Good evening, there.
Your AI learns how you work — transparently, privately, and under your control.
Recent chats
Chats & Projects
Resume work with the correct project scope, profile overlay, and learning permissions.
DCS Engineering OS
Control Tower, Adaptive Workspace, Connector OS, Mission OS, and R-Series integration.
AIE Workspace
Final UI design contract, profile controls, memory, intent, receipts, and developer handoff.
TRD Mind & Mart
Private AI workspace and open AI marketplace connected through TRD Console.
Recent conversations
| Conversation | Project | Scope | Learning | Updated |
|---|---|---|---|---|
| No conversations yet — start a chat and it will appear here. | ||||
My AI Profile
Everything the system has learned is reviewable, editable, exportable, and receipted.
Applied
Shapes every answer · ≥ 0.90Suggested
AIE thinks this · NOT acting on it · 0.70–0.89AIE inferred these but has not confirmed them with you — so it shows them and refuses to use them until you say yes. Every other memory silently acts on its guesses; this one does not.
Communication
AppliedWork Intelligence
Read overlayExecution Rules
User Controls
Learning Timeline
Point-in-time replay over the existing Time Machine receipt and version chain.
Events
—Loading your learning timeline…
Point-in-time profile
—Loading…
Memory & Knowledge
Bi-temporal facts, episode summaries, entities, and prevention rules. Raw conversation turns are never stored.
Active facts
—Loading what AIE has actually stored about you…
Corrections & Prevention
Mistake → correction → reason → prevention rule → measurable accuracy improvement.
Active prevention rules
—Loading your prevention rules…
Feedback fusion
Weights not shownThis card showed four fusion weights — 1.00 / 0.99 / 0.22 / 0.18. The last two exist nowhere in config.ts or adaptation.ts. They were invented and presented as the engine’s parameters. Fusion is real (corrections are weighted by recurrence and recency, reversals demoted below the floor) — but the numbers are in the code, and a screen that makes them up teaches you to trust a value nobody set.
Correction history
Your correction rules render above, from GET /aie/corrections. This box used to show an invented example rule with a receipt id that does not exist.
Correction: use the approved TRD asset automatically for every TRD project surface.
scope=project:trd · asset=trd_logo_locked · replacement=blockedIntent Lab
Test discussion vs instruction, quoted examples, abstention, and correction-derived prevention rules.
Classifier
No execution. Quoted/example span excluded by Boundary B5.
Execution thresholds
Config · DECIDE receiptsAdversarial restraint suite
Not asserted hereThis card showed three green PASS badges and “IntentEnginePort conformance green”. Nothing on this page ran a test. The adversarial suite is tests/adversarial.test.ts and it runs in CI — a test result asserted by HTML is not a test result, it is a claim. Use the box above to classify a real message and see the real confidence and abstention.
Transparency
See what was learned, why it was learned, the evidence used, and every change in its validity window.
Why this preference?
—Loading your highest-confidence belief…
Confidence explanation
—Click a confidence badge on any fact to see why — occurrences, confirmations, contradictions, recency, and the receipts behind it. Every value is fetched from /aie/explain; nothing here is hardcoded.
Learning boundaries
Zero-leak gateSession facts expire. Durable promotion requires recurrence or explicit pin.
Regex + model redaction before persistence; content never logged.
Single occurrence remains evidence-only.
Learn pipeline aborts at step 0.
Quoted, pasted, and example spans are non-executable.
Assistant turns never become user beliefs or opinions.
Receipt Explorer
Every belief AIE holds was written with a receipt. Verify any of them — the check is cryptographic, and AIE cannot forge one or quietly change a fact without breaking it.
Your receipts
—Loading your receipts…
Verification
—Select a receipt, or paste an id, and press Verify. Nothing here is generated. What you see is what the service returned.
Privacy & Consent
Consent is enforced at the pipeline mouth—not only in the UI.
Learning controls
—Structural guarantees
Sensitive categories
Keyword filter · NN7AIE drops candidates whose predicate, value or source text matches a sensitive keyword — in English, Hindi and Hinglish — before anything is written. takes_medication, salary and passport_number are all refused.
It is a filter, not a guarantee. A predicate that names no keyword — blood_type, bank_account — can still pass it. Switching a category on below adds a second, explicit refusal, and is recorded in your consent record with a receipt. Widening the filter itself is tracked as CW23.
Connectors & Warm Start
Warm-start imports are consent-gated, boundary-filtered, and receipted.
Import rules
P5 · DARKRecommendations
Built from your confirmed preferences only — nothing below the 0.90 act gate. Every card names the belief it came from and its receipt. Accepting one executes nothing: it records your approval. AIE was already applying the underlying preference — that is what confirmed means.
For you
—Loading your recommendations…
What this tab will never do
Enforced in codeHealth & Benchmarks
What the service actually reports, probed live. Not what a designer typed.
Feature flags
—Loading…
Capabilities
—Loading…
Evaluation gates
Not asserted hereThis card used to show seven green PASS badges. Nothing on this page ran a test. Grades come from npm run golden against the live model, and they live in eval/out/ — a test result asserted by HTML is not a test result. The latest run is in the repo, not in this tab.
How AIE Works
AIE learns what you prefer, shows you everything it believes, and can prove where each belief came from. This page explains exactly how — including what it will not do.
The one-line version
You talk to AIE normally. When you state a standing rule about yourself — "always reply in English", "never use emojis", "confirm before deploying" — it remembers. Every other product asks you to fill in a settings page. AIE learns it from the conversation, then shows you the receipt.
It does not remember your questions, your one-off requests, or anything you did not say about yourself.
What it will remember
Durableprefers_language.project_stack, never as "you like Rust".What it will refuse to remember
Blocked in codeConfidence — and why AIE stays quiet when it is unsure
The productEvery belief carries a confidence score. What AIE does with a belief depends entirely on that number:
This gap is the whole point. A memory system that acts on a 0.72 guess is worse than one with no memory at all — it is confidently wrong about you, and you cannot see why.
It forgets, on purpose
A belief AIE only inferred loses confidence over time — half-life 90 days. If you stop behaving that way, it fades.
But a preference you stated out loud never decays. Silence is not retraction. If you told AIE something once, it does not quietly forget because you did not repeat yourself.
And decay has a floor (0.30) — it can weaken a belief, never erase the evidence that you said it.
Correcting it
When AIE gets something wrong, tell it — in Corrections & Prevention. A correction can become a prevention rule: a standing instruction the intent gate checks before doing anything consequential.
Rules are matched by meaning, not keywords. "Never replace locked assets" also fires on "swap out the pinned files".
If the matcher cannot decide, it blocks. A safety rule that fails open is not a safety rule.
You can check everything
ReceiptsWhat is NOT built yet
Honest listBecause a manual that oversells is the same lie as a button that does nothing:
Every tab, and what it does
Audited — every control clickedaie_sessions has no column for them, and a filter that always returned nothing would be a lie about your data.Try it in sixty seconds
always reply in Englishprefers_language: English is now there, with a confidence and a receipt.