How per-bullet impact scoring works

Every Experience bullet in the workspace gets a small circular chip showing its impact score from 0 to 100. Click the chip and you see a three-axis breakdown plus a one-line suggestion. This page is the methodology. It documents how each axis is graded, what calibration anchors keep scoring stable across runs, and what scoring DOES and DOESN'T claim.

The three axes

Every bullet is graded on three dimensions, each scored 0-100. The headline score is a weighted blend: 40% specificity, 30% quantification, 30% action.

Quantification (0-100)

Does the bullet name a number, percentage, scale, or measurable outcome?

• 0-39: No metric of any kind. "Worked on the checkout team" earns this.
• 40-59: A directional metric without numbers. "Improved checkout latency" — sense of magnitude implied, no number.
• 60-84: One specific number. "Cut checkout latency 40%" — number present, scope clear.
• 85-100: Multiple measured outcomes or a scaled metric ("Cut p99 checkout latency 40% across 4M weekly transactions").

Quantification is the most-easily-improvable axis. Most bullets that score below 60 here can be lifted into the 70s by adding one number — even an approximate one.

Action verb strength (0-100)

Does the bullet lead with a strong, specific verb that implies ownership?

• 0-39: Weak verbs that signal dispersed responsibility — "helped", "supported", "worked on", "responsible for", "involved in", "participated in".
• 40-59: Generic ownership verbs — "managed", "led", "created", "developed" — without a more specific alternative available.
• 60-84: Specific verbs that name the work — "redesigned", "rearchitected", "negotiated", "consolidated", "rolled out", "deprecated".
• 85-100: Distinctive verbs that signal a particular kind of ownership — "championed" (for influence work), "instrumented" (for observability), "blocked" (for security/compliance), "salvaged" (for turnaround).

We do not penalise "led" or "managed" by themselves — they're sometimes the most honest verb. The penalty is for using them when a more specific verb was available.

Specificity (0-100)

Does the bullet name a concrete system, technology, team, dollar amount, or named result?

• 0-39: Generic phrasing that could describe a thousand candidates. "Built scalable systems" — nothing about THIS candidate.
• 40-59: One specific noun, no scale. "Built the orders service" — names the system, no scope.
• 60-84: System + scale OR system + technology stack. "Built the orders service in Go on Kubernetes".
• 85-100: System + scale + outcome named uniquely. "Built the orders service that now handles 40k qps for the checkout fleet, halving the latency tail."

Specificity is what protects you from being parsed as a hundred-other-engineer template. The grade signals whether the bullet could have been written about anyone, or specifically about you.

How the headline score is computed

headline_score = 0.4 × specificity + 0.3 × quantification + 0.3 × action

The weight asymmetry is deliberate. Specificity carries the most weight because it's what distinguishes a memorable bullet from a forgettable one — quantification and action verbs can be added later, but a bullet with zero specificity needs a structural rewrite, not a polish.

The chip colours map straightforwardly:

| Score range | Colour | Meaning | |---|---|---| | 75-100 | Green | Strong. Recruiter-readable, distinctive, measured. | | 50-74 | Amber | Okay. Either weak verb, missing metric, or generic phrasing. The suggestion line tells you which. | | 0-49 | Red | Needs work. Usually missing both metric AND specificity. |

How we keep scoring stable across runs

A scoring system that gave a bullet 67 today and 81 tomorrow would be useless. Three mechanisms hold scores stable:

1. Low temperature. The Gemini call runs at temperature: 0.1 — low enough that the model is essentially deterministic on the same input within a session.
2. Calibration anchors in the prompt. The system prompt includes explicit anchor examples — the 85-100, 60-84, 40-59, and 0-39 bands each have named example shapes. The model is told to align its scoring to those bands, not to free-rank.
3. Content-keyed caching. We hash the bullet + the job description and cache the resulting score across users. The first user to score a given (bullet, JD) pair pays the Gemini cost; everyone else reads the cached score directly. This means scoring is not just stable across runs for one user — it's literally identical for any user with the same input.

What scoring does NOT claim

The same way the heatmap is a simulation rather than a measurement, the impact score is a rubric grader, not a hiring oracle. Specifically:

• It does not predict whether the bullet will get you an interview. That's a function of the whole CV, the role, the company, the recruiter, and ten other things the scorer can't see.
• It does not understand context for senior roles. A staff engineer's "led the platform redesign" can be a 95-impact bullet if the platform is huge and the redesign was their call. The scorer reads the surface; it can't always read the seniority context.
• It does not catch domain falsehoods. If a bullet says "trained 12B parameter model" and you actually trained a 1.2B parameter model, the scorer doesn't know.

The scorer's job is to nudge you toward bullets that are harder to write but easier to read. That's an honest scope, and it's what the chip is for.

When to ignore the chip

If a bullet scores red but you have a specific reason to keep it — it's the bullet a particular hiring manager once told you they remembered, or it's the only bullet that surfaces a specific certification — keep it. The scorer is advisory, never prescriptive.

If a bullet scores green but reads stiff, rewrite it anyway. The model rewards measured specificity; it doesn't reward warmth.

The score is a tool, not a verdict.