Imposter-signal detection and mitigation¶

How the bench detects when a user is undervaluing themselves and adjusts curriculum + framing in response. Filed in response to operator observation 2026-05-18.

Problem¶

The bench's audience — research-engineer interview prep, mid-to-senior level — is exactly the population where imposter syndrome is most common. Accomplished people with strong credentials repeatedly under-rate their own competence, then push back the daily drill onto problems too hard for the morale-building they need.

Two observable signal patterns:

High "I'm not sure" rate on the onboarding quiz — the user defers on questions they could probably answer if they pushed.
Wide credentials-vs-self-assessment gap — the LLM-derived credential analysis (from their CV / transcript / diploma) implies a score profile noticeably stronger than the self-assessment they just filled out.

Without intervention, the picker hands them medium-to-hard compose problems on day one, they fail or stall, the imposter feeling calcifies. The intervention: detect the signal, surface evidence, ease the early curriculum, and accelerate as the user demonstrates capability.

What the science actually says¶

The relevant body of work is decades old and well-replicated. Highlights:

Clance & Imes (1978) identified the "impostor phenomenon" — accomplished people experiencing persistent self-doubt despite objective success. Their original intervention recommendation: cognitive reframing + externalize the inner critic.
Bandura (1977, 1997) — self-efficacy is built through mastery experiences. Early successes calibrate "I can do this"; early failures calcify "I can't." Mastery-learning literature (Bloom 1968) shows ≥80% early success rates produce durable confidence; ≤50% produces learned helplessness.
Dweck (2006) — growth mindset reduces imposter feelings by reframing difficulty as growth signal rather than evidence of incompetence. The "yet" framing ("I don't know transformers yet") is empirically supported.
Kross & Ayduk (2017) — self-distancing (third-person inner speech) reduces affective intensity around self-criticism.
Moore & Healy (2008) — underconfident people specifically benefit from calibration training: predict outcome → observe actual → recalibrate. Repeated cycles narrow the prediction-vs-outcome gap.
Gollwitzer (1999) — implementation intentions ("if X, then Y") bridge the intention-action gap. For imposter syndrome: "if I feel like I don't belong, then I'll review my credentials and log one win."
Cokley et al. (2013–2017) — mentor relationships and identity-affirming experiences mitigate imposter feelings, especially for underrepresented populations. The bench's mentor-voice system is a partial substitute.

The bench's interventions below all map to one or more of these mechanisms. We are not building a clinical tool; we're applying well-replicated learning-science to curriculum sequencing and framing.

What the bench will do¶

1. Detect — `app/services/imposter.py`¶

Pure-functional detector. No side effects, no DB access. Inputs:

quiz_answers: list[Literal["yes", "no", "not_sure"]] — onboarding self-assessment responses
assessment_scores: dict[str, float] — self-rated 0..1 per axis (from the quiz)
credential_scores: dict[str, float] | None — LLM-derived axis profile from credentials (may be None if no upload yet)

Outputs:

@dataclass
class ImposterSignal:
    state: Literal["none", "mild", "strong"]
    not_sure_rate: float          # fraction of quiz answers that were "not_sure"
    credential_gap: float | None  # mean(credential - self) where credential exists; None if no cred
    triggering_axes: list[str]    # axes where the gap exceeds the per-axis threshold
    detected_at: datetime

Detection thresholds (tunable via config.toml):

not_sure_rate >= 0.5 → mild signal
not_sure_rate >= 0.75 → strong signal
credential_gap >= 0.20 (i.e., credentials suggest the user is 20+ points stronger than they self-rated) → mild
credential_gap >= 0.35 → strong
Either signal alone is enough; both together strengthens but doesn't multiply

The state is persisted to user_settings.imposter_signal so downstream services can read it without recomputing. It auto-decays when (a) the gap narrows to below the mild threshold OR (b) the user has completed N mastery experiences (see §3).

2. Acknowledge — onboarding follow-up screen¶

When detection fires at the end of onboarding (the user has just rated themselves), instead of jumping straight to Today, intercept with a single screen:

Header: "Your credentials say something different" (if credential-gap fires) or "It's OK to be unsure here" (if only the not-sure rate fires).
Body (concrete, evidence-based, NOT preachy):
The specific axes where the gap is widest, with the credential-derived score next to the self-rated score
One sentence framing it as common ("This pattern is the most common shape we see in senior-engineer prep" — true)
One sentence framing the bench's response ("We're going to start you on warmups for the first week, then accelerate as the data comes in")
Action: continue to Today (no opt-out; the framing IS the intervention, not a wall to climb).

Copy is in plain English. Avoid the words "imposter syndrome" — the framing matters; clinical labels reduce agency. The intervention is "we noticed a pattern in your data; here's how the bench will adapt," not "we've diagnosed you."

3. Adjust curriculum — scheduler bias¶

When imposter_signal.state != "none", the daily-flow picker biases toward:

Warmups first — exhaust the warmup queue before any compose problem
Easier compose problems — when warmups are exhausted, prefer difficulty: easy over medium / hard
Mastery framing — the home page's "Up next" card shows "Today's win" instead of "Today's challenge" while the signal is active
Skip the readiness scoring against the active JD — the gap visualization can amplify the imposter feeling; defer it for the first week

This continues for a mastery window: until the user completes N (default: 5) practice attempts where solved_without_lookup=true AND confidence_1_to_5 >= 4. After that, the signal auto-decays to "none" and the normal scheduler takes over.

If the user is breezing through warmups (>80% solved-without-lookup, high confidence), the signal can decay earlier — the bench notices and stops treating them gently. The "accelerate" half of the operator's ask.

4. Calibration ritual — weekly Reflect¶

Adds one optional Reflect prompt while the signal is active:

"Pick one problem from this past week. Before the answer was revealed, what did you predict — would you solve it, and how long would it take? What actually happened? What's the gap?"

This is the Moore & Healy calibration loop in narrative form. Over time, the prediction-vs-outcome data trains the user's self-rating to be more accurate. The prompt stays even after the imposter signal decays — calibration is a permanent skill, not a remediation.

5. Surface evidence — /aim¶

When the signal is active, the /aim "Where you stand" section shows an additional small panel:

What your credentials say Your CV shows strong signal on: {top 3 axes from credential analysis}. Self-rated lower: {axes where gap > 0.20}. Use the difference as a starting hypothesis, not a verdict.

The credentials-gap analysis already exists at /credentials/{id}/gap. We surface a condensed version on /aim, prominently, while the signal is active.

Implementation plan¶

Six sub-deliverables, sequenced so each is independently shippable:

#	Sub-issue	Complexity	Notes
A	Detection service (`imposter.py` + tests)	`complexity:m`	Pure-functional, no DB, easy to unit-test
B	Onboarding intercept screen	`complexity:m`	Template + route; needs the detection from A
C	Scheduler / daily-flow bias when signal active	`complexity:m`	Touches `daily_flow.py` and scheduler scoring
D	/aim evidence panel	`complexity:s`	Template addition; reads `imposter_signal` from settings
E	Onboarding copy edit (growth-mindset framing)	`complexity:s`	Pure copy/text
F	Reflect calibration prompt	`complexity:s`, human-only	Touches `voices.py` or `hamming.py` — CODEOWNERS

Total: 4 Charlie-eligible + 2 small (one CODEOWNERS-protected).

Order of landing¶

A lands first (everything else reads from it).
B + E can ship in parallel (no overlap).
C + D ship after A.
F is a human-author follow-up after the rest stabilize.

What this is NOT¶

Not a clinical assessment. The bench is not a therapist. The "imposter signal" is a data-shape pattern, not a diagnosis.
Not permanent gentling. The point is to ramp the user up to where their credentials suggest they should be — the signal decays as they demonstrate capability.
Not a substitute for actual learning. Warmups still teach; the credential gap still surfaces; the bench is honest about what the data shows.

Open questions¶

Tunable thresholds: the 0.5 not-sure rate and 0.20 credential gap are first-pass calibrated. Worth revisiting after 2-3 users have run through the flow.
Display name on the framing copy: should it use the operator's display_name from settings, or stay impersonal? Impersonal is safer (less "the bench knows you" creep), but personal may be more affecting. Empirical question.
Should the operator be allowed to opt out / opt back in? Imposter syndrome literature suggests not offering an opt-out is more therapeutic for the target audience — the framing is the intervention. But power-user override might be courteous. Default: no opt-out; revisit if it's a real annoyance.

References¶

Clance, P. R., & Imes, S. A. (1978). The Impostor Phenomenon in High Achieving Women. Psychotherapy: Theory, Research and Practice.
Bandura, A. (1997). Self-Efficacy: The Exercise of Control. Freeman.
Bloom, B. S. (1968). Learning for Mastery. Evaluation Comment.
Dweck, C. S. (2006). Mindset: The New Psychology of Success. Random House.
Kross, E., & Ayduk, O. (2017). Self-Distancing: Theory, Research, and Current Directions. Advances in Experimental Social Psychology.
Moore, D. A., & Healy, P. J. (2008). The Trouble With Overconfidence. Psychological Review.
Gollwitzer, P. M. (1999). Implementation Intentions: Strong Effects of Simple Plans. American Psychologist.
Cokley, K., et al. (2013). An Examination of the Impact of Minority Status Stress and Impostor Feelings on the Mental Health of Diverse Ethnic Minority College Students. Journal of Multicultural Counseling and Development.