The formula
Information gain minus burden cost. The only prompt-selection system optimizing for both signal and tolerance simultaneously.
// CORE C2D Insight-Guided Follow-Up Prompting
// Selects the question the vault most needs answered
IGFP_score(q) = EIG(q) − λ × B(q)
// where:
EIG(q) = expected information gain // how much does this answer resolve?
B(q) = burden cost // is this question answerable? will she disengage?
λ = configurable penalty weight // tunable per user engagement history
// Burden cost components:
// - question complexity (cognitive load)
// - historical response rate for this user on this domain
// - time since last prompt in same domain
// - emotional sensitivity weight of domain
select q* = argmax IGFP_score(q) // ask only q*, not all candidates
// Selects the question the vault most needs answered
IGFP_score(q) = EIG(q) − λ × B(q)
// where:
EIG(q) = expected information gain // how much does this answer resolve?
B(q) = burden cost // is this question answerable? will she disengage?
λ = configurable penalty weight // tunable per user engagement history
// Burden cost components:
// - question complexity (cognitive load)
// - historical response rate for this user on this domain
// - time since last prompt in same domain
// - emotional sensitivity weight of domain
select q* = argmax IGFP_score(q) // ask only q*, not all candidates
How it works — one loop, one question
1
Vault scans open contradictions and drift events
The C2D engine surfaces the current list of unsealed contradictions and active drift events. Each represents a gap in the vault's evidence picture.
Open items: contradiction(cognition, 0.81) · drift(B12, 1.94) · gap(pain, 3 weeks)
2
IGFP scores all candidate questions
Every question the system could ask is scored on EIG minus burden cost. A question with high information gain but low expected response accuracy is penalized. A question Joy reliably answers well scores higher.
Candidates scored: 14 questions evaluated in this loop
3
One question selected. One question asked.
Not a questionnaire. Not a checklist. The highest-scoring question only. Joy's engagement history shows she disengages after question 3 in any session — the system has known this for 6 months.
Selected: q* = "Did Mom take her B12 supplement every day this week?" · IGFP_score = 0.74
4
Answer sealed. Vault updated. Next loop recalculates.
Joy's answer is sealed as a new vault entry with its own reporter attribution and timestamp. The open drift event is updated. The next IGFP loop runs with the updated evidence picture.
Drift event B12: new evidence appended · contradiction(cognition): still open · next q* recalculated
What a questionnaire asks vs. what IGFP asks
Standard weekly caregiver questionnaire
"How would you rate Mom's memory this week?"
"Did she have any falls or near-falls?"
"Did she take all medications as prescribed?"
"How was her sleep this week?"
"Did she engage socially?"
"How was her appetite?"
"Did she seem confused or disoriented?"
"Any mood changes?"
IGFP — this week's selected question
"How would you rate Mom's memory this week?"
IGFP: 0.31 · vault already has 14 entries on cognition this week
"Did she have any falls or near-falls?"
IGFP: 0.44 · no open drift or contradiction in mobility domain
"Did Mom take her B12 supplement every day this week?"
IGFP: 0.74 · resolves open drift event · Joy answers supplement questions accurately
[Questions 4–8 not asked]
Joy answers one question. The vault learns what it needed to know. She does not disengage. Next week, a different question rises to the top.
Investor objection
"You're just doing what any good doctor does when they ask a follow-up question."
Why that misses the point
A good doctor asks follow-up questions from intuition, pattern recognition, and whatever they can remember from the last visit. They operate on 15 minutes of face time and whatever the patient chooses to report. IGFP operates on 10 years of sealed evidence, a live contradiction graph, active drift events, and a measured model of this specific caregiver's response patterns and engagement thresholds. The doctor's intuition is good. IGFP is the same intuition applied to evidence the doctor has never had access to — and optimized mathematically against a caregiver who will stop answering questions if you ask too many.
Investor objection
"Adaptive questionnaires already exist. Survey tools do this."
Why that misses the point
Adaptive questionnaires (IRT-based systems, branching surveys) optimize for measuring a latent trait — a score on a scale, a diagnosis probability. They select the next question to narrow confidence intervals on a predetermined outcome. IGFP optimizes for a completely different target: resolving specific open contradictions and drift events that exist in a sealed evidence vault. The optimization target is not a score. It is a named, timestamped, sub-scored discrepancy in a real person's longitudinal health record. No prior art teaches this architecture. The prior art search confirmed it: HIGH novelty.
1
Question asked per IGFP loop — the minimum necessary to advance the vault's evidence picture without triggering caregiver disengagement
"What if the caregiver ignores the question?"
The non-response is itself sealed as a vault entry. The IGFP loop factors in historical response rates for this user on this domain. If Joy has never answered questions about medication adherence, the burden cost rises and a different question is selected next time. The system learns from silence.
"Can't someone build adaptive prompting without the patent?"
They can build adaptive prompting. The moment they optimize against open contradiction and drift events from an immutable sealed vault — using the EIG minus burden-cost formula — they are inside Cluster 4. And Cluster 4 requires Clusters 1, 2, and 3 to exist. The entire architecture is load-bearing.
Why the four clusters cannot be separated
Remove any one cluster. The system stops working.
Cluster 1 · The Vault
Without reporter attribution and immutability, there is no evidence to score contradictions against. C2D has nothing to compare.
Cluster 2 · C2D Algorithm
Without the contradiction graph, IGFP has no open items to resolve. The optimization has no target. Drift detection has no anchor events.
Cluster 3 · Drift Detection
Without personal baseline drift events, the IGFP loop cannot ask the question that would surface a 9-year decline. The vault knows less than it could.
Cluster 4 · IGFP
Without adaptive prompting, the vault waits passively for evidence. The contradiction graph grows stale. Drift events go unresolved. The caregiver gets asked everything or nothing.
A competitor who wants to route around this patent must give up at least one of these four pillars. Any one they give up makes their system clinically inferior — and unable to do what Joy just showed you.
Four claims. One architecture.
No way around it without giving something up.
No way around it without giving something up.
CORE C2D · Patent Pending · Nubellum Research Inc.
Return to the demo