that it can execute the task — but also so it can know when it cannot finish the task, and when it needs to escalate."
She demonstrated the mechanism live by asking the system what was missing from its own understanding: gap-awareness as the structural basis for an agent that knows the bounds of its own worldview.
In an April 8, 2026 strategy session, Scarlet Dame codified the specific triggers that govern when an agent must hand off to a human:
"Perspective-grounded agents escalate on four types: situation not covered, principle contradiction, claim tension, and ungrounded recipient input."
The four triggers make escalation structured rather than arbitrary, and the adjacent positioning frames this disciplined, perspective-grounded escalation as the product's competitive wedge: the trust contract that makes agents delegatable.
In the April 13, 2026 reachability sessions, Scarlet Dame articulated the philosophy behind delegating work to AI:
"Principled escalation (knowing limits) is the trust contract that makes delegation real."
Experts are interrupted only when the agent is genuinely out of its depth. The principle extends the goal-directed primitive: perspective plus goal plus termination criteria plus escalation criteria equals a bounded autonomous agent.
In the April 13, 2026 pilot engagement proposal, Scarlet Dame introduced the observation surface as a core component of the deployment phase:
"Pilot (Organization tier): Transparent proxy deployed on SMS/email/Slack. Internal bot + client-facing bot. Observation dashboard."
The dashboard lets the organization watch its deployed agents and verify every claim and citation its recipients saw. The model extends to consumer-facing practitioners, whose higher tiers use the same observation-and-escalation loop to step in exactly when the AI's grounding runs out.