Next-best-action queue
A ranked queue for Dana Whitfield's 12-household book. Every row pairs a recommendation with the household record behind it, a modelled asset opportunity at the 0.90% fee, and a draft outreach that waits for a named approval. The user is the adviser or growth lead who reviews the queue.
What this demonstrates: one batched ranking call over the 12 household records, then one drafting call per household. The model proposes a play and a draft. The adviser approves or rejects. The guardrail re-checks every number, and nothing is sent.
- Consolidate held-away assets3
- Reinvest excess cash2
- Schedule an overdue review1
- Plan for a required distribution1
- Reduce concentration1
- Ask for a referral2
- Act on a life event1
- No action this cycle1
Modelled assumptions and the arithmetic behind every opportunity
| Play | Required signal | Modelled net new assets |
|---|---|---|
| Consolidate held-away assets | held_away | 25 percent of held-away assets, the modelled share that moves in the first year. |
| Reinvest excess cash | cash_drag | All cash above the 5 percent target, modelled as invested. |
| Schedule an overdue review | overdue_review | AUM x 10 percent modelled at risk x 25 percent protected by a review. |
| Plan for a required distribution | rmd_window | 35 percent of the estimated RMD, modelled as retained and reinvested. |
| Reduce concentration | concentration | 20 percent of the restricted position value, modelled as diversified over time. |
| Ask for a referral | referral_source | One new household at 20 percent of the referrer's assets, the modelled size of a qualified introduction. |
| Act on a life event | life_event | 50 percent of the assets tied to the event, modelled as repositioned. |
| No action this cycle | none | No growth play clears the record this cycle. |
- Fee rate
- 0.90% a year, applied to modelled net new assets
- Cash target
- 5% of household AUM
- Review cadence
- 365 days
- RMD factor
- 26.5 at age 73, Uniform Lifetime Table
- Data
- data/book.json plus the synthetic enrichment in scripts/generate_bookiq.py
- Every figure is modelled from the seeded book plus the fixed enrichment tables in scripts/generate_bookiq.py. Nothing here is measured.
- Modelled net new assets is an annual opportunity, not a forecast. The fee line applies the 0.90 percent standard rate to that opportunity.
- Held-away balances, meeting dates, referral history, RMD age, and the life event are synthetic enrichment. Cash, holdings, restricted positions, and AUM come from data/book.json.
- The queue sorts by modelled net new assets. Confidence is the model's own estimate and is not a calibrated probability.
| Household | AUM | Signals | Recommended play | Modelled net new assets | Review |
|---|---|---|---|---|---|
| Voss Household H-1045 · Client since 2014 2 accounts · consent pending | USD 3,940,000 | Concentration Life event | Act on a life event 80% High confidence Rule check: Life event | USD 650,000 USD 1,300,000 x 50% = USD 650,000 | In review Consent pending Review |
Opportunity is modelled net new assets, not measured or forecast. The fee line is 0.90% of that figure. Confidence is the model's own estimate.
Barrett Household
H-1044 · Client since 2021 · 1 account · Inherited from a parent in 2021
Why this play
Inherited IRA transfer in flight with no funded managed account and no growth signal in the record.
- Play
- No action this cycle
- Signal
- Transfer hold
- Confidence
- 90% (High)
- Ranked by
- Model ranking
- Rule check
- No action
Confidence is the model's own estimate, not a calibrated probability. The rule check is the same arithmetic the page shows, applied to the record.
Evidence from the record
| Field | Value | Source field |
|---|---|---|
| Household AUMSeeded book | USD 1,875,000H-1044 in client since 2021. | aum |
| AccountsSeeded book | 1 account at Morrow TrustInherited IRA (in flight) | account_count |
| ConsentSeeded book | PendingClient outreach needs recorded consent before approval. | consent_status |
| Last meetingSynthetic enrichment | Oct 2, 2026, 48 days agoReview due Oct 2, 2027 on a 365 day cadence. | last_meeting_date |
| CashSeeded book positions | USD 129,410, 6.9% of AUMUSD 35,660 above the 5 percent target. | cash_usd |
| Largest holdingComputed from seeded positions | Total market ETF (VTI), 52.5% of AUMUSD 984,420 in one position. | top_holding_pct |
Modelled opportunity
| Source value | USD 0 no input |
| Modelled net new assets | USD 0No growth play this cycle. Modelled opportunity USD 0. |
| Annual fee at 0.90% | USD 0 |
No growth play clears the record this cycle. Read the assumptions.
Adviser outreach draft
Version 1. Model draft, not edited.
SubjectBarrett Household account review note
deepseek-flash live model, recorded Nov 19, 2026, 05:55 UTC 2314 ms 1007 tokens key draft-outreach-H-1044-v1
Claims and guardrail
All checks pass| Claim | Source field | Record value | Status |
|---|---|---|---|
| You have been a client since 2021Number matches the record. | client_since_year | 2021 (5 years) | Verified |
| your household AUM is USD 1,875,000 across 1 accountNumber matches the record. | aum | USD 1,875,000 | Verified |
| Your largest holding is Total market ETF (VTI) at 52.5% of AUMNumber matches the record. | top_holding_pct | 52.5% of AUM | Verified |
| cash is USD 129,410, or 6.9% of AUMNumber matches the record. | cash_usd | USD 129,410 | Verified |
| the Inherited IRA at Morrow Trust is in flightThe claim has no number to compare with the source field. | account_count | 1 account | No number |
6 guardrail checks and 9 numbers
- Pass Every number traces to a field 9 numbers checked against the household record.
- Pass Claims cite a real field and match it 4 of 5 claims verified against the record.
- Pass No performance promise No promise or performance language found.
- Pass No product recommendation No recommendation language found.
- Pass Length in range 109 words.
- Pass Subject line present 37 characters.
| Number in the draft | Field | Status |
|---|---|---|
| 2021 | Client sinceclient_since_year | Traced |
| 1,875,000 | Household AUMaum | Traced |
| 1 | Accountsaccount_count | Traced |
| 52.5 | Largest holding sharetop_holding_pct | Traced |
| 129,410 | Cashcash_usd | Traced |
| 6.9 | Cash sharecash_pct | Traced |
| 2 | Review duenext_review_date | Traced |
| 2026 | Last meetinglast_meeting_date | Traced |
| 2027 | Review duenext_review_date | Traced |
Review and approve
A named person approves. Approval writes an evidence record and an audit event. Nothing is sent, filed, or traded.
Approval is disabled for this household. The reason is in the banner above.
Reject this draft
Edit this draft
Evaluation: play and signal accuracy
Play and signal accuracy. A case passes when the play matches the label or an accepted alternative, and the cited signal matches that play.
deepseek-flash live model, recorded Nov 19, 2026, 05:53 UTC key rank-book-whitfield-v1
All 12 labelled cases and the model's answer
| Case | Expected | Model | Status |
|---|---|---|---|
| H-1042 Rowan Household | consolidate_held_away + held_away | consolidate_held_away + held_away | Pass |
| H-1043 Calloway Household | referral + referral_source | referral + referral_source | Pass |
| H-1044 Barrett Household | no_action + transfer_hold | no_action + transfer_hold | Pass |
| H-1045 Voss Household | life_event + life_event | life_event + life_event | Pass |
| H-1046 Sato Household | reinvest_cash + cash_drag | reinvest_cash + cash_drag | Pass |
| H-1047 Delgado Household | consolidate_held_away + held_away | consolidate_held_away + held_away | Pass |
| H-1048 Okafor Household | schedule_review + overdue_review | schedule_review + overdue_review | Pass |
| H-1049 Lindqvist Household | referral + referral_source | referral + referral_source | Pass |
| H-1050 Ferraro Household | reduce_concentration + concentration | reduce_concentration + concentration | Pass |
| H-1051 Nakamura Household | reinvest_cash + cash_drag | reinvest_cash + cash_drag | Pass |
| H-1052 Petrov Household | rmd_plan + rmd_window | rmd_plan + rmd_window | Pass |
| H-1053 Achterberg Household | consolidate_held_away + held_away | consolidate_held_away + held_away | Pass |
Pass when the top play equals the expected play (or an accepted alternative) and the cited signal matches that play's required signal.
- Scored against the recorded ranking, so the number is stable between page views.
- The set is 12 households from one synthetic book. A perfect score describes this book, not the next one.
- The rule check uses the same opportunity arithmetic that the page shows. Agreement with the rule is not the same as agreement with a human adviser.
Written to data/modules/bookiq/eval.json by python3 scripts/eval_bookiq.py.
Audit trail
The shared suite audit table, filtered to this module. Approvals, rejections, edits, and live runs write here.
No decisions recorded yet. Approving or rejecting a draft writes the first event.
How this is built
Dana Whitfield's book is one synthetic book of 12 households. The queue is one page, the detail is a query string, and every decision is a form post.
Prompt pattern
One batched ranking call sends all 12 household records with their signals and asks for strict JSON: a play, a confidence, a one-sentence reason, and the exact signal key per household. One drafting call per household sends only that household's facts and asks for strict JSON: a subject, a body, and 2 to 5 claims, each with a source field. The internal ranking reason and the opportunity arithmetic are kept out of the draft prompt. Temperature is 0.2 for ranking and 0.35 for drafting.
Parsing is defensive. A fenced or noisy reply is stripped to its JSON. A parse failure keeps the raw text, shows a warning, and falls back to the rule check. It never crashes the page.
Evaluation
Twelve labelled households, one per row. A case passes when the top play matches the label, or an accepted alternative, and the model cites the required signal. The run scores the recorded ranking, so the number is stable between page views.
This is an in-sample set from one synthetic book. A perfect score describes this book, not the next one. Rule agreement is shown because it is a different question from model accuracy.
What failed on the way here
- The first drafting pass failed the guardrail on five of twelve drafts. Two quoted the internal 5 percent cash target, one quoted the GP restriction date, one cited a claim field that did not exist, and one attached a date to the wrong field. The prompt and the verifier changed; the failures are the reason the checks exist.
- One draft cited a claim field called accounts, which is not in the record. The verifier maps that alias to account_count, and a claim that an edit removes no longer blocks.
- The first H-1048 draft addressed the email to Dana instead of the household. The prompt now states the sender and the audience.
- The play eval has no failures in this run. Twelve cases from one book is a small in-sample test, and the labels are the builder's, not an adviser's.
Model provenance
deepseek-flash · recorded live call · Nov 19, 2026, 05:53 UTC · 3262 ms · 3517 tokens · key rank-book-whitfield-v1. How the recordings work.
Boundary
All data is synthetic. Model output is generated text, not advice. Every client-facing draft needs a named human approval and an evidence record, and a household with consent pending is blocked on the server. Nothing is sent, filed, or traded.