Methodology
Why Signal can be
trusted.
Signal is a continuous, anonymous early-warning system for organizational risk inside a portfolio company. This is how it works — what it treats as evidence, how it decides something is real, and where automated analysis ends and human judgement begins.
Signal detects. Beacon escalates. Bearing interprets. Beacon and Bearing are capabilities of Signal, not separate products.
The unit of evidence
What counts as an observation.
An observation is a brief, structured piece of input from a verified participant, submitted on a recurring cadence. It is deliberately small — a short, specific read on a defined aspect of how the organization is operating, not an essay, a grievance channel, or an open-ended survey.
A single observation is never treated as a finding. On its own it is noise: one person's read of one moment. Signal is built on the opposite premise — that meaning lives in the pattern across many independent observations over time, not in any one of them. Nothing an individual submits surfaces as a conclusion, and no participant's input is ever shown to the portfolio company or to other participants.
Who contributes
Participants, coverage, and protection.
Participants are selected for balanced coverage of the organization — across function, seniority, tenure, geography, and business unit. Coverage matters because a pattern that only one part of the company can see is weaker than one that independent corners describe in the same terms. The objective is a representative read, not a self-selected one.
Participants are named into function slots by someone inside the business and reached through single-purpose links, so input comes from real people in the organization rather than anonymous outsiders. The platform itself holds no identity to protect — a participant is a function label and a single-use code — so no observation can be joined back to a name in anything that surfaces.
Aggregation is also a floor, not just a mechanism. Minimum cohort requirements must be met before anything about a pattern is shown — enough distinct participants and enough distinct functions that no individual can be inferred from the result. Below that floor, nothing surfaces at all.
How Signal decides something is real
The six-dimension validation model.
Signal does not reduce a pattern to a single, unexplained score. Every pattern is scored against six distinct dimensions, and the overall confidence is composed of them — so confidence is always decomposable back into the reasons behind it. There is no opaque number.
These four ideas are kept deliberately separate: confidence (how sure we are the pattern is real), coverage sufficiency (whether enough of the organization has been heard from), materiality (whether it matters), and urgency (how quickly it may need a response). Collapsing them into one figure is exactly what Signal is designed not to do.
Recurrence
Has the same underlying condition appeared repeatedly, rather than once. A single observation is never treated as signal — recurrence is the first threshold anything must clear.
Reach
How many independent participants, and how many distinct functions, describe the condition. Breadth across people who do not coordinate is far stronger than volume from one corner.
Persistence
Is the pattern present across multiple reporting periods, not confined to a single moment. Persistence separates a structural condition from a passing reaction to one event.
Specificity
Is the pattern concrete enough to interpret and act on. Vague sentiment is held back; specific, describable conditions are what a board can reason about.
Consistency
Do separate observations describe a materially similar condition, or merely a similar mood. Independent accounts that converge on the same specifics carry the most weight.
Materiality
Could the condition affect execution, leadership stability, forecasting, governance, or enterprise value. A pattern can be real and well-corroborated yet immaterial — materiality is assessed separately.
Illustrative
How a confidence figure is composed
81/100An illustrative example, not a real finding. It shows the shape of a composition — a headline figure that breaks down into named, inspectable parts.
- Recurrence
- 22 / 25
- Reach
- 20 / 25
- Persistence
- 17 / 20
- Consistency
- 14 / 20
- Coverage sufficiency
- 8 / 10
The exact production weighting is proprietary. What is not proprietary is the principle: every score can be opened up and explained in these terms.
Beacon
Escalation: nothing surfaces until it earns attention.
Beacon is the escalation layer inside Signal. Nothing is escalated on a single reading. Two conditions must hold together: confidence must cross the program threshold — which a theme reaches only when independent participants across functions corroborate it, not when one voice repeats — and the pattern must persist across consecutive weeks rather than spike once and fade. Severity is then assigned by how far a theme has run past its threshold, and shapes how urgently the escalation is raised — but a one-week signal does not escalate however sharp it looks.
Before escalation, contradicting observations and alternative explanations are handled explicitly rather than discarded. If part of the organization describes the opposite of an emerging pattern, that contradiction is weighed into confidence, not filtered out to make the story cleaner. Plausible benign explanations are tested against the evidence before a pattern is treated as a risk.
False-positive controls guard against ordinary variance being read as a trend, and coordinated-manipulation controls guard against a small group attempting to manufacture a signal. Because escalation depends on independent corroboration across functions rather than raw volume, a coordinated push from a narrow group does not clear the reach and consistency thresholds on its own.
Bearing
Interpretation: what Signal knows, and what it does not.
Bearing is where an escalated pattern becomes board-ready guidance — drafted with AI assistance, reviewed and published by a person, never automatic. It begins by separating what Signal actually knows from what it does not — drawing a clear line between corroborated evidence and inference so the reader can see where confidence ends.
Each interpretation sets out, in order:
- The most-likely explanation given the evidence
- Plausible alternative explanations that remain open
- The business exposure if the most-likely reading is correct
- Decision options available to the board or operating team
- A plain confidence statement about the read itself
Bearing never claims that a specific executive is definitively at fault on the basis of anonymous evidence alone. Anonymous, aggregated observation can establish that a condition exists and that it is material; it cannot, by itself, assign individual culpability. Where the evidence points toward a person, that is framed as a question for the board to examine — never as a verdict Signal has reached.
Where the line sits
AI and human review.
AI does the work it is genuinely good at: finding structure in large volumes of language. It is used to cluster themes, recognize when differently-worded accounts describe the same condition, detect recurrence and contradiction across periods, summarize trends, and draft explanations for review. This is what makes continuous, organization-wide listening tractable.
AI does not independently make employment decisions or determine executive culpability.
Human review is required at every consequential point. A person reviews material and urgent escalations before they leave the system, owns all board-ready output, and handles anything touching individual personnel or sensitive legal and HR matters. AI proposes; people decide what reaches a client and what it means.
Where AI is used
- Thematic clustering of observations into candidate patterns
- Semantic similarity to link differently-worded accounts of the same condition
- Recurrence and contradiction detection across periods
- Trend summarization as a pattern strengthens or decays
- Draft explanations and suggested follow-up prompts for review
Where human review is required
- Any material or urgent escalation before it reaches a client
- All board-ready outputs and recommendations
- Sensitive legal or HR-adjacent cases
- Any interpretation touching individual personnel
- High-impact recommendations with real business consequence
Signal operates within clear limits. Anonymous, aggregated evidence can establish that a condition exists, that it is widespread, and that it persists — it cannot read intent or adjudicate individual conduct, and it is not a substitute for formal investigation where that is warranted. Ethical and legal boundaries, particularly around personnel, are treated as hard constraints on what the system will assert.
Models are recalibrated over time as understanding of what does and does not predict organizational risk improves, and data retention is governed by defined policies rather than left open-ended. The detailed security and data-handling model is documented separately.
See it in practice
Understand how Signal would read your portfolio.
We are glad to walk through the method in detail and show what an escalated pattern and a board interpretation actually look like.