DynamoDB

The user’s float/loan configuration, versioned by created_on. Fields: user_id, is_float_enabled, is_loan_enabled, cfi_enabled, floats (array of {id, amount, is_enabled} — amounts in dollars), loans (array of {id, amount_cents, is_enabled} — cents), created_on, reason, notes, last_cfi_rule_name, 200_test_available (true on every saved version), 200_test_offered ($200 qualification).

Semantics, defaults, and fees: Float Profiles.

One row per rule per user (the SK is the stored rule name, so a re-run overwrites). Fields: rule_name, user_id, loan (result encoding below), error, features (map of inputs and intermediates the rule saw), updated_date, ttl.

The aggregated decision for one rule run. Fields: result_id (<unix>_<uuid>), user_id, item_id, account_id, float_results and loan_results (Evaluation, below), cfi_state (CFIData snapshot), created_date, ttl, the five top-level early_advance_* attributes, and early_advance_experiment (below).

Evaluation (float_results/loan_results): approved, amount, deciding_rulebook, status (statuses), errors, results (RulebookResult array), result_id.

RulebookResult has no dynamodbav tags — items store Go field names (RulebookID, Result, Status, ApplyTo, Priority, Features, Approved, IsApplicable), while JSON responses use snake_case. Queries against stored items must use the Go names.

CFIData (cfi_state): current_limit, sub_rank, float_rank, total_float_rank, highest_float, account_balance, paid_subscription_count, is_reactivating_user, ewa_stats, is_feature_flag_enabled, two_hundred_test_available (always true), two_hundred_test_offered (mirrors FloatProfile’s 200_test_offered marker — $200 qualification).

Permanent copy of the evaluation that decided a taken float/loan, written by the float-created handler. Fields: EvaluationResult’s field set (minus ttl) plus amount, float_id, loan_id, two_hundred_test_float (true for a $200 float granted by the $200 qualification, not by an enforced Float Tiers assignment), and float_tier_assignment (copied from the decisive evaluation’s MLResult — Float Tiers). CreateHistoricalEval copies both early-advance shapes; records written before early_advance_experiment existed never age out, so history readers must handle both shapes indefinitely.

One payback-predictor-v2 prediction per evaluation. Fields: user_id, evaluation_result_id, version, prediction_id, risk_score, authoritative (flag value snapshotted at prediction time), rank, term, threshold, invoked_at, float_tier_assignment (the Float Tiers snapshot: tier, tier_amount_cents, assigned_amount_cents (cents, unlike float amounts elsewhere), approved, guardrail, guardrail_amount_cents, last_repaid_amount_cents, last_repaid_days_late, current_offer_amount_cents, authoritative_v2, guardrails_applied), reasons (the model’s reason codes, most negative first; absent on records predating them), float_id, ttl (168h). When a float is created under an authoritative prediction (or an enforced float-tier assignment), Settle records the float_id and removes the TTL so the record is kept for audit. Flow: Payback Predictor V2. The record is exposed read-only as model_result on the summary-history and historical-evaluation responses (schema: API spec, ModelResult).

Global rule-set configuration. Fields: rulebook_id, rulebook_name, type (floats or loan), rules (array of {rule, rule_arn, props, ab_group}), apply_to (basis points, 10000 = 100%), priority (higher evaluated first), superseding, last_updated, ab_group. Loaded from YAML by CI: Rulebooks.

Audit record per rulebook sync: update_user, update_time, and a full rulebooks snapshot.