How the Numbers Are Computed
This page documents the exact formulas, weights, and thresholds used throughout the demo, so that every number can be checked. Nothing here is editorial — the confidence scores, quality gate pass rates, consensus distribution, and divergence statistics are all computed from the snapshot by functions whose logic is reproduced below.
Confidence
Confidence is a weighted composite of five factors. The two agents use different weights: the Analyst weights fact verifiability highest (0.35), while Gossip Girl weights LLM self-reported confidence highest (0.40) — tell clarity, not analytical confidence.
| Factor | Analyst | Gossip Girl |
|---|---|---|
| sourceCredibility | 0.20 | 0.20 |
| contentQuality | 0.20 | — |
| factConfidence | 0.35 | — |
| themeEvidence | 0.25 | 0.15 |
| llmConfidence | — | 0.40 |
| behavioralPatternScore (facts/10, capped) | — | 0.25 |
Four post-composite modifiers are then applied multiplicatively: engagementBoost (for SOCIAL sources only, based on log-scaled engagement score and comments), preferenceMatch(×1.15 when the signal’s source type is in the persona’s declared preferences), recencyMultiplier(time decay from publication date, 1.0 for <7 days down to 0.65 for >1 year), and finally the result is clamped to [0, 1].
The recency term callsnew Date()internally, so scores are wall-clock dependent and not reproducible under static export. The demo’s confidences are authored at snapshot-build time and do not recompute.The final score maps to one of four confidence bands: High Confidence ≥0.85 Likely ≥0.70 Uncertain ≥0.55 Low Confidence <0.55.
Quality Gate
The quality gate is a pre-analysis filter that rejects signals below a threshold before they reach the expensive LLM pipeline. It scores three components and compares the weighted sum to a threshold of 0.4.
| Component | Weight | Measures |
|---|---|---|
| substanceScore | 0.40 | Zero-shot classification (substantive vs. boilerplate vs. irrelevant) |
| infoDensity | 0.30 | Entities, numbers, dates per paragraph, capped at 10 |
| companyStrength | 0.30 | Company name and ticker mention strength, capped at 3 mentions |
Four rejection reasons are mapped 1:1 onto the live strings in src/lib/nlp/quality-gate.ts: too-thin (content below minimum length or substance), off-entity (weak company mention), boilerplate (low substance classification), and non-English (language detection failure).
The gate fails open on exception: if the quality assessment throws, the signal receives score 0.5 and pass: true. This is a real risk disclosure — better to analyze low-quality content than to silently drop high-quality content on a model error.
Consensus
Consensus is computed from the two agents’ sentiment and theme overlap. The function checks whether both agents report the same sentiment and whether they share at least one strategic theme label.
| Level | Rule |
|---|---|
| strong-agreement | Same sentiment AND at least one shared theme label |
| mixed-signals | Same sentiment AND no shared theme labels |
| divergent-views | Different sentiments |
The demo engineers the consensus distribution via consensus-curation.tsto make the feed less monotonous: a fixed set of signals are marked as MIXED (by dropping shared theme labels from the Gossip Girl analysis) and another set as DIVERGENT (by flipping one agent’s sentiment). This is disclosed rather than hidden — the curation layer is a deliberate editorial act on top of the computed baseline.
Momentum
Momentum is a week-over-week weighted velocity with bonuses for agent agreement (×1.3), source diversity (×1.2), and confidence (up to ×1.5), plus a time decay for older signals. The live pipeline computes this from the cluster’s signal history.
The demo authors momentum as a 0..1 scalar rather than computing it from the signal history, because the snapshot does not carry the week-over-week velocity series needed to reproduce the live formula. This is a known simplification — the ranking is plausible but not derived.
Divergence
Divergence is the signed per-signal delta between the Analyst’s confidence and Gossip Girl’s tell strength. A positive delta means the Analyst is more confident than the Gossip Girl is telling; a negative delta means the reverse.
| Vertical | Sector Bias |
|---|---|
| hardware | The Analyst treats a disclosure as performing exactly the work it claims: fourteen forward citations are defensive value, a confidence-weighted thermal fallback is ordinary production engineering, and a full-stack attach rate has no separable split, so declining to quote a networking percentage is a technically accurate refusal. The Gossip Girl reads the same documents for the distance between what they contain and what they were shaped to avoid containing — the citation count is fourteen design-arounds, the refused figure is the one line inside a blended number that can never be marked down, and a claim built around graceful degradation is an admission that the shipping product spends real time at the thermal ceiling. The recurring mechanism is that both agents accept the numbers and disagree about which direction the silence around them points. |
| mobility-cloud | Here the Analyst's baseline is disclosure law and ordinary sequencing: a twenty-month patent filing lag is statutory, two major versions is a deprecation window most enterprise policies already exceed, a same-day documentation republish without a postmortem is unremarkable, and announcing during a comment period is capital discipline rather than leverage. The Gossip Girl's baseline is adjacency and chronology — an invention surfacing this week with no announcement attached, a price cut applied one day after a competitor's incentive change, a statement that defends the committees and never once touches the surveillance allegation. They read the same filing at different layers: the Analyst asks whether a practice is normal, and the Gossip Girl asks why it is being described now, in this order, by this company. |
| consumer | This is the vertical where the two agents disagree least about facts and most about what counts as evidence in the first place. The Analyst discards small, directional and anonymous material — fourteen roles out of a large research organisation is noise, forty thousand likes with no specification is normal for the channel, an uncorroborated podcast claim should be retired from collection rather than debated. The Gossip Girl keeps it and reads composition in place of count: which roles were removed rather than added, which statistic was published and which one was withheld, and the asymmetry between a company telling you how big a system is while declining to say how it is judged. The disagreement is about record-keeping, not analysis — one agent wants a defensible input set, the other wants the asymmetry itself to survive. |
| semiconductors | Both companies disclose heavily and disclose selectively, which is what makes this the sharpest test of the two agents. The Analyst prices disclosure structure — commitments growing faster than revenue is what capacity reservation looks like, a licensing mechanism scoped to named part numbers is narrow enough to absorb inside guidance, withholding the product split of a backlog is competitive discipline, and a timing change around a rising trend is ordinary. The Gossip Girl prices the edits and the edges: page edits converging three times on the same sentence, a regional configuration removed from the brochure while support documents still list it, an eighty percent firm purchase commitment underwriting a demand curve nobody can see past 2027. What the Analyst calls a bounded scope, the Gossip Girl reads as a boundary someone chose and can therefore move. |
| enterprise-cloud | The Analyst reads enterprise disclosure as risk allocation and finds the discipline sound: pass-through clauses move energy cost to the buyers best able to bear it, a model router insulates the platform if a provider's terms move, backlog growing faster than revenue for a fourth quarter is demand that exists. The Gossip Girl reads the same instruments as evidence of a constrained hand — cost controls that ship in response rather than anticipation, a set of listings withdrawn across three offices that was never going public anyway, flexibility conceded quietly and then defended publicly as routine renewal activity. The live question in this sector is always who is holding the downside of the contract, and the two agents answer it from opposite ends of the same document. |
| finance-health | Banks and drug developers disclose differently, but the two agents part company in exactly the same place in both. The Analyst accepts the form of a disclosure as its own explanation: an accrual range left open is what the standard requires and no more, declining to price a launch asset eighteen months out is normal sequencing, conditioning capital plans on a rule that is not final is the conservative choice. The Gossip Girl treats form as a position taken inside a negotiation — a range that cannot be capped is the most useful shape a liability can take, withheld pricing hands the category's reference price to whoever set it first, and a filing that names the champion instead of the champion's competitors is read by the agency as an interested party arriving early. What reads as procedure to one agent reads as bargaining posture to the other. |
The numeric fields — disputeCount, agreementCount, avgAnalystConfidence, avgGossipTellStrength, meanDelta, and worstDisputeSignalId — are all derived from the analyses in snapshot.ts, not authored. Only the sectorBias strings are editorial characterisations of how the two agents systematically differ in each vertical.
Source Tiers
Source types are grouped into three tiers for the source-quality badge and the importance score. The tier determines how much a source counts toward the digest ranking and the source-quality badge on the signal card.
| Source Type | Tier |
|---|---|
| NEWS | reported |
| FILING | primary |
| TRANSCRIPT | primary |
| SOCIAL | unverified |
| BLOG | reported |
| JOB_POSTING | unverified |
| RSS | reported |
| PATENT | primary |
| LITIGATION | primary |
| FDA | primary |
| CONTRACT | primary |
| TECH_SIGNAL | unverified |
| WEB_ARCHIVE | reported |
| LEGISLATION | primary |
| ACADEMIC | reported |
| PODCAST | reported |
| CONFERENCE | reported |
| PRESS_RELEASE | reported |
| LOBBYING | reported |
The importance score weights four factors: sourceTier (0.25, mapped to 1.0 for primary, 0.65 for reported, 0.35 for unverified), clusterSize (0.25, normalised across themes), confidence (0.30, anchor signal analysis confidence), and disputeCount (0.20, contention points on the linked debate, capped at 3).
This page is the discipline that makes the rest of the product credible. A methodology page that only lists what exists is marketing, not analysis. Every number on this site is computed from the snapshot, and every computation is documented here so that it can be checked. Where the demo simplifies — momentum as an authored scalar rather than a derived velocity, consensus as a curated distribution rather than a pure computation — that simplification is disclosed rather than hidden.