L2 · Intelligence
No signal exits as a mention.
Every signal that enters the system exits as one of four dispositions — ignore, watch, brief, act — with a named owner, a materiality score, a precedent match and a simulated consequence. No signal exits as a mention.
Pipeline stages
12
Materiality drivers
6
Sentiment dimensions
4
Feedback loops
5
The next-best-action engine
Twelve stages from raw signal to decision
Stages after DEDUPE are interpretation — the work the function currently does by hand.
Signal → disposition · twelve stages
Raw signal with channel, language, market, author resolution attempt.
Always include "do nothing"
With its projected consequence. An action engine without a null option manufactures work and trains its users to ignore it.
Show the ranking logic
Every recommendation exposes its drivers. A comms leader who cannot see why the engine ranked something first will override it once, then stop reading it.
Deadlines scale with posture
4 hours normal, 30 minutes elevated, 8 minutes crisis. On expiry, escalate one level automatically. A recommendation with no deadline is a suggestion, and suggestions accumulate.
“Even on the media monitor, you can do all kinds of volume [work]. But then we've got to go and still do a lot of manual interpretation of does this matter or not.”
Supports · Stages 3–8 exist to move interpretation into the machine without flattening it.
Materiality model
Six weighted factors, one number, named drivers
Velocity, stakeholder influence and business exposure carry the most weight. Move the sliders to change the disposition.
Six weighted drivers · 0–100 each
Audience actually exposed, not impressions. Weighted by audience quality against the Stakeholder Graph.
Rate of spread and — critically — the second derivative. An accelerating small story outranks a large flat one.
Distance from the Canon position. A story that contradicts what the company has said is material regardless of size.
Graph centrality of the voices carrying it. One regulator outranks ten thousand consumers on a regulatory topic.
Revenue, operations, talent or license-to-operate at stake, from the Policy Register exposure map.
How the closest historical analogs actually resolved. Pattern-matched, not asserted.
Materiality
Named drivers · ranked by contribution
The drivers matter more than the score. This is what a gate-holder reads first.
Factor weights (×100)
Weights sum to 1.00. Velocity is weighted above reach deliberately: rate of change is the single best early predictor.
Calibration
- method
- Rank correlation against the CCO's own ordering of a 200-story human-labelled set.
- target
- 0.8
- cadence
- quarterly
- failure signature
- Rising story count with a falling act rate. Recall optimized over judgment.
Sentiment model
Positive/negative is a vanity metric. Four dimensions, scored independently, are the minimum that supports a decision.
Never averaged into a single score. Reported as a four-vector with a trajectory.
Valence
-1 → 1Favorable or unfavorable?
−100 to +100, per cohort, never aggregated across cohorts.
The conventional dimension, kept — but only as one of four, and never reported as a company-level average.
Salience
0 → 1Do they care, and how much?
0–100, from engagement depth, repeat exposure and unprompted mention rate.
Mild negativity that nobody cares about is noise. Intense positivity about something irrelevant is also noise. Salience is the filter that makes valence actionable.
Attribution
0 → 1Is this about us, our sector, or our category?
Company / sector / category / individual actor, with confidence.
The dimension whose absence causes the most wasted effort. Half of what a monitoring tool flags as "negative brand sentiment" is category sentiment the company cannot move and should not try to.
Narrative alignment
-1 → 1Is what they believe what we said?
Distance from the Canon position, by topic.
The dimension no vendor offers, because it requires a machine-readable corporate position — which is the Canon. This is where the platform's architecture pays off directly.
Four-vector read · franchisee cohort, value dispute
Reported as a four-vector with a trajectory, never averaged into one number. High salience with rising attribution is the shape that predicts regulatory follow-on.
Cohort resolution, always
Sentiment is reported per cohort and never as a single company number. For McDonald's the mandatory cohorts are franchisee (by tenure and market), crew, customer value-seeker, customer quality-seeker, investor, regulator, journalist, advocacy, and supplier. A number that averages franchisee anger with customer contentment is true of nobody.
Velocity and the second derivative
The level matters less than the rate of change, and the rate of change matters less than whether it is accelerating. A cohort at −20 and stable is a fact. A cohort at −5 accelerating downward is an event. The engine ranks on the second derivative.
Driver decomposition on every move
The output is never "sentiment fell 8 points"; it is "franchisee valence fell 14 points in three markets, driven by these four operator-forum threads about renewal standards, originating from these two association communications."
The forecast, not the history
Where the Precedent Library has comparable trajectories, the engine projects where a cohort is heading and what typically arrests the slide. A sentiment system that only describes the past is a rear-view mirror sold as a windshield.
Market scanning
Four cadences, one queue
The continuous layer is narrow and crisis-grade; the weekly layer is wide and structural. Nothing in between is sent unless it clears a threshold.
Continuous · crisis-grade
Sub-5-minute pollingWatchlist terms, named executives, active situations, regulatory dockets in open comment.
A narrow, configured watchlist only: named crisis triggers, regulatory dockets with live matters, executives’ names, active litigation, and the top 200 Stakeholder Graph nodes. Threshold is deliberately low; output is a single alert, routed by severity.
CRISIS extends continuous scope to all brand terms in all operating markets.
Hourly
60 minutesTier 1 media, sell-side notes, platform-level social volume, review platforms.
Full earned and social sweep with materiality scoring. Outputs only what clears the briefing threshold. This is the layer that replaces the morning clip book, and it replaces it by refusing to send anything that does not matter.
Daily
Fixed, pre-briefTrade press, local-language sources across all operating markets, employee channels, competitor output, AI answer surfaces.
Synthesis, not aggregation: what changed, what it means, what to do. Competitive SOV movement, cohort sentiment deltas with drivers, new early-warning cards, and the ranked action set. Maximum one page plus drill-downs.
Weekly / Monthly
ScheduledRegulatory horizon across the top 20 markets, legislative calendars, activist and NGO campaign pipelines, narrative-share benchmarking.
Trend and structure: narrative health and drift, SOV position by topic and geography, scenario library refresh, stakeholder graph changes, campaign performance, platform health, and the Precedent Library additions.
"The ability for us to get a really comprehensive, tailored to our business model view of the global regulatory processes in place in the top 20 markets. I have it in half an hour. That used to be months of agencies working country by country and consolidating."
Under-monitored
Where the story actually starts
Conventional monitoring covers earned, broadcast, wire, trade, social and review platforms. These seven are where the signal appears first.
Walled gardens
Substack, Beehiiv, Patreon and equivalent. The highest-influence commentary in many sectors has moved behind them and conventional monitoring does not see it at all.
AI answer surfaces
Scheduled API-based probing of major assistants across a fixed question set, in multiple languages — A5's remit. For a consumer brand this is already a higher-volume answer channel than the corporate newsroom.
Franchisee and operator communities
In a franchised model this is the earliest-warning channel that exists, and almost no corporate monitoring program covers it.
Regulatory and litigation dockets
In all operating jurisdictions. Not the top five markets. Twenty is the current frontier; 114 is the target.
Review and local-rating platforms
For a restaurant system these are the leading indicator of an operational issue becoming a communications issue, and they are geo-resolvable to individual restaurants.
Employee and crew channels, where permitted
Subject to a jurisdictional permission matrix; off by default in works-council markets. Gated by G3 with HR and Legal present.
Prediction markets and short-interest signals
Noisy, and also directionally honest in a way survey data is not.
B5 · reaction simulator
Predict named-audience reaction before release, not measure it after.
Five rows build a persona. The simulator never predicts an individual's behavior and never holds a gate.
The five-row persona model
| Who am I? | Role, tenure, mandate, incentives, what success looks like in this seat. |
| How do I relate to the company? | Nature and history of the relationship; dependency; leverage in both directions. |
| What do I care about? | The two or three things that actually move this person, ranked. |
| What company data reaches me? | Which disclosures, communications and coverage this seat actually sees — often far less than the company assumes. |
| Live data | Current Signal Lake and Stakeholder Graph state for this actor or actor type: recent statements, current posture, open issues. |
Free selection of class, segment and seat
More selections produce a blended persona; fewer seats produce more specific reactions; fewer classes produce more rationale detail.
A scrutiny slider
Models the difference between a skim and a forensic read. The same statement survives one and not the other, and knowing which is the actual decision.
A context layer
Import files or prompt context specific to the situation — a draft, a prior transcript, a regulatory letter — so the simulation is grounded in the actual material.
Multi-pane review
Every persona's reaction shown side by side with its rationale and the specific passage that triggered it. Passage-level attribution is what makes the output editable rather than merely interesting.
Versioning against the draft
Re-run after an edit and see which reactions changed. This turns simulation from a one-shot check into an iteration loop.
Hard rules
- Real identities are never passed to the simulator. Pseudonymized graph nodes only.
- A simulator output is never presented as a prediction of an individual's behavior.
- Simulation is advisory input to a gate. It is never a gate.
Value anchor
In an analogous, heavily scrutinized and regulated industry, BCG research observed 4%-15% negative share-price impact following large-scale negative media attention events.
D1 · intake and inbox
The inbox is not a communications channel. It is the function's unstructured work queue, and it is where the coherence failure is most visible.
Eight steps. Seven of them run without a human. The one that does not is the release gate.
Feedback
Five feedback loops
Every loop names what it measures, what it feeds and who owns it.
Outcome loop
continuousObserved outcome vs. predicted outcome for every action taken.
owner · F1
Human-override loop
continuousEvery gate rejection, every edit, every ceiling ratchet — with the reason.
owner · Agent Ops Lead
An override with no recorded reason is a lost training signal. Reason capture is mandatory at every gate.
Canon-drift loop
weeklyPositions past expiry; evidence past expiry; voice edit distance trend; unclassified rate.
owner · Canon Steward
Simulator-calibration loop
quarterlySimulated reaction vs. observed reaction, by audience class, segment and seat.
owner · Simulation Calibrator
Slop Sentinel
weeklyOutput volume vs. engagement per output, per channel. Detects the failure where the system produces more and lands less.
owner · Agent Ops Lead
The fifth loop · the Slop Sentinel
Runs on every piece of content before it reaches a gate and reports three things: density of generic construction (the tell-phrases, the tricolon, the "it’s not just X, it’s Y"), distance from the house or principal’s voice corpus, and information density — the ratio of specific, checkable assertions to total words. Content below threshold is returned to the agent rather than passed forward.
This is a quality floor enforced by software, and it exists because the alternative is discovering the problem in aggregate, eighteen months later, when someone notices that everything the company publishes now sounds the same.
The meta-loop: gate load as the health metric
The single number that tells you whether CHORUS is working is gates per person per week, trending down while output volume trends up. Rising gate load means autonomy ceilings are too low or materiality thresholds are miscalibrated. Falling gate load with rising override rates means autonomy has been pushed past what the humans trust. Falling gate load with falling override rates is the platform working as designed.