01
02
03
04
05
06
01Beacon

Understand a market you cannot read.

Beacon measures what a market says — unprompted, in its own language, as it happens.

Global monitoring platforms are built for high-resource languages. In small-language markets their coverage thins, their language models misfire, and the conversation that actually decides outcomes stays invisible. Beacon is the instrument for exactly that gap: depth-first collection, native-language classification, and propagation modelling in markets no incumbent tool reaches.

See the method Request a briefing ↗
≈17×
Retrieval depth · COMPUTED
More local-language discourse retrieved per topic than leading global monitoring platforms, in head-to-head benchmarks. Benchmark set and date stated on every engagement.
24/7
Continuous collection
Narratives are registered as they form, not after they land.
0
Individuals profiled
Every export is aggregate-only, with a minimum cell size enforced in code.
↗ ↗ ↗
02Method

Two instruments.
Two different objects.

A survey and a body of public discourse get treated as rival answers to one question: what does the public think? They are not two routes to one number.

A survey estimates the distribution of elicited attitudes inside a defined sample, at a chosen moment. It is a reactive instrument — the subject responds to being studied — and it hands the researcher four degrees of freedom the respondent does not have: the agenda, the frame, the menu, and the timing.

Discourse analysis observes what a public expresses, unprompted, in its own terms, with no researcher in the room. Different object. Different failure modes.

Reactive — the survey

Elicited attitudes

Estimates the aggregate of private attitudes. Representative, quantified, periodic. Expensive to repeat, and so repeated rarely.

Non-reactive — Beacon

Expressed discourse

Observes the public sphere: what a population articulates, circulates, and contests. Live, generative, relational — and available before a survey could be fielded.

“Discourse tells you what to ask, and when. The survey tells you how many.”
We do not sell a cheaper poll. We sell the layer a poll was never built to see.
03Areas

What Beacon
does.

01

Coverage

Question-led collection in the local language and script, at a depth global tools do not reach. Not a keyword sample of an unbounded system — the bounded discourse of one market, with the coverage record stated on every deliverable.

02

Narrative mapping

What is circulating, in whose words, with what framing. Frames are extracted natively, not translated in from an English taxonomy that was never built for this market.

03

Propagation

How far, how fast, and whether it is self-sustaining. Cascade dynamics are modelled, not eyeballed — the branching ratio tells you within hours whether something fizzles or runs.

04

Authenticity

Organic uptake separated from coordinated amplification. Scored at cluster level, with the organic-alternative hypothesis always stated. Never per account.

05

Causal impact

Did the event actually move the conversation, or would it have moved anyway? Counterfactual estimated with credible intervals, reported both pointwise and cumulatively.

06

Emerging frames

The thing a message becomes once it leaves your hands. New frames are detected as they split off — early warning, hours before a reframe dominates.

“Coordination is computed. Inauthenticity is judged.”
Doctrine · D7
See the process ↗
04Process

How it
works.

Observe [+]Classify [+]Model [+]Report
.01

Observe

Native-language collection across the platforms where the market actually talks. Script normalisation, deduplication, entity resolution, and coordination scoring all run before any inference. No model consumes raw harvest.

.02

Classify

Frames, stance, and discrete emotion extracted per utterance and per author-bundle, against a versioned local taxonomy. Every classifier is a registered instrument with a version, a gold set, and a calibration record.

.03

Model

Four families, four questions. Point processes for cascade size. Compartmental models for adoption versus rejection. Structural time series for causal impact. Frame clustering for mutation. Coupled into one system, not four dashboards.

.04

Report

A predictive distribution and explicit tail flags — not a point estimate dressed as certainty. Every figure carries its evidence class and its denominator.

05Evidence

Every number
traces back.

The reason to trust a measurement is not the confidence of the vendor stating it. It is whether the number can be reconstructed. Every figure Beacon reports is rebuildable from a run ID: the data snapshot it read, the instrument version that produced it, the calibration artifact in force, and the gate results that let it through.

That property — not any single model — is the asset. It is cheap to build in and prohibitively expensive to retrofit, which is why most of the category does not have it.

Measured

Directly observed

Directly observed in collected data. The discourse layer itself.

Computed

Deterministically derived

Counts, rates, network metrics — derived from measured data with no inference step.

Modeled

Inferred or projected

Inferred, classified, or projected. Never enters the measured layer.

Vendor · Attributed

Third-party enrichment

Always attributed, always stamped modeled-class.

06Clients

Built for
the outsider.

Multinationals & operators
  • > Market entry reads
  • > Standing watch
  • > Stakeholder mapping
  • > Incident briefs
Investors & emerging-market capital
  • > Country reads
  • > Sector narrative risk
  • > Early-warning alerts
  • > Diligence support
International media
  • > Story verification
  • > Source provenance
  • > Coverage context
  • > Local framing briefs
Communications & risk advisors
  • > White-label reads
  • > Campaign context
  • > Reputational monitoring
  • > Evidence packs
Development & multilateral institutions
  • > Programme perception
  • > Information-integrity assessments
  • > Baseline and repeat reads
Research partners
  • > Corpus access under agreement
  • > Methodology review
  • > Co-published validation
07Stack

No black box.

The methodology is stated, not implied. A buyer who cannot interrogate the method cannot rely on the output.

Point process

Cascade dynamics

Reshare activity is self-exciting: each event raises the probability of the next. Fitting a Hawkes process yields the branching ratio n — the expected number of direct offspring per event. Below one, the cascade dies out. At one, it is self-sustaining. Estimated from the first hours, it is how a fizzle is separated from a run before either is obvious.

Compartmental

Adoption and rejection

SEIZ tracks flows between belief states: susceptible, exposed, spreading, and skeptic. The skeptic compartment is the one that matters in a contested space — it captures the substantial share of a population that sees a narrative and actively rejects it. The fitted rates answer whether a message is converting the middle into spreaders or into skeptics.

Structural time series

Causal impact

Fit a state-space model to the pre-event period using control series that were not exposed, project the counterfactual forward, and report the gap with full posterior intervals. This is what permits "this raised conversation by an estimated Y%, interval [a, b]" instead of "conversation went up afterwards."

Frame extraction

Mutation

Narratives do not merely spread; they mutate. Rolling-window clustering detects a frame as it splits off, and the cross-excitation term between frames measures one frame igniting another — a capture signal that is monitorable rather than merely narratable after the fact.

Reference class

Forecast priors

A proprietary episode library records how past episodes started and how they ended. A new episode retrieves its nearest historical analogues and uses their trajectories as an empirical prior. It works from the first dozen cases and improves monotonically. It cannot be purchased in these languages.

Evaluation harness

Native-language scores

Every language-dependent component reports a measured score on a held-out, human-labelled local benchmark before it ships. Regressions block release. The gold sets exist and are versioned: label protocol, set sizes, and splits are recorded in an artifact register, summary scores are disclosed in the methodology paper, and the composition is available for review under agreement. This is the difference between running a language model over local text and knowing what its precision is, and where it degrades.

08Coverage

One market,
at depth.

Beacon operates today at national depth in Georgia — a bounded discourse system collected question-led rather than keyword-sampled, benchmarked at ≈17× the per-topic retrieval of global platforms on Georgian topics. Depth in a bounded system is not a larger sample. It is a different class of asset.

The method is portable. Where a market meets the conditions — a knowable platform set, a language the incumbents underserve, and a legal basis for collection — Beacon can be stood up as an operated engagement or licensed to a local operator who owns their own national corpus.

Standing watch

An annual retainer. Continuous monitoring with alerting against agreed thresholds. Most of the value on most days is the confirmation that nothing is forming.

Market read

A bounded engagement. One market, one question, one delivery window. Full methodology appendix included.

Method licence

For operators building in their own market. Architecture, taxonomy governance, evaluation harness, and calibration discipline, transferred and supported.

Pricing is annual-first and quoted in round numbers. No charm pricing.

09Boundary

The limits,
stated.

An instrument is only as credible as its account of its own limits. These are structural, not incidental — and stating them plainly is the condition of using the instrument well.

.01

We do not profile individuals.

Individual-level inference does not leave the inference boundary. Exports are aggregate-only with a minimum cell size enforced in schema, not in a policy document. Psychographic and emotion-inference layers are not offered on this line.

.02

We do not attribute.

Beacon produces graded, evidenced inference of coordinated activity. It does not produce courtroom-grade attribution to a named actor or agency — that requires data no external system will ever hold. Any vendor claiming otherwise is selling marketing ahead of the mathematics.

.03

We do not forecast the move.

Beacon sees the move forming. The language will escalate only when a public prediction ledger validates it, and not before.

.04

We do not replace polling.

Discourse analysis does not produce a representative estimate of what a population privately believes. Any claim that it does is a misuse of the instrument. Where you need a weighted number, field a survey — Beacon tells you which survey is now worth fielding.

.05

We do not run campaigns.

Beacon is a measurement instrument. We do not conduct advocacy, influence, or persuasion work on this line, for any client, in any market.

.06

We do not hide the population problem.

The population that speaks is not the population. Participation is dominated by a small and atypical minority, and public speech is performed for an audience. Every deliverable carries its coverage record: who was actually measured, and who was not.

10FAQ

Questions,
answered.

Is this social listening? +
No. Listening platforms are built for breadth across high-resource languages and thin out precisely where a small-language market needs depth. Beacon is built the other way around: depth in one bounded system, question-led, with native script handling, a local frame taxonomy, and a measured evaluation score on local text.
How is this different from a synthetic-audience or AI-simulation tool? +
Those systems simulate a population and infer reaction from model-people. Beacon observes the real discourse and measures what actually happened and how it moved. Different objects — comparing accuracy across the two is usually a category error. Where a simulation's hard problem is proving its synthetic population resembles a real one, ours is coverage and calibration, and we report both.
Can you tell us who is behind a campaign? +
We can tell you, with graded confidence and the specific evidence behind it, that a cluster is behaving as a coordinated network, when it started, what it is pushing, and whether it moved domestic conversation. Naming the sponsor is an investigative conclusion for a human analyst, informed by our evidence and never asserted by the system.
Do you work with political clients? +
Not on this line. Beacon's export practice is a measurement product sold to commercial, institutional, and research buyers. Political work is a separate entity, separate paper, and separate brand, and the two never co-appear on a client deliverable.
What is the legal basis? +
Collection is of public expression, under a documented retention policy and an impact assessment completed before any client-facing deployment. Aggregate-only export and minimum-cell-size enforcement are architectural, not procedural. We route data-protection questions through counsel and will share the posture in a briefing.
How fast is a read? +
A standing watch alerts continuously. A bounded market read is scoped in weeks, not months, because the collection substrate already exists. What we will not do is compress the calibration step to hit a date.
What accuracy do you claim? +
Per component, measured on a held-out local benchmark, and disclosed. Not a headline percentage. Vendor-stated accuracy figures in this category are marketing claims until someone shows you the gold set and the split — and we show ours: register, protocol, and splits, under agreement.
Got another question?Send message ↗