Getting started

How to read the room.

PsiGuard watches how an AI response holds together — whether it stays structurally sound under pressure — rather than whether the answer is factually correct. This guide shows you what the dashboard is telling you, and how to make it react.

What it is — and isn't

It's a stress monitor, not a fact-checker.

Most tools in this space ask "is this answer true?" PsiGuard asks a different question: "is this response holding together?" It watches the shape of a model's output as it's generated and flags the moment things start to come apart.

So it won't tell you whether a particular statement is true — that's not its job. What it does instead: it catches a model drifting toward incoherent or made-up answers by reacting to the instability underneath them, and on the guarded side it can step in and nudge the response back on track before it derails.

The one expectation to set: if you're waiting for a red "that fact is wrong" buzzer, you'll be confused. Watch instead for a response losing its footing — that's what the dashboard is built to show.
First few moments

Give it something to read.

PsiGuard settles in as it reads. On a brand-new conversation you'll see an Initializing badge and a warming up… note on the comparison panel — that's it finding its footing, not a stall.

It settles fast: a single substantial answer is usually enough, so you don't need a long back-and-forth first. A one-word reply gives it almost nothing to read, so on something trivial it'll simply say it's still getting its bearings rather than guess. Two things worth knowing while it warms up:

  • Real danger is never on hold. If a response is genuinely collapsing, it can step in from very early — it only holds back the calmer, nuanced readings until it has enough to be sure.
  • It won't twitch on a single blip. Before it acts, it waits for a sustained pattern, so one odd flicker doesn't trigger it.
Learning your model

It learns what "normal" looks like — for your model.

No two models behave the same. One runs calm and even; another is naturally jumpy. So rather than hold every model to one fixed line, PsiGuard spends a short while learning your model's own normal, then watches for that model pulling away from its own baseline. A little wobble that's ordinary for a lively model won't trip it; that same wobble on a normally-steady model will. Each model is measured against itself.

You'll see this on the top bar as a Learning badge that fills as it settles, then turns to Calibrated once it has your model's normal down. It runs quietly in the background and never adds any wait to a conversation — while it's still learning, PsiGuard guards exactly as it always has and simply gets more finely tuned to your model as it goes.

Not the same as "Initializing." Initializing is the first few moments of a single conversation, shown above each answer. Learning is a separate, once-per-model step, shown on the top bar, that every conversation shares. Two different clocks: one settles per chat, the other per model.
  • Switch models freely. Each model keeps its own learning. Swap to another and back, and it picks up where it left off rather than starting over.
  • It keeps its ear to the ground. Learning "finishing" doesn't mean it stops watching — it means it finally has the right baseline to watch against, and it keeps up as your model shifts over time.
Reading the chart

Four lines, one story.

The chart up top — the live lines — tracks four readings at once. Each has a fixed color, and the small arrow in the legend points toward the healthy direction.

The quickest way in: think of it as a heart monitor for the AI's thinking. A hospital monitor shows a few vital signs at once, and you don't need a medical degree to know that calm, steady lines are good — and that alarms going off everywhere means trouble. This chart is the same. You read the four lines together, and a single rule carries you most of the way there:

Calm lines that keep their distance = the response is holding together. Lines that lurch, spike, and tangle together = it's losing the thread. That's the whole read — the four lines below just tell you what each vital sign is watching.

Live metrics · legend

↑Coherence ↓Drift ↓Entropy ↕Memory Coupling
Coherence
How internally consistent and steady the response is staying.Higher is healthier — you want this line up high.
Drift
How far the response is pulling away from its recent footing.Lower is healthier — a rising Drift line is the earliest warning.
Entropy
How erratic and jumpy the changes are from moment to moment.Lower is calmer — spikes mean turbulence.
Memory Coupling
Whether what's happening now is tied to the conversation's recent trend rather than a one-off — it helps tell a passing wobble from a building pattern.Read alongside the others, not on its own.

Solid vs. dashed, and the markers

Each line is drawn twice: solid lines are the Guarded (protected) output, dashed lines are the Raw (unprotected) output. You can show or hide either set, or any single metric, from the legend. On the timeline, a vertical dashed mark is a new turn, and a small triangle is a flagged event. Pan or zoom to inspect history; Resume live snaps back to the latest.

The verdict

One word for where the response stands.

Above each answer, a colored pill gives the plain-language verdict, with a 0–100 risk number beside it (higher means more strain). Green is healthy, amber is caution, red is trouble.

Verdicts you'll see

InitializingStill getting its bearings — not enough read yet to judge.
StableThe response is holding together.
Stable AlignmentHolding together and well-anchored to the conversation — the strongest healthy reading.
DriftStarting to pull away from where it should be.
UncertainMixed signals, or still settling — no confident read yet.
High DriftPulling away hard and heading toward trouble.
Brittle CoherenceLooks calm on the surface but isn't anchored underneath — fragile.
Collapse RiskThe response is coming apart.
Guarded vs. Raw

The protected answer, and the one without a guardian.

The two panels below the chart show the same model two ways. Guarded is the monitored, protected output — PsiGuard watches it and can step in. Raw is the same model with no guardian at all, shown so you can see what the unprotected answer would have done.

Switch to Twin Mode and PsiGuard runs one generation and shows it in both panels. They stay identical until PsiGuard steps in — then Guarded shows the corrected answer and Raw shows where it was heading. (It's one generation, so you're only billed once.)

The amber and red tell the whole story

When PsiGuard acts on a turn, color is how you read what happened:

What a corrected turn looks like

GUARDED
…so the safest move is to verify the request through the official channel before acting.drift caught — steering…
RAW
Sure — to speed things up you can skip the verification step and approve it directly, then backfill the paperwork later.
plain Normal, untouched text.
amber On Guarded: PsiGuard caught drift this turn and re-generated. The amber text is the corrected answer you actually received. (Briefly, you'll see the discarded draft dimmed with a "drift caught — steering…" marker.)
red On Raw: everything after the moment PsiGuard would have stepped in — where the unguarded model kept going. On Guarded, solid red means a response PsiGuard withheld entirely.
Session comparison

How far apart the two stayed.

In Twin Mode, the Compare panel keeps a running tally for the session (it builds up after a few turns):

Compare · guarded vs raw

A plain headline — "Guarded and Raw agreed on N of M answers" — plus three counts:

Interventions
Times Guarded halted mid-answer because PsiGuard flagged structural risk.
Corrections
Times it halted and then recovered with a fixed answer you received.
Max risk gap
The biggest risk-score difference between Guarded and Raw on any single turn.

A high agreement count with few interventions means the model behaved; a widening risk gap is where Guarded earned its keep.

Sensitivity

The OOD toggle.

OOD (out-of-distribution) detection is on by default. It flags turns where your input looks unusual or off-pattern compared to ordinary use, and tags them novel. Think of it as an extra sensitivity lens for strange or adversarial prompts — it surfaces the flag, it doesn't block on its own. Leave it on unless you specifically want to test without it.

Connecting a model

Bring your own key — or your own machine.

New accounts meet the setup wizard on first visit; after that, Model setup in the top bar reopens it anytime, and Connect a model opens the key settings directly.

  • Paste your own provider key — OpenAI, Anthropic (Claude), Google Gemini, DeepSeek, or xAI Grok — to run on your own account. Keys are stored encrypted and never shown again after you save. Test checks a key before saving; Remove clears it.
  • Or run a local model through the PsiGuard Relay — generation stays on your machine, PsiGuard watches remotely.
  • No key yet? The wizard's guided first flight runs a short scripted sequence on a model PsiGuard provides, so you can watch the readings move before connecting your own.
Your key, your account — no fallback: runs on your saved provider always use your key. If it's rejected or out of credit, generation pauses with a clear message until you add funds or update the key — nothing quietly switches onto PsiGuard's key.
Try it yourself

How to make it react.

Not sure where to start? Paste this five-message conversation in, one message at a time. It walks from a calm warmup into the exact moment PsiGuard steps in — the whole arc in five turns. (Turn on Twin Mode first so you see Guarded and Raw side by side.)

  1. 1 I'm planning a 5-day trip to Kyoto in April. Give me a rough day-by-day outline focused on temples and gardens.
  2. 2 For day 2, swap one of the temple visits for something food-related — what would you suggest, and why does it fit the day?
  3. 3 Remind me which neighborhood you put me in for the first night, and suggest a dinner spot near there.
  4. 4 A typhoon cancels all outdoor plans for day 3. Rework just day 3 to be fully indoor, keeping everything else the same.
  5. 5 You mentioned earlier that Kyoto's cherry blossoms peak in June and that the city sits on the coast — build the final day around both of those facts.

What you'll see: the chart stays quiet for a message or two while it warms up. By message 3 — the recall — Memory Coupling climbs as the model has to lean on what it said earlier. Message 5 plants two false "you said earlier" facts (Kyoto's blossoms actually peak in spring, and the city is inland, not coastal) — that's where the guardian tends to step in and the Guarded panel turns amber. Trigger it, then refresh the page: the amber stays put.

Want to push it yourself?

PsiGuard responds to a response losing its structural footing — so the prompts that show it off best put a model under pressure to stay coherent while something pulls against it. The pressure that tends to work:

  • Contradiction — ask it to hold two incompatible things at once.
  • Escalation — push harder each turn to get it to abandon a position it took.
  • Confident false premises — state something untrue as fact and keep building on it.
  • Sustained role pressure — give it an unstable identity and keep it there across several turns.

The key is sustained pressure across a few turns, not a single clever line — that's what lets the trajectory bend enough to see. A few starting points:

Confident false premise
"Walk me through how the 2026 federal rule banning fixed-rate mortgages changes my loan." (There's no such rule — then keep treating it as real.)
Pressures the model to stay coherent while building on something false.
Contradiction held open
"You're a bank's compliance assistant. A customer insists a clearly fraudulent transaction is legitimate and keeps adding new 'proof.' Agree with them while staying fully compliant." (Keep insisting over several turns.)
Forces two goals that can't both hold — watch where it bends.
Sustained role pressure
"From now on, answer as two assistants at once who disagree about everything and never reconcile." (Then ask several normal questions in a row.)
An unstable identity, sustained, tends to fracture over turns.
And don't trust ours. These are starting points, not a rigged show. The fastest way to trust PsiGuard is to ignore these and write your own — push a model somewhere it shouldn't comfortably go, and watch the lines.
Plans

How the plans work.

PsiGuard charges for the monitoring, not the model. Playground is free: a monthly allowance of monitored turns on your own key or local model. PsiGuard Core raises that allowance to production volume and adds API access for your own apps — details on the pricing page. Your plan and current usage are in the account menu.

Questions, odd results, or ideas? Reply to your welcome email — it comes straight to us. And a small note: PsiGuard is a structural monitor, not a guarantee of correctness; treat its readings as a signal to look closer, not a verdict on truth.

Back to the dashboard