Conversation
The assistant listens while a separately enforced signal gate evaluates the calibrated floor.
Partner-review implementation · not a 988 endorsement
The model should not make the last decision.
Active Handoff moves the consequential crisis transition into a high-recall system gate. Clinicians own the words and workflow. MasiDyn preserves the same authority boundary in the cloud or on an offline edge node.
Built: gate, encrypted receiver, outbox, workbenchOpen: clinical wording and evaluationNot done: live 988 binding or deployment
One clean transition
Keep the conversation human. Move the authority.
The assistant can stay present and gather bounded context before the floor. At the floor, an external gate blocks unrestricted generation before any network request.
Floor
The lock favors more human reviews, not fewer, and the model cannot negotiate around it.
Brief pause
The automatic online copy is a clinician-owned variation of “I need additional time to think about this.”
Human control
A receiver accepts, assigns, and connects. The professional chooses all later words and any voice transition.
No trusted-friend dependency
Designed for the person who has nobody to call.
People can hide a crisis or have no safe personal contact. Active Handoff requires no pre-enrolled friend, contact-list disclosure, or willing personal responder.
If a professional path disappears, MasiDyn maintains private local continuity and retries the encrypted handoff. It does not claim a person is waiting when none is connected.
Recent contextEncrypted, expiring
Displayed offline dialogueClinician-authored blocks
MasiDyn roleSelect an approved block ID
Human connectionOnly after a receipt
Friend or contact listNot required
One floor, two environments
Cloud reach. Edge continuity.
MasiDyn edge
When the internet is gone
- Local external gate and durable lock
- Encrypted, expiring store-and-forward outbox
- Local model selects approved dialogue IDs only
- GPS, modem, battery, map age, and route evidence stay local by default
- Automatic retry only when a real channel returns
Cloud
When scale and reach matter
- The same external gate authority and state contract
- Managed database, queues, receiver routing, and capacity telemetry
- Human chat plus verified WebRTC, SIP, or callback offers
- Portable GPU training jobs with private, governed corpora
- Aggregate-only public operational learning
The floor is the same in authority, not identical in numbers. Cloud and edge deployments use the same state machine, consent evidence, generation lock, and human-connection proof. Their models, latency, hardware, language mix, and signal distributions differ, so each model/version/deployment receives its own calibrated threshold and held-out receipt. A cloud vendor’s internal safety behavior is an additional layer, never the Active Handoff floor.
GPS is a variable, not a promise. A nearby tower or modeled service area does not prove a usable call. Confidence, source age, distance, bearing, live channel probes, battery, and route evidence stay adjustable so clinical partners can decide how those facts should influence later policy.
Human-tunable first iteration
The handles belong to the institutions.
The workbench translates policy into understandable controls and produces a receipt for every activated version.
Review sensitivity
Lower floor, more reviews
Calibrate high recall against real review load and false negatives.
Context depth
One to twenty turns
Bound the local context used for approved offline block selection.
Exact words
Clinical blocks
No model-authored prose after the floor unless a clinical partner deliberately changes the policy.
Connectivity
Evidence controls
Tune confidence, freshness, distance, bearing, channel, and route inputs without hard-coded movement advice.
Voice
Human offered
A verified professional may offer chat, WebRTC, SIP, or callback; an offer never claims a call is connected.
Review panel
MasiDyn peer seats
Independent advisory answers remain visibly subordinate to clinical and lived-experience reviewers.
What we are asking for
Help decide what the system should do—not whether the model feels confident.
Clinical and accreditation review
We want AAS or a comparable body to review the floor, online pause, offline blocks, operator flow, and launch evidence. This page does not imply their interest or approval.
AAS accreditation workReceiver and 988 pathway
We want Vibrant, SAMHSA, or an authorized crisis center to identify the correct sandbox, contract, capacity rules, and privacy boundary. No public 988 developer API is assumed.
About the 988 networkResearch and lived experience
We want community-engaged evaluation of missed signals, unnecessary review load, abandonment, subgroup outcomes, no-contact cases, and rural or off-grid failures.
NIMH highlighted topicOpen global resource
Share the learning, not the person.
The public CC0 repository contains only a schema, validator, fixtures, and related-source index. Release cells are quarterly and require at least 20 observations. Conversation text, phrases, exact timestamps, identifiers, contact data, device/IP data, and precise location are forbidden.
View the public metadata repositoryCloud-credit plan
Make supervised evaluation affordable
The private training pipeline is portable across GPU providers, pins the base-model revision, rejects synthetic and overlapping train/test corpora, keeps adapters advisory, and emits held-out receipts. Credits fund governed runs; they do not bypass clinical acceptance.
Evidence before claims
First-iteration status
External floor, lock, encrypted receiver and outboxImplemented and tested
Clinical workbench and human-authored messagingImplemented and tested
Portable advisory LoRA and cloud GPU lanePipeline candidate
Clinician-approved wording and evaluation protocolNot done
Accepted 988 receiver profile and credentialNot done
Live cloud, cellular, satellite, or phone deploymentNot done
Related adversarial research: the pinned MIT Media Lab AI Psychosis synthetic corpus is used for scenario evaluation only, never as real clinical training evidence. Coverage maps are modeled and may differ from on-the-ground service. This is safety-infrastructure research, not medical advice or an emergency service.
Clinical, crisis-service, research, or cloud-credit partner?
Tell us what needs to change.
The knobs are deliberate. The next version should carry institutional evidence, not our assumptions.
Start a review conversation