This is the conceptual reference for what's "under the hood" of the workspace. Knowing what Compass produces makes every PI view easier to read and helps you spot when something looks off.
You won’t see the prompts or the model internals here — Compass uses advanced language models for narrative analysis and a dedicated sentiment engine for tone — but every signal it produces is summarised below.
The four families of signals
Compass generates four broad families of analysis, each feeding different parts of the workspace:
Review extraction — Per-review analysis of every completed review. Drives the Reviews tab and feeds Trajectory.
Trajectory synthesis — Cross-cycle synthesis per employee. Drives the Employee Trajectory tab.
Objective quality scoring — SMART analysis per objective. Drives the Objectives Quality & SMART page.
Fast feedback NLP — Sentiment and topic analysis per feedback entry. Drives the Fast Feedback Sentiment and Research pages.
1. Review extraction signals
When a review is marked complete and its PDF is generated, Compass reads the full review and extracts a structured set of signals. These all live in a single per-review record.
Performance level: a six-value classification:
Exceptional: clear standout, well above peersStrong: consistently delivers above expectationSolid: meets the bar, dependableConcerning: falling short in important areasCritical: serious performance issuesInsufficient data: not enough content to judge
Overall quality score: a 0–1 decimal capturing how high-quality the review content is overall (clarity, evidence, depth).
Reviewer quality score: a 0–1 decimal capturing how well the reviewer wrote the review (structure, specificity, evidence) — independent of the reviewee's performance.
Sentiment score: a −1 to +1 decimal capturing the tone of the review's narrative (negative = critical, positive = supportive).
Self-awareness gap: where self and manager ratings diverge:
Aligned: self and manager broadly agreeOver-rater: self rates themself higher than the manager doesUnder-rater: self rates themself lower than the manager doesMixed: inconsistent patternInsufficient data: not enough to judge
Primary development area: a short free-text label of the main thing this person needs to work on.
Trajectory relevance: how "weighty" this review is for cross-cycle analysis:
Primary: a formal full review; counts heavily in trajectory analysisSupporting: a lighter review; counts but with less weightMinimal: insufficient to feed trajectory at all
Lists of themes: Compass also extracts several narrative-driven lists per review:
Contribution descriptors (how people are talked about)
Key themes
Skills signals
Strengths
Growth opportunities
Risk indicators (retention, burnout, disengagement, etc.)
Collaboration evidence
Culture / people-investment evidence
These lists are aggregated up across the cohort to populate cards like "Top strengths", "Top development areas", and "Risk by department".
2. Trajectory synthesis signals
Once an employee has at least two primary-classified reviews across cycles, Compass synthesises their trajectory. These signals live per-employee and update whenever new extractions arrive.
Trajectory: the headline direction, a pure read of performance-level movement across cycles:
Accelerating: improving across cyclesSteady: holding the same level (the most common direction — it isn’t negative)Declining: performance trending down
The "stuck despite coaching" pattern isn't a trajectory value — it's captured by the Stalled archetype below.
Archetype: a richer pattern label that combines trajectory with current performance level:
Rising Star: accelerating at strong or exceptionalEmerging: accelerating from a lower level; the upward arc is the signalPillar: steady at strong or exceptional; the consistent high performerSolid Core: steady at solid; the reliable middleCooling: declining but still strong or exceptional; the early-warning archetypeStalled: steady with persistent, closeable gaps that aren't closing despite developmentAt Risk: declining with worsening risk signals, or stuck at a concerning levelVolatile: alternating direction across three or more cycles
Progression readiness: promotion / progression signal:
Ready Now: could be promoted todayDeveloping: making clear progress toward readinessConsolidating: recently promoted or stabilisingNot Ready: not on a clear progression pathInsufficient data: not enough substantive review data to make the call confidently
Risk trajectory: direction of risk signals over time: increasing, stable, decreasing, new, or none.
Persistent gaps: development gaps that have shown up cycle after cycle without resolution.
Resolved gaps: gaps that appeared in a prior cycle and don't appear now (development is working).
New gaps: gaps that just appeared in the latest cycle.
Persistent strengths: strengths that show up across multiple cycles.
Arc summary: a short narrative line summarising the person's trajectory in plain language.
Confidence: a 0–1 decimal Compass attaches to the trajectory assessment.
3. Objective quality & SMART signals
For every active or completed objective, Compass scores it on SMART quality.
Overall quality score: a 1–10 score (the average of the five SMART dimension scores), plus an overall rating of Excellent, Good, Fair, or Poor.
SMART dimensions: five individual sub-scores, each 1–10 with a specific issue noted when the dimension falls short:
Specific: is it clear and unambiguous?
Measurable: can success be quantified?
Achievable: is it realistic given context?
Relevant: does it tie to a strategic priority?
Time-bound: is there a clear deadline?
Admin flags: specific issues Compass detected (e.g., "Missing deadline", "No success metrics", "Vague language").
Coaching focus: a one-line recommendation for what to improve.
Compass also generates objective embeddings and extracts clients, matters, themes, and keywords for each objective — these are used behind the scenes for alignment and matching, not exposed directly on the PI views.
4. Fast feedback NLP signals
For every fast feedback entry submitted, Compass runs sentence-level sentiment and topic extraction.
Sentiment classification: per sentence: Positive, Neutral, Negative, or Mixed, with confidence scores for each.
Net Sentiment Index (NSI): an aggregated score that combines positive, neutral, and negative counts into a single number per segment (department / role / location). Used throughout the Sentiment view.
Aspects: topics extracted from the feedback (e.g., "communication", "delegation", "client relationships"). These power the trending topics chart and the Research view.
Qualifiers: descriptive terms paired with aspects (e.g., "clear communication", "lacking delegation"). These drive the word clouds on the Sentiment page.
Quality score and dimensions: for each feedback entry, Compass also scores how well-written the feedback is, with dimension breakdowns (similar to objective quality).
How signals refresh
Most signals are produced automatically by background jobs. Fast feedback, objective, and collaboration-points jobs are queued the moment the source item is completed. Review extraction and trajectory jobs are queued by an hourly discovery sweep that scans for completed reviews that haven't been analysed yet — the same sweep also re-queues anything the other triggers missed. You'll see jobs in flight via the queue status banner on each module's Introduction page (covered in the next article).
For Trajectory specifically, Compass requires at least two primary-classified reviews for an employee before it'll produce a trajectory record. New employees and people with only minimal reviews won't show up on Trajectory pages.
Reading attribution headers
Every AI-driven view in PI carries a small line near the top, for example:
Analysis by Compass AI · Talent Grid · 142 employees placed from 156 reviews · AI can make mistakes
That line tells you three things at a glance:
What kind of analysis is on the page (Talent Grid)
The scope of the data (142 placed, 156 reviewed)
A reminder that AI can be wrong — treat it as a starting point
Tips
The trajectory signals are the most opinionated. Compass is making judgement calls about archetype and readiness — read them as a recommendation to investigate, not a settled answer.
Sentiment can be misleading at the individual level. It's good at the cohort and segment level (across hundreds of entries the noise cancels) but unreliable for a single review or feedback item.
Quality scores are about the writing, not the person. A "poor quality" review doesn't mean a poor performer — it means a poorly written review. Don't conflate.
Look for divergence between signals. When trajectory says "declining" but the manager rating says "exceptional", that's worth investigating — could be a brilliant person on a bad year, or a manager who's overrating.