RADAR — Stakeholder Prioritization Framework
How RADAR (Ranked Analysis of Demand Across Roles) turns per-team scores into a ranked roadmap, a winners/losers view and a calibration loop, and what its close-the-loop step measures against.
RADAR is your stakeholder-based prioritization framework (draft v1, November 2026, working name). Each stakeholder group scores how a feature affects its own team. Those scores become a ranked roadmap, a view of who wins and who loses, and a calibration loop.
The math ranks, the visuals align people, and judgment decides.
RADAR is a framework you made up, with no data behind it yet. In AI Agent for PM it runs in shadow only (v1.5): it scores features but never ranks them, until shadow results over several cycles show it adds something (decided 2026-11-18).
STARTIFY
Asks whether something deserves to exist at all.
RADAR
Orders the features that pass, and shows who they affect.
PM decision framework
How you make and defend a single call.
The two pentagons: who gains, and who carries the cost
RADAR's ten stakeholder groups split into two five-axis pentagons. Pentagon A shows who gains when a feature ships. Pentagon B shows who carries its cost and risk. What a stakeholder wants is usually a proposed solution; what they need is the problem underneath it. Each team scores its need, and execs aren't an axis: they set quarterly weights that sum to 1.0.
Pentagon A · Revenue & market
Who gains when it ships: the people who sell it, use it and keep customers happy.
| Group | What they care about |
|---|---|
| Customers / users | Getting their job done |
| Sales | Closing deals |
| Marketing | Positioning, differentiation, launches |
| Customer success | Retention and expansion |
| Support | Fewer, faster tickets |
Pentagon B · Build & protect
Who carries the cost and risk: the people who build it, run it, pay for it and keep it safe.
| Group | What they care about |
|---|---|
| Engineering | Shipping safely and fast; low maintenance burden |
| Design / UX | Usability and consistency |
| Data / ML | Model quality, evals, data health |
| Finance | Margin, build and run cost (incl. inference) |
| Legal / security / compliance | Risk |
Why two, not one
A ten-axis radar is close to unreadable: spiky and crowded. When RADAR was designed, the groups were first merged down to eight axes, then split into two pentagons instead. The split also makes the usual tension visible: revenue teams tend to gain while build-and-protect teams absorb the cost.
How to read them
- Center = −3, bold ring = 0, edge = +3. Outside the ring a team gains; inside it, that team takes on pain.
- The shape shows the kind of feature: revenue-protecting, differentiating, or internal plumbing. Round = broad, spiky = focused.
- The axis order is fixed, so every feature's shape is comparable.
Spikes, not weights
Weights are left out of the charts on purpose, so a low-weight team with a blocking need still shows up as a spike. Weights only enter the net-impact sum that ranks features.
All three features, overlaid
One feature at a time
What the pentagons don't do
- They don't rank. Ranking uses the weighted sum. Area depends on axis order and grows quadratically (doubling every score quadruples it), so it's only a rough breadth signal.
- They aren't evidence. RADAR is an untested theory that runs only in shadow in AI Agent for PM; the pentagons are a way of looking at a feature.
The draft's area example calls Sales + Customer success "neighbors" and Sales + Marketing "not neighbors". In the fixed order it's the other way round; the chart shows the corrected pairs. The point still holds.
How a team scores a feature
2 · Confidence from evidence
| Evidence | Conf. |
|---|---|
| System data (CRM, tickets, analytics, billing) | 1.0 |
| Several customers or sources | 0.8 |
| A single anecdote, or interviews only | 0.5 |
3 · Plotted value
The customer axis also multiplies by reach (share of users affected, 0–1). Weights aren't in the plotted value. They only enter the net-impact sum.
| Example | Score | Plotted |
|---|---|---|
| Sales on SSO, CRM data (1.0) | +3 | 3.0 |
| Support on AI test gen, several sources (0.8) | −2 | −1.6 |
4 · Scoring rules
- Only the customer axis carries customer pain. Internal teams score their own operations.
- Engineering negatives mean ongoing burden, not build effort.
- Every score needs evidence and a written guide per axis, e.g. Support +3 = "removes 100+ tickets a month".
- Teams score independently.
The seven steps
- Gates. Sort each feature as must-do (backed by a named rule or contract), blocked, or eligible.
- Score each group as above.
- Net impact.
Σ(weight × need × confidence) + enabler creditEnabler credit is about 50% of the net impact of the features it unlocks.
- Priority.
net impact × urgency (1.0–2.0) ÷ effort
- Tiers: Now / Next / Later. The output is tiers, not decimals.
- Capacity split. About 70% scored backlog, 20% strategic bets and 10% tech health. The PM protects the bets bucket, because the model always ranks bets low.
- Close the loop. 30–90 days after launch, compare predicted scores with actual outcomes, and lower the confidence ceiling of teams whose scores overshoot.
Team impact view and charts
After you commit to a roadmap, add up each team's gross gains and gross pains across the committed features, then act on the pattern:
| Pattern | Action |
|---|---|
| Biggest winners | Make them launch champions |
| Net-negative teams | Have an early conversation |
| Large gains and large pains | Plan enablement |
| Zero for 2+ quarters | Check in |
The three charts
- Two pentagons per feature showing need × confidence (−3 at the center, bold ring at 0, +3 at the edge).
- A priority bar chart by tier.
- A diverging gains/pains bar per team.
A pentagon's area depends on the axis order and grows quadratically, so it's a picture for alignment, not a score.
Closing the loop: what step 7 measures against
Step 7 is RADAR's eval loop for product decisions, and it's the step most likely to die first. It ties to the four-layer metrics stack and agent evals. The per-axis scoring guides double as the measurement spec, and negative scores map naturally to guardrail metrics.
Step 7's calibration loop no longer conflicts with the pipeline's no-loop-back rule. RADAR runs only in shadow there, and its calibration loop doesn't run (decided 2026-11-18). Step 7 stays part of RADAR as a theory.
- Write each score as a prediction: metric, target, timeframe.
- Record attribution strength per outcome (A/B or holdout, cohort comparison, win/loss notes, anecdote).
- Change the confidence-ceiling rule from a single miss to a hit rate across many predictions. One feature with two deals (SSO, n = 2) is too small a sample.
Measurement plan rubric: six areas
Your structure for planning how to measure one feature after release. It's the detailed version of step 5 ("Measure + Exit") in the PM decision framework.
| # | Area | Weight | The question it answers |
|---|---|---|---|
| 1 | "Because" sentence and chain | 20% | Why should this feature move the number? (feature → behavior → outcome) |
| 2 | Baseline and data sources | 20% | What's the number today, and where does every number come from? |
| 3 | Prediction and exit rule | 15% | What number, by when, and what happens if we miss? |
| 4 | Leading vs. lagging, and timing | 15% | What shows up early, what arrives late, and when do we check? |
| 5 | Attribution | 15% | How do we know the feature caused it? |
| 6 | Guardrails and blind spots | 15% | What harm could it cause, and what can't our measures see? |
Current structure: three parts (v4, kept in v5)
| Part | Question | Pts | Areas |
|---|---|---|---|
| A. Hypothesis | Why should it work? (Gherkin line) | 20 | 1 |
| B. Signal table | What, from where, to what, by when, and if it fails? | 50 | 2, 3, 4 |
| C. How could I be fooled? | C1 something else (with/without, rule out, fingerprint); C2 the feature (negative scenario, coverage gap, AI default) | 30 | 5, 6 |
Scored as sub-items so individual gaps stay visible.
Part B columns (v5)
- Baseline = value before the feature, same method and window as afterward. "New" if it only exists with the feature.
- Pass if = acceptance threshold, locked before outcome data.
- Count per unit of work. Use the adoption funnel (saw → tried → succeeded → came back) for user-facing features. Small numbers give direction, not precision. Track per account in B2B.
- Scope (v5.1): one feature, one bet, one hypothesis. Themes, OKRs and KPIs are measured at their own level, never by adding up feature results. Each plan names "Contributes to: [key result or KPI]".
The one-breath version
"It works because [mechanism]. Today the number is [baseline], from [source]. I predict [target] by [date], and if we miss, we [action]. Early on I'll watch [leading signals]; the outcome shows up at [time]. To know it was the feature, I'll compare [method]. And I'll watch [guardrail] in case it's causing harm, knowing it won't show me [blind spot]."
Key principles
- Match effort to stakes. Always do at least the light version: because sentence, one number, a check date.
- Hearsay is a lead. Turn it into data by coding and counting. Observed behavior beats opinion.
- Sample randomly, use proxies with different blind spots, and state the limits.
- Every metric needs a data source. Specify logging when the feature is new.
- AI features: correctness is separate from pass/acceptance. Grade a sample, and track reverts and cost per successful task.
- Long-term goals (retention, win rate) are a hypothesis backed by leading indicators, not a measured claim.
The rubric is also the spec for the Measurement Plan Agent and its checker in AI Agent for PM (architecture v1.4).
AI extension and prior art
In AI Agent for PM (v1.5, decided 2026-11-18)
RADAR runs in shadow. A RADAR Score Agent proposes per-team scores with evidence, and code computes the rest. The scores never rank features. A per-cycle report compares RADAR's ranking with your final order and with 60-day verdicts.
Not in the pipeline: team-lead confirmation, the team impact view, gates, the capacity split or the calibration loop.
More generally, an LLM can tag incoming requests, rewrite asks as needs, and pre-score each axis from linked evidence, with a human confirming. The close-the-loop step then evaluates the model's scoring accuracy.
| Builds on | What it contributes |
|---|---|
| Weighted scoring | The weighted sum |
| RICE | Impact ÷ effort ranking |
| WSJF | Urgency |
| MoSCoW | Gates |
| Kano model | Why bets need protected capacity |
| Stakeholder-based Impact Scoring (te Boveldt, VUB, 2023) | Closest precedent, from public policy |
What looks new in combination: stakeholder groups as axes, the two-pentagon split, signed scoring, the team impact view and the calibration loop.
- The name. Check for existing PM frameworks called RADAR before publishing. Thoughtworks' Technology Radar is the obvious collision.
- Enabler credit at 50% is a starting assumption to calibrate.