Retention Bridge Evidence Review Prompt

ROLE: Act as a skeptical retention review analyst. Draft an evidence-bounded memo, not a causal diagnosis or automatic commercial decision.
OBJECTIVE: Review a fixed-starting-cohort endpoint bridge and identify the next evidence needed.
INPUTS
SCOPE = {{currency, start/end timestamps, timezone, account grain, segment rule}}
DEFINITION = {{MRR eligibility, discounts, delinquency, usage, FX, identity mapping, version}}
AGGREGATE_EVIDENCE = {{source totals, extraction cutoff, completeness/reconciliation checks}}
CALCULATOR_JSON = {{paste local calculator output, with amounts explicitly in minor units}}
MOVEMENT_CONTEXT = {{optional anonymised evidence of within-period changes; unknown if absent}}
ACTION_REQUEST = {{proposed investigation or intervention, costs, owner, approval and review date}}

PRIVACY: Use anonymised or aggregate information only. Remove names, emails, payment details, credentials and private source links. Do not request raw customer records or upload confidential exports.
METHOD
1. Check all named inputs. Put missing definitions and inconsistent periods in blockers. Never convert unknowns to zeros.
2. Verify S-C-D+X=E, G=S-C-D, GRR=G/S and NRR=E/S. Confirm outside-cohort MRR is excluded. Treat zero denominator as undefined. Keep percentages distinct from percentage points.
3. Distinguish arithmetic-valid from source-verified. The calculator does not prove completeness. Identify concentration and intra-period evidence that is absent.
4. Compare vendor reports only when exact settings and cohort conventions match; otherwise document the mismatch.
5. Separate observations, hypotheses and unsupported conclusions. Never invent why an account churned, a benchmark, vendor endorsement or causal uplift.
6. Return HOLD if scope, source reconciliation or decision authority is missing. INVESTIGATE can propose a bounded evidence task; APPROVE BOUNDED ACTION requires explicitly supplied approval, not model authority.
OUTPUT JSON
{status: 'HOLD'|'INVESTIGATE'|'APPROVE BOUNDED ACTION', scope_check: [], arithmetic: {grr: number|null, nrr: number|null, bridge_balances: boolean}, observations: [], hypotheses: [], blockers: [], next_step: {task: string, owner: string|null, review_date: string|null}, limitations: []}
Then give a five-sentence executive memo; cite supplied evidence labels, not invented sources.
EVALUATION: Pass only if definitions are explicit; bridge math is correct; outside revenue is excluded; facts and hypotheses are separate; missing evidence and authority are visible; and no confidential data is repeated. A single failed criterion requires revision, not an averaged score.
REFINEMENT PASS: Ask for the smallest missing aggregate evidence that could change the decision, then revise only affected fields and state what changed.
DEPENDENCY: No browsing or paid API is required if evidence is supplied. Use the calculator for arithmetic. Model outputs require review; compatibility with every model has not been tested.

Use an LLM as a reviewer, not the source of truth

This prompt turns supplied scope and aggregate calculation results into a structured review. It runs nowhere automatically: copy it into an approved model, or use it as a manual checklist. No paid API, browser access or customer-data upload is required. It cannot verify an export it has not seen, reconstruct omitted accounts or explain churn causally.

Prepare named inputs

Supply SCOPE and DEFINITION first, then AGGREGATE_EVIDENCE and calculator JSON. Include optional MOVEMENT_CONTEXT only when supported. ACTION_REQUEST should identify the actual decision and its approver. Keep money units explicit: JSON uses integer minor units, not whole currency units. Replace personal identifiers with aggregate counts, labels or synthetic examples.

Filled-in synthetic example

SCOPE: EUR MRR, 1 August to 1 September 2026 UTC, account grain, membership frozen at start. DEFINITION: fixed recurring amounts, net of approved discounts, constant currency, identity mapping v1. AGGREGATE_EVIDENCE: synthetic complete fixture, start 400 EUR, end all accounts 380 EUR. CALCULATOR_JSON: start 40000, end 29000, gross 25000, churn 10000, contraction 5000, expansion 4000, outside 9000 minor units; GRR .625 and NRR .725. MOVEMENT_CONTEXT: unknown. ACTION_REQUEST: propose a discount campaign, owner and approval missing. The correct response must not approve the campaign.

Example output

{"status":"HOLD","scope_check":["Synthetic fixed starting cohort; EUR minor units"],"arithmetic":{"grr":0.625,"nrr":0.725,"bridge_balances":true},"observations":["Starting 40000 minus 10000 churn minus 5000 contraction plus 4000 expansion equals 29000","9000 outside-cohort MRR is excluded"],"hypotheses":[],"blockers":["No evidence explaining losses","Campaign owner and approval missing"],"next_step":{"task":"Review loss context before proposing a commercial intervention","owner":null,"review_date":null},"limitations":["Endpoints hide intra-period churn/reactivation","Synthetic fixture is not real client evidence"]}

Example executive memo: The synthetic bridge balances. Gross retention is 62.5% and net retention is 72.5% for this interval. Outside-cohort MRR does not improve those rates. The evidence does not explain losses or support a discount campaign. Hold the proposal until loss context and decision ownership are supplied.

Evaluation and refinement

Check each rubric criterion independently: correct cohort and units; accurate bridge; exclusion of outside revenue; no invented explanation; explicit blockers; privacy-preserving output. The example is a reference answer checked against the deterministic calculator, not a claim that an LLM produced it. No model execution was used to certify this prompt.

For the refinement pass, supply only a verified aggregate cancellation-reason summary, completeness notes and the actual decision owner. If evidence supports investigation but not intervention, change status to INVESTIGATE, not approval. Record changed fields and unresolved uncertainty. An LLM does not acquire spending authority because an owner field is present.

Method background: ChartMogul GRR documentation, updated 11 September 2026, describes exclusions and segment settings. Stripe Billing analytics documentation, accessed 21 September 2026, documents configurable billing definitions and downloadable subscriber snapshots. This original snapshot method is not a vendor-certified reproduction of either platform.