Signal Intelligence · Viska Conviction Engine

Signal News Scan

Each morning, after the council settles its convictions, the scan reads the day's news against every live thesis and reports back three things: what confirms it, what threatens it, and what merely surrounds it.

StatusLive in production
Sub-workflowTS80bLFkGCb1IG4V
Host reportyWvHQU8q8hvmIkEV
CadenceDaily · 06:00 UTC
First live proofexec 39870 · 16 themes
01

Why it exists

The conviction council is excellent at forming a view — it weighs the brokers, weighs the theses, and lands a net signal per theme. What it does not do on its own is check that view against what happened in the world today.

A conviction is only as good as its exposure to disconfirming evidence. The Signal News Scan closes that gap: it turns each standing thesis into a daily question — did reality move toward this thesis or away from it? — and answers it with cited, current developments, right inside the same client report the conviction already ships in.

The core idea in one line

The council decides what we believe. The scan tests, every day, whether the news still agrees.

02

The concept: news, judged against a thesis

Most news pipelines summarize. This one adjudicates. Every candidate development is paired with a specific theme's thesis and its stated invalidation trigger, and the classifier returns one of three stances — nothing else.

VALIDATES

The development moves reality toward the thesis. Evidence the council's read is holding.

INVALIDATES

The development cuts against the thesis, or trips its invalidation trigger. The signal to look at first.

CONTEXT

Relevant and worth seeing, but not decisive either way. Deliberately un-opinionated — it does not inflate the case.

Grounding rule

The classifier may only cite items from the day's fetched pool — never its own memory. A thesis it can't support with a real, dated, linked article gets CONTEXT or nothing. No invented evidence, ever.

What a live thesis looks like

theme · ai-infrastructure · bull 0.23 · 7 brokers

Thesis. Panel is net bullish on structural AI-hardware demand — cleaner semiconductor setup, power-architecture change — against Zerohedge's conditional bear that AI capex is cyclical and vulnerable if the largest spenders change behavior.

Invalidation trigger. A meaningful pullback or behavioral shift in capex plans by the largest AI spenders.

03

How it triggers & runs

The scan is not a separate schedule to babysit. It is a sub-workflow the daily report calls inline, immediately after the conviction council has produced its signals — so it always classifies against that morning's convictions, never stale ones. Its output is handed back as a rendered HTML block and injected straight into the report body, after the conviction section.

Trigger
Daily Report
Calls the scan after the council, passthrough inputs
Read
Read Signals
theme_intelligence — thesis + invalidation per theme
Shape
Build Targets
Theme keywords + scoped search queries
Source A
Curated RSS
Reuters · AP · CNBC · FT · EIA
Source B
Web Search
Thesis-scoped, gpt-5.5 built-in tool
Pool
Merge & Dedupe
One citable pool, hashed by url+title
Judge
Classify vs Thesis
VALIDATES / INVALIDATES / CONTEXT
Store
Persist
signal_news_developments (history)
Return
Render → Report
developments_html, back into the email
Fail-open by design

Every stage tolerates emptiness. A dead feed, an empty search, a malformed model reply — none of it can break the daily report. Worst case the section reads "No material developments," and the rest of the report ships untouched.

04

How the data is shown

The reader never sees a pipeline. They see one new section in the daily report, grouped by theme, with the most consequential stance first — invalidations, then confirmations, then context. Each line is a plain-language read plus a cited, clickable source. Quiet themes say so, rather than padding.

Below: the actual rendered output from the first live run (exec 39870), verbatim.

Latest developments vs thesis · injected after conviction · 2026-07-11
ai-infrastructure · bull 0.23
  • VALIDATES SK Hynix's Nasdaq debut surge and management's "demand is enormous" comment support the structural AI-hardware-demand framing. — CNBC
  • VALIDATES Relaxed US export controls on advanced chips for the UAE back the constructive hardware-demand side of the thesis. — FT
  • CONTEXT Waning demand for longer-dated AI debt signals investor skepticism on long-term profitability — not an actual capex pullback. — CNBC
ai-memory · neutral −0.04
  • INVALIDATES SK Hynix's sharp debut gain and "demand is enormous" point to renewed AI-memory momentum, challenging the cautious, hedged read. — CNBC
oil-short · bear −0.07
  • VALIDATES Renewed fighting over the Strait of Hormuz directly supports the concern that Hormuz supply risk can challenge oil-short positioning. — CNBC
gold · neutral 0.00

No material developments since analysis.

05

Sources: what we weighed

The design question was never "how much news can we pull" — it was how much signal per unit of noise, without contaminating the corpus that drives conviction. Three constraints shaped every source decision.

Constraint · grounding

Citable or it doesn't exist

A source only counts if it produces a dated, linkable item. That rules out anything the model can't attribute — no "general sentiment," no un-sourced claims.

Constraint · corpus purity

Never touch the broker corpus

News lives in its own store. The broker research embeddings that feed the council are held completely separate — a hard invariant, so daily noise can never leak into conviction.

Constraint · cost & cadence

One pass, bounded spend

A single daily run: one scoped search, one classifier pass, one write. Breadth is added by editing an allowlist, not by widening an expensive net.

Curated vs open — the split

Curated RSS gives reliable, reputable baseline coverage at near-zero cost. Thesis-scoped web search reaches the specific, the breaking, and the long-tail a fixed feed list misses. Running both, then merging into one deduped pool, is what produced the SK Hynix and Hormuz reads above — neither was on an RSS front page.

06

Sources today & candidates

The starter allowlist is intentionally small and reputable — chosen to prove the loop, not to be exhaustive. Adding a wire is a one-line change; the table below is the roadmap, not a limit.

SourceKindWhyStatus
ReutersWire · RSSBroad, fast, reputable market & macro baselineLIVE
APWire · RSSGeopolitical & breaking, low-bias baselineLIVE
CNBCMarkets · RSSEquity & single-name market coverageLIVE
FTMarkets · RSSMacro, policy, and cross-border finance depthLIVE
EIAEnergy · RSSPrimary-source energy data for energy/oil thesesLIVE
Web searchOpen · scopedThesis-targeted long-tail & breaking, per themeLIVE
Bloomberg / WSJMarkets · RSSDeepen equity & macro; some paywall frictionCANDIDATE
Sector trade pressVertical · RSSSemis, defense, mining, grid — match named themesCANDIDATE
Regulatory / filingsPrimary · feedSEC, central banks, statutory registries — hard evidenceCANDIDATE
Earnings / transcriptsPrimary · APICapex & guidance — direct on invalidation triggersCANDIDATE
Curated X / socialSocial · APIEarliest signal; high noise — needs strict filteringCANDIDATE
Adding a source

Every candidate is judged on the same three constraints from Section 05. A wire that emits clean, dated, linkable items and stays out of the broker corpus is a one-line allowlist edit. Anything noisier (social, transcripts) earns a filtering layer before it's trusted.

07

What it unlocks for the conviction engine

Today the scan is a daily read. Its real value is that it makes conviction a closed loop — beliefs that are continuously tested against reality and can be scored, tracked, and acted on over time. The stances are already persisted per theme per day; that history is the raw material for everything below.

→ forms
Council

Brokers & theses land a net signal per theme.

→ tests
Scan

Each thesis met with today's cited developments.

→ scores
Stance history

VALIDATES / INVALIDATES stored per theme, per day.

↺ feeds back
Conviction

Reality-support informs the next read.

AInvalidation alerts
An INVALIDATES on a high-conviction theme is the single most valuable line in the report — surface it as a standalone alert, not just a buried bullet.
BReality-support trend
Roll the stance history into a per-theme confirmation trend — is the news drifting toward or away from a conviction over weeks, not just today.
CConfidence adjustment
A theme accumulating invalidations while the council stays bullish is a divergence worth flagging into the score — the scan as an independent second opinion.
DThesis-durability backtest
With dated stances persisted, ask historically: which theses held up under news pressure, and which decayed — a measurable track record for the engine itself.
The shape of it

Conviction stops being a snapshot the council prints each morning and becomes a position that has to keep earning its place against the day's evidence. The scan is how it's held accountable.