DreamerOS

Pipeline architecture

Under The Hood

What happens between your message and the response.

Five stages

One message. Five stages. A receipt attempt.

This is the normal checked-chat path. Raw mode and specialist endpoints can take a different route. The response shows which checks ran, and the depth depends on the plan and request.

Stage 01

Pre-checks

Injection scan. Ambiguity score. Sensitive topic flag. Multi-question detection. Your message gets read before any model sees it.

Stage 02

Intent restructure

Intent recovery. Filler strip. Multi-question numbering. Context injection. Your prompt becomes precise before it goes out.

Stage 03

Engine Router

Smart routing selects from the engine families available to the plan. When independent verification runs, one lineage-separated verifier can review the result. The record names what actually served the request.

Stage 04

LLM Engine Call

The model runs. Streaming or non-streaming. Your direction set loaded as system prompt. Tier-gated model selection.

Stage 05

Output post-checks

Synchronous post-checks run on the normal path unless raw mode is selected. Deeper drift, second-opinion, and silent-drop checks can finish after the answer and join the audit record.

in: your message out: verified response

DreamerOS attempts to create a receipt after an eligible answer. When DreamerOS creates a receipt, it creates an append-only integrity record. Database triggers reject any UPDATE or DELETE. If receipt creation fails, the answer returns without receipt fields. minted records: no UPDATE. no DELETE. no rewrite of history.

Storage architecture

Three storage planes.

Account records, credentials, and hot cache have separate jobs. The public page describes those jobs; deployment region, backup schedule, and failure recovery need operational evidence, not a marketing promise.

Plane 01 / Truth

System of record

Account data and minted receipts use the DreamerOS system of record. Queries are scoped to the authenticated account where the route enforces that boundary. Public proof exposes signed fields only; signed-in accounts can see their own private context. Public verification of deployment region remains In Beta 2.0.

Plane 02 / Credentials

Credentials and backup

Credential storage and application-data storage use separate code paths. Encryption depends on the configured vault key. Backup schedule and restore proof are operational controls and remain In Beta 2.0 on this public surface until the current evidence is published.

Plane 03 / Hot Cache

Hot cache layer

The hot cache supports speed and rate-limit enforcement; it is not the system of record. On a cache miss, the normal path can reload from storage. Receipt minting is attempted separately, so a mint failure leaves the answer without receipt fields. Recovery during a real cache outage remains an operational test, not a public guarantee.

Sign up at app.dreameros.app and send your first message through the pipeline.

Start free See the journey

The difference

Same question. Wildly different experience.

See what happens when integrity checks are built into every layer.

Raw AI

"ugh so my boss wants me to figure out if we should switch to postgres or keep mysql, we have like 50k users and its getting slow"

Sure! Here are some things to consider when choosing between PostgreSQL and MySQL...

1. PostgreSQL has better JSON support
2. MySQL is more widely used
3. Consider your team's experience
4. Both are open source
5. Performance depends on use case

Generic. No context. No urgency. No decision framework.

DreamerOS

"ugh so my boss wants me to figure out if we should switch to postgres or keep mysql, we have like 50k users and its getting slow"

Quick answer

At 50K users with performance degradation, migrate to PostgreSQL -- but the Q3 deadline is the real constraint. Start with a read-replica strategy while planning the full migration.

Deep dive

Decision framework: current bottleneck analysis, migration complexity score, team readiness assessment, and phased rollout plan with rollback strategy...

Follow-up forecast

You probably also need: a load testing plan, connection pooling strategy, and a cost estimate for managed PostgreSQL hosting.

Structured. Actionable. Anticipates your next question.

What your AI is not telling you

Your AI can sound certain when it is wrong.

Not maliciously. A fluent answer can hide a weak premise, missing source, or forgotten constraint. Tone alone does not tell you which one happened.

She agrees with you too much

Ask a model "are you sure?" and it can abandon a correct answer to agree with you. An earlier draft used 86% and 98% figures that this audit could not trace to a primary study, so those numbers are not evidence. OpenAI separately documented rolling back an update it called overly supportive but disingenuous.

On Pro and Elite, integrity signals can trigger a second independent check. The verifier does not inherit the first answer's conclusion. Its result joins the audit record when it finishes; it does not silently rewrite the delivered answer.

She is most confident when she is most wrong

The AI voice can sound the same whether it opened a source or invented one. A 2026 large-scale citation audit checked 111 million references and conservatively estimated 146,932 nonexistent citations in 2025.

A confident-fabrication check can flag claims that lack grounding when that check is selected. The record says whether it ran and what it found.

She forgets everything the moment you close the window

Long context is not the same as reliable recall. The Lost in the Middle study found that model performance often drops when relevant information sits in the middle rather than at the beginning or end.

Supported DreamerOS routes can load account memory and working preferences before an engine call. Past-chat ingestion and full cross-client parity remain In Beta 2.0.

She does not know what she does not know

Models have coverage limits. Topics with plenty of training data can produce stronger answers than topics with limited training data. A confident tone does not reliably reveal which side of that line a question landed on. Sources and explicit checks give you a better signal before a downstream failure does.

When the coverage check is selected, it can ask targeted questions to surface missing context. The record says whether that check ran.

Published figures above link to their sources. Untraceable draft figures are named as unverified and are not used as evidence. See the source list.

Read the deeper reason verification exists

The pipeline

The normal chat path, in three parts.

Your input goes in. The route shows which shaping, generation, and checks actually ran before and after the answer.

You type ugh should we switch to postgres

Question shaper Layer 1 - intent locked Should we migrate from our current database to PostgreSQL? What are the tradeoffs for our stack?

5 engines Layer 2 - Build / Review / Synthesis / Signal / Research

Build Research Review Synth Signal

Integrity checks Selected before and after

! hallucination caught: confident guess ! fabrication caught: confident guess ✓ drift monitoring ✓ source verification ✓ intent match

Checked answer + receipt attempt (every plan; Light 25 a day)

Postgres is worth it if you need JSONB queries or row-level security. Here is what changes for your stack. receipt attempted

Open the engineering view of the three integrity layers

The pipeline

Three layers, when the route calls for them.

The normal checked-chat route uses shaping, generation, and visible check status. Raw mode and specialist endpoints can differ, and deeper checks can finish after delivery.

The question shaper restructures

The question-shaping layer takes your messy input and rebuilds it into a structured, domain-aware prompt. 12 enrichment layers including context loading, risk awareness, and confidence calibration.

Engine generates

Your restructured prompt routes among the engine families the plan can use. Five engines, five roles - Build, Review, Synthesis, Signal, Research. DreamerOS routes each question to the best one and checks the answer. Full detail: what DreamerOS does. The receipt names the model that actually served the request when that field was recorded.

app -> gateway | route each question to Build, Review, Synthesis, Signal or Research | best-fit routing | full spec on the What DreamerOS does page

Integrity checks verify the output

The registry contains more than 30 integrity rules. A given response runs the checks selected for its path, plan, and mode. The result shows which checks ran, and background checks join the audit record when they finish.

DreamerOS Light is free Try it now