Trust should be inspectable.
An assessment like Stronghold asks people to be honest about things they may never have said out loud. That honesty deserves more than a privacy policy written by lawyers and a promise to be careful. It deserves running systems — oversight, hard rules, and records that cannot be quietly rewritten.
What follows is the public version of The Stronghold Standard, the internal protocol every Stronghold product is built and operated under. Each item describes something operating on the platform today.
Gideon watches the platform — including our own AI.
Gideon is Stronghold's independent oversight system. It runs continuously, and its only job is to observe and report to a human. It never modifies data, never messages a client, and never takes an action on its own.
System health
Failed emails, broken syncs, stalled report pipelines — caught across every practitioner, usually before anyone notices.
AI conversation quality
Samples the coaching assistant's real conversations for fabricated claims, out-of-scope clinical framing, or drift from its rules.
Scoring integrity
Any score built from too little data or low confidence is flagged before it ever reaches a practitioner.
Assessment experience
Where people slow down or drop off, so friction gets fixed instead of costing someone their moment of clarity.
And the watchdog has a watchdog. A second, deliberately simple system — independent of Gideon and immune to the failures that could silence it — checks that Gideon is actually reporting. If oversight ever goes quiet, that silence itself triggers an alert.
Hard rules for our AI. Not suggestions.
Every AI assistant in the Stronghold family operates under the same non-negotiable rules — written into its operating instructions and checked by Gideon against real production conversations:
- No fabricated predictions. The assistant never states a made-up number as if it were measured. If the platform didn't measure it, the AI doesn't claim it.
- No diagnosis, ever. Patterns are described in plain language. A self-report assessment is not a diagnosis, and our AI will not pretend otherwise.
- Context before analysis. The assistant asks whose situation it's looking at before analyzing anyone — no instant psychological profiles of third parties.
- Hard safety escalation. Suicidal ideation, abuse, or severe crisis content routes to a licensed provider or 988 — every time, without exception, regardless of what else is being discussed.
- Coaching, not treatment. The AI supports a certified practitioner's judgment. It does not replace it.
What never leaves the platform.
- Access is restricted at the database level. A practitioner can only ever query their own clients' data — enforced in the database itself, not just hidden in the interface.
- Sensitive content never travels by email. Automated reports and alerts are stripped of client-identifying detail before anything is sent. Full records live only inside the authenticated, access-controlled system.
- Confidential disclosures stay confidential. What a client or partner shares privately is flagged practitioner-eyes-only and structurally excluded from anything joint or client-facing.
- Calendar privacy by default. Scheduling shows a first name and last initial — never a full name, never notes.
- Consent is logged, not assumed. Assessment consent and disclosure acknowledgments are recorded with a timestamp.
Detect, fix, verify — and prove it.
- Detect. Gideon's checks usually catch an issue before a client or practitioner reports it.
- Alert. Critical issues notify a human instantly; routine findings roll into a daily review.
- Log. Every issue is written to a permanent record — nothing lives only in an inbox.
- Resolve. A human fixes it and records what changed.
- Verify. If a "fixed" issue recurs, its record automatically reopens. A fix isn't done until it stops recurring.
And the history can't be quietly rewritten. Every status change is copied to a separate, append-only ledger where each entry is cryptographically chained to the one before it — altering any past entry breaks the chain and is immediately detectable. We tested this the honest way: we deliberately altered a historical record to try to fool our own system. The verification check caught it on the next run and pointed to the exact entry.
Specialized tools are earned, not sold.
Every client starts with the core Stronghold assessment. The deeper modules — couples work, business owners, and beyond — require the practitioner to hold the relevant certification before the system will let them administer it. That's enforced server-side, not politely requested.
And assessment results are never shipped to a stranger with no practitioner attached. Stronghold's wellbeing check-ins can surface things that deserve a trained human's attention — so the platform is built on the assumption that a human is watching. Quality control lives with certification, not with who has a credit card.
We publish our own corrections.
The Stronghold Standard is a living document with a public spine: when outside review finds a gap, we verify it against the running system, fix it, and record the correction — dated, specific, and in the open. Recent revisions came directly from external scrutiny: alert handling was tightened, and a sanitization layer was added after a review caught a raw identifier in an outbound alert. Both were verified live the same day.
Get checked. Verify against reality. Fix it. Write it down. That loop is the standard.