Truth and discernment

DreamLayer's flagship capability is a live read on what is being said to you: Veritas checks the content of claims, Truth Lens reads the delivery, and Discernment fuses the two — plus the speaker's history — into one graded stance. The three are deliberately separate engines answering different questions on different inputs, composed at the end. A fourth conversational copilot, answer-ahead, rides the same caption stream.

Everything in this chapter is engineered to underwhelm on purpose: strict claim filters, confidence floors, per-speaker cooldowns, and an explicit "unverified" verdict. A fact-checker that fires constantly, or overclaims, is worse than none.

Veritas — the live fact-checker

Off by default; enabled with set_factcheck(True) (the phone's "Live fact-checker" toggle). While on, every other-speaker line landing in ingest_caption runs two passes (orchestrator/veritas.py):

Pass 1 — self-contradiction (fully offline)

The line is compared against the same speaker's own earlier lines from the conversation ledger, using Candor's contradicts predicate (at least two shared content words with an opposing assertion). A clash produces the strongest verdict in the system — self_contradiction, confidence 0.9 — with the receipt quoted on the card: earlier: "we settled at two million". No network, no model, no cloud: this pass works on a plane.

Pass 2 — the world check

Only a checkable claim is ever escalated:

• Not a question (trailing "?" disqualifies).
• Not hedged — a leading "I think / maybe / probably / it seems / in my opinion..." disqualifies.
• And it must either contain a number-like fact (years, percentages, quantities, money, distances) or a factual predicate (an is/are/was assertion touching a fact word: capital, invented, founded, discovered, largest, president, boiling point...).

The claim goes to _verify_claim, which is real and tiered: it wraps the claim in a tightly-shaped verification prompt (ai_brain/verify.py: VERIFY_PROMPT demands exactly one line — VERDICT: SUPPORTED|DISPUTED|UNVERIFIED — reason) and routes it through brain.ask — your local model first, the cloud tier only if you have opted in, and never while incognito. The reply is parsed into {verdict, basis, confidence}: unverified maps to confidence 0.3; supported and disputed map to 0.8, knocked down to 0.55 if the model hedged. If no tier can answer (offline, no model), the seam returns None — and the offline self-contradiction pass still stands alone.

When it actually speaks

Sparingly, by construction:

• One verdict per speaker per 45 seconds (per-speaker cooldown).
• Above the floor only: a disputed needs confidence >= 0.55; a supported needs >= 0.85 to bother congratulating anyone.
• Held during Focus; silenced by the Veil.

The card

Verdict tone is resolved on the device renderer itself, so a wire-delivered card can never lose its color: green VERIFIED (chime earcon), amber CHECK THIS (urgent hark earcon, double haptic, flash), red THEY SAID DIFFERENT BEFORE (hark, double haptic, flash), ghost UNVERIFIED (hark). Dismisses after 7 seconds.

Truth Lens — the delivery read

Off by default; enabled with set_truthlens(True). Where Veritas asks is the claim true, Truth Lens asks how is it being delivered — a nine-stage multimodal pipeline (truth_lens/): face embedding, action-unit detection, prosody, linguistic analysis, baseline lookup, fusion, rendering, baseline update, and a disabled fact-check stub (Veritas owns content).

Three channels, honestly labeled:

Calibration is the point

Every read is scored as a z-score against that person's own baseline (set_contact), built with an online Welford update. A baseline needs 10 samples to calibrate; until then the person is treated as a stranger, and a stranger's micro-expressions stay noise: stranger fusion only alerts if all channels exceed 0.75, and its confidence is fixed at 0.2. The output is a CredibilityVector — deception probability, confidence, per-channel z-scores, dominant channel, stranger flag — with labels stepping CREDIBLE (< 0.40), UNCERTAIN (< 0.65), ELEVATED (< 0.85), HIGH ALERT (>= 0.85), and CALIBRATING below 0.3 confidence.

On glass it renders as the testimony thread — nine stages around the ring, truthful stages as continuous arc, deceptive ones torn:

With Truth Lens on, the orchestrator runs a delivery read for the current speaker on every other-speaker line (_read_delivery), without the HUD display gate — because Discernment wants the reassuring reads too — and feeds the result in via note_credibility(speaker, vector).

Discernment — three lenses, one read

orchestrator/discernment.py: discern(fact, credibility=None, history=0) composes content, delivery, and history into one stance. The scoring is plain and inspectable:

• Content concern: self-contradiction 0.85, disputed 0.80, unverified 0.20, supported 0.
• Delivery counts only once the baseline is calibrated (vector confidence

  = 0.3): deceptive at deception probability >= 0.65, credible below 0.40. Delivery contributes at half weight.

• History: each prior flag on this speaker adds 0.15, capped at three.
• Synergy: a false claim delivered deceptively earns an extra 0.2 — content and delivery agreeing is the strongest signal there is.

The weight maps to a stance: flag (>= 0.8), caution (>= 0.5), note (>= 0.25), trust — and a human headline that says what the fusion actually thinks:

"Doesn't add up — and it didn't sound like it, either." "The claim is off, but they seem to mean it." — a false claim delivered sincerely reads as an honest mistake, not a lie. "Checks out — but the delivery was uneasy." "They contradicted their earlier words. Not the first time."

The stance and headline land on the FactCheckCard, and the footer carries the tag ("Marcus - elevated - seen before").

Answer-ahead — the conversational copilot

Off by default; enabled with set_copilot(True) (the phone's "Answer-ahead" toggle). When someone else asks a question in the caption stream, orchestrator/answer_ahead.py decides whether it is a real, answerable question — a wh-question, or one aimed at you — and never a rhetorical "...right?" or "you know?". If so, it pre-fetches the answer from your own knowledge through the same tiered router (cloud only when opted in) and flashes an AnswerAheadCard you can read and say yourself.

Its manners are the feature: no wake word, no earcon — silent by design (a tick haptic only), a cooldown between fires, a confidence floor beneath which nothing surfaces, held during Focus, gated by the Veil, and a clean miss (offline or low confidence) surfaces nothing at all.

What is implemented versus seam, precisely