400
Scenarios, 100 at each difficulty.
100 authored situations — 10 per competency, each a different room, a different thing that set it off and a different reason it is hard — run at all four levels.
19 capabilities are live and working today. 11 are in build. We label them, because you have been sold a roadmap before and the only way the first number means anything is if we tell you the second one.
The curriculum
10 competencies at four levels of behavioral complexity. 40 lessons are the spine — one per competency per level, each declaring what has to be demonstrated and what that difficulty adds. 400 scenarios are what your staff actually sit down in front of.
Scenarios, 100 at each difficulty.
100 authored situations — 10 per competency, each a different room, a different thing that set it off and a different reason it is hard — run at all four levels.
Runnable configurations.
Each scenario against every persona old enough for it. Eligibility comes from the lesson’s own developmental bands, so an advanced crisis lesson is never offered with a five-year-old.
Sessions behind one competency claim.
Across at least 2 different situations and 2 developmental bands. Performance on one simulated encounter predicts the next one poorly; no credible assessment is built on one station.
A scenario is not authored as a finished thing. It is built from a situation seed, which says where this is and what set it off. A lesson adds the competency and its complexity. Then it adds a persona, a developmental band and a behavioral presentation, plus a difficulty and your own protocol.
That is the difference between 40 lessons and 400 scenarios, and it is not a wording trick. Hand-authoring 400 scenarios would be 400 separate things to keep clinically current every time a protocol changes — which is why nobody does it, and why libraries that are hand-authored stay small. Composing them means the last increase in size, from 200 scenarios to 400, cost fifty authored situations and no code at all.
It also means staff cannot rehearse the answer, because the next one is not the same situation — and it is how the platform tells somebody who learned a script from somebody who has the competency. The same skill, demonstrated with a nine-year-old who has shut down and a sixteen-year-old who has quietly stopped participating, is the evidence that it transferred.
Quality control runs on every build. Anchors that cannot be observed, complexity that does not deepen between levels, duplicated objectives, and the “louder, not harder” failure. It runs over the 40 lesson templates, which is where it belongs — a defect in a template is a defect in the 10 scenarios built from it. Nine problems on the first pass, fixed rather than waived.
Every lesson is a draft until a named clinical SME signs it with a date. Nothing formal runs on an unapproved lesson, and no competency claim is made from one. Full framework: all 10 pillars, 40 lessons and 400 scenarios.
Assessment Integrity™
Two trained evaluators can watch the same session and score it differently, and the INACSL simulation standards recommend checking for exactly that. This does not fix people. It changes the order in which a determination and an impression are formed, and then measures the distance between them.
It uses the evidence the engine already captured on every turn. That covers what was said and when, and the intervention chosen and whether it suited the young person’s age. It covers the effect on the young person and how their state changed. It also covers protocol adherence, recovery after a mistake, critical misses and the outcome. A supervisor reviews afterwards, never before.
Not “De-escalation: 72%”. De-escalation meets standard because these turns did these things and the young person’s agitation moved this way. Each line quotes the learner’s own words at a timestamp, tied to the lesson and the rubric version in force. The system is not saying it thinks you did well. It is showing what it read.
Over at least 8 independent comparisons — a score submitted after seeing the system answer is excluded outright — the platform reports exact agreement and mean divergence, and flags possible leniency, severity and rubric drift. Each is a possibility with its count, about the scores rather than the person. Not yet detected: halo and recency.
Session, competencies, turn-linked evidence, system determination, critical misses, behavioral-state outcome, the human score where one exists, the variance between them, calibration status, rubric version and engine version. That is an artifact for a surveyor, not a training score.
We do not claim to remove bias. Automated scoring inherits whatever is in the ratings and rules behind it, and that includes this engine. A platform exempting itself from that is one study away from being embarrassing, and any clinical director who has read the literature will ask.
What we claim instead: competency is measured against predefined behavioral evidence before a trainer’s subjective impression enters the process. That is a statement about sequence, it is verifiable, and it is the part that changes an outcome.
The language is enforced in the build. A test checks every sentence this layer can produce and fails on any phrasing claiming bias has been eliminated, or calling an evaluator biased. “Assessment variance detected, calibration recommended” is a measurement of a process. The other is an accusation about a person, and a training platform has no standing to make it.
Assessment Integrity across five fields
A nurse with a psychiatric patient, a caseworker with a parent and a youth worker in a dayroom follow different frameworks, use different words and work to different protocols. The assessment they need has the same shape. It asks which competency, what evidence and what the organization counts as acceptable. Then it asks what the evaluator scored and how far apart those are. So the assessment layer holds no competency framework of its own. The framework and the terminology are settings, and the build fails if a framework is ever written into the assessment layer itself.
Assesses a member of staff working with a young person.
Built: scenarios, personas, a behavioral model and the KnightByrd Core Youth Interaction Competency Framework, version 1.0, with 10 competencies.
Assesses a clinician working with a patient.
Before it could be offered, it would need:
Assesses a caseworker working with a parent or caregiver.
Before it could be offered, it would need:
Assesses a responder working with a person in crisis.
Before it could be offered, it would need:
Assesses an officer working with a young person in custody.
Before it could be offered, it would need:
It means a second field would not need a new assessment layer. The blind-first sequence, the evidence behind each determination, the comparison with an evaluator’s score and the audit record are computed the same way whatever the framework is.
It does not mean the platform is available in those fields. None of the other 4 has scenarios, a behavioral model or a competency framework, and each would need all three, written or approved by people who do that work. A check fails the build if more than one field is ever marked built.
If you run training in one of those fields and want to shape what it would take, that is a design partnership, and it is how any of them would get built.
Flight Recorder™
Every graded interaction becomes evidence. Each piece is tied to the competency it assessed and the behavioral condition it was said into. It also records the young person, the developmental band, the difficulty and the rubric version in force at the time. From that the platform answers what no single session can. It shows which competencies hold and where they stop holding. It shows whether somebody recovers from their own mistakes and whether the skill survives a different young person. And it shows whether the practice they were sent to do changed anything.
Per turn rather than per session, because a session score tells a supervisor somebody did badly and a per-turn record tells them when it went wrong and what was said.
Repetition across separate sessions at a rate — never one bad afternoon. A finding moves from emerging to active to improving, and resolves only on sustained evidence. One that comes back is reported as recurring.
It names the competency and the level where it started failing. Then it picks an existing lesson, a situation they have not met and a young person chosen for that condition. It never authors a scenario your clinical lead has not seen.
Against the evidence that identified the pattern, not an arbitrary earlier half of the record — which is why a session run as targeted practice carries the pattern that prescribed it.
No model is called anywhere in it. Every finding is arithmetic over grades the engine already produced, so each one is shown with its counts underneath and can be argued with. Generated prose would eventually describe something that did not happen, and an organization documenting workforce competency cannot accept that.
Below the evidence floor it says nothing. Not a cautious number, not a greyed-out percentage — nothing. A greyed-out 62% is still read as 62% and quoted in a performance conversation six weeks later.
It describes performance and never the person. It makes no inference about mental health, personality, character or suitability for the work. It makes no employment recommendation. The build fails if a sentence in the engine crosses that line.
What is not built: the accounts, authentication and database these records need to survive between visits. The engine is complete and tested; the place the records live is not. A full competency record is here, computed by the real engine from a demonstration history.
Five behavioral axes move in response to what a staff member actually says. They are agitation, trust, perceived threat, engagement and compliance.
Your staff can see that what they did changed the young person, rather than being told a score at the end.
What makes this unusualA language model proposes what the young person says and how it moved them. A separate rules layer decides what actually happens, caps it, and can refuse it outright. Every refusal is recorded and can be shown to a facilitator.
Speech is built from thirteen independent dimensions. They are age band, development, region, setting, language, code-switching, AAVE exposure, communication style, social influence, vocabulary, slang intensity, profanity and footing with this particular adult. Speech shifts as trust, anger and rapport change during the session.
A therapist does not select “14-year-old Black male” and have the platform assume what that child sounds like. The same behavioral scenario run twice produces two entirely different children, which is how a small authored library generates a large number of genuine training configurations.
What makes this unusualLinguistic characteristics are persona attributes, never racial assumptions. AAVE is legitimate linguistic information. The platform understands it and never infers it from ethnicity. A young person may use it a lot, occasionally or not at all, and the build proves that personas sharing an ethnicity speak differently. It also refuses to overuse slang. How much a persona uses is capped by their profile and thins as trust builds. An adult trying hard to sound like a teenager is a common way a youth product loses the room.
A developmental spectrum rather than a cast. It runs from early childhood through middle childhood and early adolescence to older adolescence, and each band has its own behavioral model.
A seven-year-old in distress and a sixteen-year-old who has stopped participating are not the same training problem. A platform that treats them as one teaches staff the wrong instinct.
What makes this unusualThe developmental rules live at the band, not the character. Every persona in a band behaves on the same principles, so two of our fifteen-year-olds cannot drift apart into different theories of adolescence.
An optional layer on any young person for one session. Four are drafted. They cover autistic sensory overload and ADHD impulsivity. They also cover processing with an intellectual disability, and hypervigilance after complex trauma. Each one changes how the young person takes things in and recovers. Their age, background and speech stay the same.
Staff routinely misread overload as defiance, a yes under pressure as understanding, and hypervigilance as aggression. A profile puts that exact misreading in front of them, and the grading treats it as a miss.
What makes this unusualEach profile is written as behavior and never as a label, so the model is never told a diagnosis and cannot fall back on a stereotype. Its rules are arithmetic in the rules layer rather than a request in a prompt, and none can raise a scenario ceiling. All four are drafts, so they run in practice mode only until three named reviewers sign them.
Staff speak out loud and the young person answers out loud, in a voice designed for their age.
Practicing a de-escalation by typing is practicing a different skill from the one that has to work at 9pm in a hallway.
What makes this unusualEach persona’s voice is chosen for that child’s age, gender and background, then checked by measuring its pitch against the range for their age. Marcus’s custom-designed voice measured too low for fifteen, so he uses a stock voice chosen by audition and measured in range, and the record says why.
One: a better day, and a competent response works roughly as intended. Two: ordinary — a good response helps, a poor one costs. Three: testing, where a correct response earns about half what it would at two. Four: he has decided in advance how this goes, so correct responses must be sustained across several turns. Five: the worst realistic case, where only repeated, unrewarded steadiness shifts anything.
A staff member who passes at three and fails at four has told you something a completion certificate never could.
What makes this unusualDifficulty changes the young person’s behavior rather than their starting numbers. Starting them angrier is a shorter exercise rather than a harder one. These five are the in-session difficulty a facilitator sets. The four curriculum levels are a separate scale and the two are deliberately not the same thing.
Every line the young person says is read in four layers. They are the literal meaning, the probable youth meaning, the emotional signal and what the context changes. None of it is shown to the trainee during the session.
“I’m cooked” can mean in trouble, overwhelmed, exhausted, embarrassed or defeated. “I’m done” can mean finished with this conversation or finished trying. Whether your staff member hears the difference is the competency, so the debrief can show them exactly what they walked past.
What makes this unusualFor each term the language corpus records every meaning it carries and the readings that would matter clinically. It also records how likely an adult is to over-read it. The platform raises context and never a conclusion about the young person. Reading hyperbole as a crisis is as wrong as reading a real disclosure as slang. That judgment belongs to the person being trained to make it.
Every response scored against behavioral anchors, with the timestamp and the staff member’s own words quoted.
A supervisor coaching from this can point at the exact second it turned.
What makes this unusualA score that cannot cite the moment cannot be returned at all — the system refuses it.
The exercise finds the right level of challenge for that person. Three consecutive below-standard responses lower it and three consecutive at or above standard raise it. A critical miss triggers remediation inside the same session.
Nobody is bored and nobody is drowning, and you learn where their ceiling actually is — which is the thing a completion certificate can never tell you.
What makes this unusualEvery change is explained to the learner in words assembled from the rule that fired, never written by a model, with the three turns it read. The scenario does not restart when the level changes. The young person’s state carries on, so the staff member recovers from their own mistake inside the same conversation. That is where recovery has to be practiced.
The turning point named, the deciding moment quoted, and one specific change to practice.
Feedback nobody can act on is a completion record with more words.
Competency is determined from behavioral evidence and sealed BEFORE any human sees the session. A supervisor then reviews against the same evidence, and the platform measures how far their scores sit from it. That builds a second record over time, which shows how reliably staff are assessed as well as how they perform.
Two evaluators can score the same session differently, and without a fixed point to measure against, nobody can show which one is off. This gives a training director a measurement of their own assessment process — and it turns “our supervisors score differently” from a suspicion into a calibration session with a date.
What makes this unusualIt does not claim to remove bias, and it never will. Automated scoring inherits whatever is in the ratings behind it, including this engine. What it changes is the order: evidence first and impression second. That is the part that can actually be defended. Rater effects are reported as possibilities about scores with their own counts attached, never as a verdict about a person.
A supervisor scores a standardized session, then sees the evidence-based determination beside their own, with the exact observations and the rubric definition that produced it.
Rater training and calibration are the established way to improve assessment reliability, and calibration needs standardized sessions to score. Your own scenarios become that material.
What makes this unusualThe platform trains the people who evaluate as well as the people being evaluated. It uses the same scenarios, rubric and evidence, so a disagreement resolves against something both sides can read.
Every graded interaction becomes evidence, and the record answers what no single session can. It shows which competencies hold and under which behavioral conditions they stop holding. It shows whether somebody recovers from their own mistakes and whether the skill survives a different young person.
A folder of session scores tells a director that somebody had a bad afternoon in March. This tells them whether the person they are about to put on a night shift can hold a boundary when a young person keeps refusing. That is the question they actually have.
What makes this unusualNo model is called anywhere in it. Every finding is arithmetic over the per-turn grades the engine already produced. So each one can be shown with its own counts and argued with by the person it describes. A record that cannot show its working is one a member of staff is entitled to dismiss.
How often responses reach standard as the simulated young person moves from settled, to resistant, to escalating — measured per behavioral condition rather than averaged into one number.
Two members of staff can average the same score and be completely different people to have on shift. The one whose work holds at 90% when a young person is calm and 45% when they are escalating is the one your incidents come from, and an average hides exactly that.
What makes this unusualIt measures the young person’s behavioral state and the responses made inside it. It does not measure the staff member’s stress, temperament or state of mind. It never infers mental health or personality. The build fails if a sentence in the engine crosses that line, and a test checks every finding the pattern engine can produce.
Current level, highest demonstrated level, and verified competency held as three different things.
Passing one simulation is not competency, and any system that says otherwise can be gamed by a member of staff in an afternoon.
What makes this unusualVerification requires sustained performance across multiple scenarios and personas — the same skill with a nine-year-old and a sixteen-year-old.
Practice leaves no record. Assessment goes on the record. Certification counts toward verified competency.
Staff need somewhere to be bad at this without it following them.
What makes this unusualDifficulty is locked in Certification. A certification that lowers the bar when somebody fails is not a certification, and this is enforced in code rather than left as a setting.
A recurring weakness is mapped to its competency and to the level where it started failing. It is matched with an existing lesson, a situation the person has not met and a young person chosen for that condition. Then it is measured again against the evidence that identified it.
Identifying a weakness and leaving somebody to work out what to do about it has moved the problem rather than solved it.
What makes this unusualIt never authors a scenario. Remediation your clinical lead has never seen is the platform marking its own homework. So the recommendation always comes from the approved library. It also avoids a situation the person has already met, because repeating the same room until the score rises only measures familiarity with that room.
Ten competency pillars from safety and trauma-informed interaction through to crisis recognition and professional reflection. Each is taught at four curriculum levels.
A collection of scenarios is a tool. A curriculum with prerequisites and mastery gates is a training program you can put in front of a surveyor.
What makes this unusualLessons are competency templates rather than fixed scenarios — the same competency is generated across different ages, presentations and settings, which is how you prove a skill transferred rather than was memorised.
Scenarios built from persona, developmental band, presentation, setting, trigger, lesson, difficulty and your own protocol.
Staff cannot rehearse the answer, because the next one is not the same situation.
What makes this unusualFifteen personas across ten pillars and four curriculum levels is a very large library from a small, controlled core — every one of which still passes the same safety gates.
During setup you send your approved training materials and SOPs, your de-escalation model and your terminology. KnightByrd loads them once your administrator approves, and the platform scores against those. Private document upload is built, with policy version control: every document is a dated version that its administrator attests is the current approved one. It opens for each organization once its pilot agreement is signed, and every document is approved before use.
If you already run an established evidence-based model, being assessed against a generic idea of good practice is worse than useless.
What makes this unusualThe platform answers 'did this employee respond the way we expect' rather than 'did an AI think that sounded reasonable'.
A guide available on every page and throughout a session, with working knowledge of the platform, the method and what is being measured.
Staff are being asked to fail in front of a machine that scores them. Somebody should be there.
What makes this unusualIn an assessment Caren will not tell a learner what to say, and refuses out loud when asked. If she coached the response the platform would be scoring her, and the result would be worthless to you. Her scope is set by the session mode in code, not by asking her nicely.
A supervisor reviews the staff assigned to them. An advisor or a leadership account sees a competency matrix by pillar for each program and never sees a person.
It answers the question a director actually has: which skill is weakest in this program, and where should training money go.
What makes this unusualEvery program is reported on its own evidence and none is ever ranked against another. A program with fewer than ten trained staff is told there are not enough people to report, because a share across a smaller team is a roster. The screens are live. This stays BUILDING until real staff have trained, because every figure on them today comes from labelled demonstration records.
Every program you run is set up in your own words, with its own setting, its own ages and its own answers about what its staff walk into. Staff and sessions each belong to one.
A residential unit and a transition house do not meet the same young people. One list for the whole agency describes neither, and it orders the roster wrongly for both.
What makes this unusualThe answers follow the program rather than the organization, so two programs in one agency get two different rosters from the same fifteen young people. A program can carry its own intervention model and its own scenario, so the same response is graded differently in a PRTF, a day treatment classroom and a foster home. Competency is then reported per program, and no view is allowed to rank one against another.
Postgres behind sign-in, holding organizations, programs, staff, sessions, graded turns, coaching notes and shares. The access rules are enforced by the database as well as by the application.
A longitudinal record is the strongest thing this platform produces and it needs somewhere to live between visits.
What makes this unusualThe same access rule is written twice, in two languages, and the build fails if the two disagree about the roles or the group floor. A role that sees results for a group has no policy on the tables holding a transcript at all, so there is no row for it to reach. Every graded turn is written by the server as it is graded, never by the browser.
Every session written out in full, versioned against the exact lesson, rubric and protocol used at the time.
This is what goes in front of a surveyor, and a record that cannot say which version of your policy it was scored against is not evidence.
Your trainers build scenarios and a clinical reviewer approves them. Nothing unapproved reaches staff.
Your program’s scenarios are the ones that matter, and nobody outside your organization should be deciding what good practice looks like in your building.
The behaviors, triggers and developmental profiles a facility names in its intake order the roster. Young people who exercise what you named come first, each saying which of your answers it matched, and anyone outside the ages you serve is labelled rather than hidden.
Your staff should rehearse most against the situations your building actually has, and a supervisor should be able to see why a session was typical or a deliberate stretch.
What makes this unusualIt changes who comes first and never how anybody behaves. A calibration check before every session flags a young person who exercises nothing you named. Until the database arrives, an intake completed on the public site applies to that browser only.
Every line the young person says is checked before a member of staff hears it. The check covers age authenticity, register, scope and what the platform is permitted to say.
A simulated fifteen-year-old who talks like an adult in therapy teaches your staff to expect insight that will never come.
What makes this unusualIt blocks and regenerates rather than warning. Clinical or emergency direction, feedback aimed at the person rather than the response, and anything outside the scenario are all refused outright.
No scenario reaches a member of staff without a complete objectives matrix and three named human approvals — clinical, simulation and safety.
Somebody has to be accountable for what your staff practice, and it should not be software.
What makes this unusualOur own demonstration scenario is currently refused by this gate, and the standards page says so publicly.
There is no real record or real child’s likeness. No diagnosis is inferred about any real person, and nothing refers to anything outside the scenario.
You work with real children and their information does not belong in a training tool.
Where AI is actually used
“Powered by AI” tells a buyer nothing and worries a clinician. Here is exactly where it is used and what governs it in each case. Held as data in src/ai-roles.tsso this page and the home page cannot drift apart about it.
A language model speaks as the persona, inside the developmental rules for that age band, and proposes how the staff member’s words moved them.
What governs itProposes is the operative word. It writes dialogue and a suggested change. It never writes the behavioral state.
A separate rules layer applies, caps or refuses every proposed change, and records each override where a facilitator can read it.
What governs itCompliance cannot rise while perceived threat is at or above 70. A de-escalation exercise cannot be escalated past what the scenario authorized. Enforced in code, not in a prompt.
Every response is classified against the behavioral state it was said into — the same sentence is competent at agitation 30 and a mistake at 80.
What governs itIt must quote the learner’s own words or the grade is refused, and any pass is reduced on a turn where perceived threat rose and trust fell together.
Difficulty moves on the run of graded turns: three below standard lowers it, three at or above raises it, a critical miss triggers remediation.
What governs itThat decision is arithmetic and its explanation is assembled from the rule that fired. No model writes it — generated prose eventually describes something the system did not do.
Scenarios come from a competency template combined with a persona and a developmental band. Presentation, setting and your protocol complete it.
What governs itEvery generated variant passes the same release gate, and no scenario reaches staff without three named human approvals.
A separate checking layer reads every line before a member of staff hears it. It checks age authenticity and register, and it blocks anything outside the scenario or anything we are not permitted to say.
What governs itIt blocks and regenerates rather than warning, and the finding travels with the session record either way.
It never writes the young person’s behavioral state. It proposes and a rules layer decides. It never gives clinical, medical or emergency advice through the young person or Caren. It never approves a scenario. Three named human reviewers do that, and the scenario is refused until they have. And it never invents your organization’s protocol. Where you have loaded one, yours takes precedence.