System of record and ledger
Who owns the truth, and how do systems reconcile?
Every industry has a place where the official answer lives, and it is rarely your database. Knowing who owns the truth tells you which system wins an argument, and where reconciliation will eat your team's week.
What it is
The system of record is the place whose answer counts when two systems disagree. A ledger is the append-only history of what changed and why. Most industries have several systems of record, each owned by a different party, and a lot of the work is keeping them in step.
Your product's data is usually a copy of someone else's truth, plus a reconciliation job.
Where you'll see it
| Industry | Whose truth you reconcile against |
|---|---|
| Payments | Your ledger vs the processor's settlement files vs the bank statement |
| Payroll | The payroll register vs the general ledger vs tax filings |
| Telco | Your number inventory vs the national porting database vs carrier records |
| Healthcare | The provider's chart vs the payer's claim history |
| Legal | The court docket vs the firm's matter system |
Questions to ask in week one
- For each key entity, which system's answer wins?
- Which of those systems do we own, and which belong to a counterparty or a regulator?
- How often do we reconcile, and who fixes the breaks?
- Is history append-only, or can records be edited in place?
- What does an auditor ask to see first?
The trap
Assuming your database is the source of truth because it's the one you can query. When a regulator, a network or a court holds the official record, your system is the copy, and "fixing" it on your side changes nothing.
In the Field Guides
AI agents: orchestration platforms
Several records describe what an agent did, and they disagree by design.
| Fact | System of record |
|---|---|
| What the agent changed | The system acted on (CRM, ERP, payments), stamped with the agent's identity |
| Why it did it | The platform's trace: sampled, metadata-only by default, retention-limited |
| What the model received | The lab's request logs: 0 to 30 days, none under zero retention |
| What it cost | The invoice, not the trace |
| What it remembers | Memory store, vector index, checkpoints, the lab's stored state |
A trace isn't a ledger. A ledger is complete, append-only and kept. Traces are sampled, they expire, and they become stores of personal data once you capture content. AWS says observability explains activity but isn't a billing report, and OpenAI warns that its usage and cost data may not reconcile.
Logs also become evidence. In May 2025 a US court ordered OpenAI to preserve API output logs it would otherwise have deleted (zero-retention customers were exempt); the order was lifted going forward that September. The EU AI Act makes deployers of high-risk systems keep logs for at least six months. I'd keep eval results outside any console that can close.
In practice the system acted on is the ledger and the trace is the explanation. Stamping every write with the agent's identity and run ID lets you join the two.
AI coding agents: writing, reviewing and shipping code
| Fact | System of record |
|---|---|
| What changed, who approved it, which checks passed | The repo host: commits, pull requests, reviews, checks |
| What the agent did and why | The session log: in the vendor's cloud (Copilot links each commit to one) or on the laptop (Claude Code keeps plaintext transcripts for 30 days by default) |
| What it cost | The invoice, the vendor's usage dashboard and your own telemetry, which rarely agree |
| Usage billed through a cloud (Bedrock, Vertex, Foundry) | The cloud bill; the agent vendor's analytics don't see it |
Git history and pull request records are what auditors test for change management. The session log is another ledger, one that may hold code and secrets, so who keeps it and for how long matters for audits and retention. Cursor's Origin is an agent vendor trying to own the main ledger itself.
AI evals and observability: how teams measure AI quality
There's no single record of quality. Four stores hold pieces of it:
| Record | Where it lives | The catch |
|---|---|---|
| What happened on each request | The trace store | Sampled, kept from 14 days to a few years depending on plan, content capture off by default |
| Test cases and scores | The eval platform | Consoles can close; OpenAI's hosted Evals ends November 30, 2026 |
| Prompt versions | Git, or a prompt-management tool | A tool that lets anyone edit prompts can bypass code review |
| What the provider saw | The lab's request logs | Kept for days or not at all, depending on the lab and contract |
| What the agent actually did | The system it acted on (help desk, payments, CRM) | The trace says "refund issued"; only the payments system knows |
None of these is a ledger. I'd keep each eval result with its bundle, model version, judge version, dataset version and date in storage you control, in files git can diff. The EU AI Act makes deployers of high-risk systems keep logs for at least six months, longer than some model versions live.
AI governance and model risk
The inventory is the ledger. US bank guidance wants enough information to understand model risk "at the individual and aggregate levels", and E-23 lists required fields. The fields I'd keep:
| Field | Why it matters |
|---|---|
| ID, name, owner | Someone answers for it |
| Purpose, users, decisions affected | Sets the tier and which rules apply |
| Vendor, model and version | Tells you what to re-test when a vendor changes something |
| Data used | Privacy and bias questions |
| Approval, conditions, review date | What was agreed and when it expires |
| Validation dates, findings, open exceptions | US guidance asks banks to track "recommendations, responses, and exceptions" |
| Monitoring metrics and incidents | Evidence it still works |
| Retirement date and archive location | What you'd show a court later |
Even banks keep it in modest tools: in Deloitte's 2025 survey of 87 banks in Europe, the Middle East and Africa, 24% still used Excel as their main model risk tool. The evidence itself lives in eval runs, traces, tickets and logs.
Litigation turns these records into evidence. Cigna produced about 2.1 million pages in a suit over its claim-review algorithm, and Workday was ordered to produce its federal equal-employment reports. Retention rules set the floor: six months for EU deployer logs of high-risk systems, four years for automated hiring data in California. The test I'd use: if a court asked for every input and output for one applicant from 2024, could we produce them?
Agent identity and delegated access: how AI agents act for people
No one record says "Maya asked this agent to do this, here, with this approval". The pieces live apart:
| Fact | Where it lives |
|---|---|
| Which agents exist and who owns them | The identity provider's registry (Entra, Okta, Google's Agent Registry) |
| Which third-party tokens an agent holds | A token vault (Auth0, AgentCore, Google's auth manager, Anthropic's Managed Agents) |
| Whether a web agent is verified | The gatekeeper's bot registry |
| Whether an agent may pay | The card network's directory |
| Whether a call was allowed, and what it did | The API or MCP server's own log |
The standard field for the agent is the act claim in token exchange, and the MCP spec doesn't mention it. Google and Entra each log the agent with its user or sponsor, in their own formats, and token passthrough makes downstream logs show the wrong caller. Gaps can go unannounced: Microsoft 365 Copilot didn't log which files it summarized until about August 2025, according to press reports.
NIST's concept paper asks how to bind each agent action back to a human's authorization. Until someone answers, I'd record for each action: agent ID, person ID, operator, token ID and audience, scope used, tool, a hash of the arguments, the outcome and any approval. For payments only, Google's AP2 mandates and Mastercard's Verifiable Intent come closest to a shared ledger.
B2B payments: cross-border payouts
Every fact has an owner, and the payout company owns fewer than its dashboard suggests.
| Fact | System of record |
|---|---|
| Customer's dollar balance and payout liability | The provider's ledger, over a pooled or safeguarded account |
| Local-currency balances at each partner | The partner's statement |
| FX position between quote and bulk purchase | The provider's treasury |
| Proof the money moved | SPEI's CEP, gpi's UETR status, Pix end-to-end ID, card settlement file |
| Gross to net, payslips | The payroll engine; in Mexico the CFDI is payslip and tax record |
Say a marketplace prefunds MXN 2 million at its Mexican partner on Monday and pays out MXN 1.8 million by Friday. Ledger, partner statement and SPEI receipts should all show MXN 200,000 left; when the statement shows 195,000, the gap is usually a netted partner fee or a return still in flight.
A public remittance company's filing warns that its prefunded balances at disbursement partners earn no interest and could be lost if a partner fails to pay out, because that money sits at the partner rather than in its own bank. I'd reconcile each currency every day against all three records: the partner's statement says what money exists, your ledger says whose it is, and the rail confirmation says whether it moved. Expect breaks from FX rounding, netted partner fees, peso returns against dollar payouts, and time zones.
Regulators test this: US state laws want permissible investments covering what's owed, the UK (PS25/12, since May 7, 2026) daily reconciliations, Canada's RPAA funds in trust or insured. When ledger and banks disagree, end users lose: the 2024 collapse of a banking-as-a-service middleware company (see spend management) left a gap estimated at up to $95M.
B2B payments: order to cash
Most facts here are owned by someone other than you.
| Fact | System of record |
|---|---|
| Quotes and contract terms | The CRM |
| Customer master, credit limits, orders, invoices, AR subledger, general ledger | The seller's ERP |
| Proof of delivery; deduction backup | The warehouse or transport system and carrier; the retailer's portal |
| Cash | The bank, in BAI2 or ISO 20022 camt.053 statement files |
| Card authorizations, funding, fees, chargebacks | The card processor |
| Check images and stub data | The lockbox file |
| Insured limit per customer; borrowing base | The credit insurer; the lender |
| Payments in flight, and who owns what in a pooled for-benefit-of (FBO) account | The AR platform's ledger, which must sum daily to the FBO balance |
| Remittance | Nobody |
The classic breaks: card deposits arrive net and batched against gross invoices; a lockbox deposits one total for fifty checks; an ACH return lands after the cash was applied; a payer deducts its wire fee. Book fees, returns and corrections as their own entries, never edits. When the ERP and the bank disagree, the bank wins on cash, the ERP on what's owed.
The ledger also carries the loss reserve. CECL (ASC 326, the expected-credit-loss rule) applies to trade receivables: lifetime expected loss from day one, pooled by risk, never zero on current balances, with price concessions and credit memos stripped from loss history. ASU 2025-05 (July 2025) lets any company assume current conditions persist for current receivables, and private companies may also count collections after the balance-sheet date, for years beginning after December 15, 2025.
Nobody is the system of record for remittance or for deduction reasons, so whoever captures them best ends up owning the matching and a clean loss history.
B2B payments: procure to pay
Every fact has an owner, and the buyer's ERP owns fewer of them than it seems.
| Fact | System of record |
|---|---|
| POs, receipts, the AP subledger, the vendor master, the general ledger | The buyer's ERP |
| Invoice images, workflow and approval history (audit evidence) | Often the AP automation tool |
| Cash out | The bank, in BAI2 or ISO 20022 camt.053 statement files |
| Which checks were issued | The Positive Pay issue file the buyer sends its bank |
| Virtual card payments and rebates | The card issuer |
| Funds in transit at a payment hub | The provider's ledger over a pooled for-benefit-of (FBO) account |
| Supplier finance balances | The ERP plus the funder's platform; disclosed under ASU 2022-04 |
The classic breaks:
- Tool and ERP disagree. The AP tool shows an invoice paid; the ERP still shows it open.
- Received but not invoiced. Goods arrived, the invoice hasn't, and the accrual is a month-end chore.
- Outstanding checks. The ERP says paid; the bank hasn't seen the check.
- Card timing. A virtual card was issued, not yet charged, and nobody can say where the cash sits.
As on the seller's side, the bank wins for cash and the ERP for what's owed; corrections post as journal entries, never edits. For audit the approval trail matters as much as the payment. Auditors test who approved and who released, as well as whether the money moved.
B2B payments: real-time payments and open banking
| Fact | Where it lives |
|---|---|
| Whether money moved between banks | FedNow: the banks' accounts at the Fed. RTP: a joint account at the New York Fed, split into each bank's position on The Clearing House's ledger |
| The supplier's cash | The supplier's bank statement |
| Which invoice was paid | The receivables platform and the ERP |
| Fraud reports | The scheme (FedNow, The Clearing House, Brazil's key directory) |
Under a 2022 schedule, the smallest RTP sending banks keep at least $25,000 in that joint account and the largest $10 million, and a payment goes out only if the bank's position covers it at that moment. FedNow needs none.
For the supplier's platform: on settlement, debit cash and credit the customer's receivable; if the bank holds the payment, debit "funds held by bank", credit suspense and apply nothing yet. A voluntary return is a new payment and a new entry. Nothing here can bounce by rule, so the ledger needs no "pending, could bounce" state of the kind ACH debits and checks need.
B2B payments: spend management
Every fact has an owner, and few of them are the platform.
| Fact | System of record |
|---|---|
| Authorizations, clearing records, card status | The issuer processor |
| The legal receivable; settlement with the network | The sponsor bank |
| Balances, limits, rewards; the participation in the receivable | The program manager's ledger |
| Prepaid and just-in-time funds | A pooled for-benefit-of (FBO) account at the bank, sub-ledgered by the platform |
| Purpose, approvals, receipts, coding | The expense tool |
| Books, accruals, close | The ERP |
| Taxable amounts (W-2, T4, Quebec's RL-1) | Payroll |
The classic breaks: the processor and the bank disagree on what settled; card spend is uncoded at close; unsubmitted out-of-pocket spend is accrued as a guess, or ignored. And pooled funds stop adding up. Synapse, a banking-as-a-service middleware company, failed in April 2024 with its ledger out of step with its banks'. The shortfall was estimated at up to $95M, and the CFPB later paid about $46M from its penalty fund to the people affected. Prepaid cards and spend-platform cash accounts use the same pooled structure.
The FDIC's 2024 proposal for daily reconciliation of custodial accounts has no final date ("To Be Determined"), but banks demand it by contract anyway: a law firm's August 2026 summary lists real-time, auditable ledger reconciliation among sponsor banks' standard asks. So four ledgers describe one purchase: the bank's wins for what's owed, the expense tool's for why and the ERP's for the books, and someone has to reconcile them every day.
B2B payments: stablecoins and blockchain payments
A pass-through payment touches four ledgers, and none of them is complete.
| Fact | Where it lives |
|---|---|
| Which invoice was paid | The payer's ERP |
| Whose coins these are, and the payment's state | The orchestrator's or custodian's books |
| That coins moved between two addresses | The chain: public, but only addresses and amounts |
| That pesos arrived | The supplier's bank statement |
The custodian's books can decide who owns the coins. When the crypto lender Celsius failed, a US bankruptcy court ruled in January 2023 that coins in its interest-bearing accounts belonged to the estate, because the terms of use moved title to Celsius; those customers became unsecured creditors. The contract's word on ownership counts for more than the chain.
Accounting follows the model. In a pure pass-through, the payer books accounts payable against cash and a fee, and never records a coin. A company that holds coins books an asset, and whether it's a cash equivalent depends on the FASB proposal's test (see How the rules work); Coinbase has booked USDC and PYUSD as cash equivalents since the end of 2025. Coins in transit need a clearing account until the off-ramp confirms, and some accounting integrations don't support stablecoin transactions yet (Stripe's docs offer a CSV export instead). Tokenized-deposit networks put the bank back as the single book of record, which is much of their appeal.
Card issuing: launching a card program
A cardholder's balance lives in four places, and they have to tie every day:
- The bank's core: the total in the pooled (FBO, "for the benefit of") account.
- The processor: the authorization balance, or open-to-buy, used to approve the next purchase.
- The program's ledger: the per-person sub-ledger that says who owns what.
- The network's settlement reports: what was actually cleared and owed.
The daily check is simple to say: the sum of every user's balance, plus pending items, equals the pooled account balance, give or take items in transit.
The bank's core, account by account
- Example
- A traditional bank
- Why choose it
- The bank sees everything
- How it fails
- Product changes are slow
The processor
- Example
- Marqeta's or Galileo's clients
- Why choose it
- Proven, and the bank reaches it through the processor
- How it fails
- Lock-in; migrations hurt
The program, in-house
- Example
- Chime's ChimeCore
- Why choose it
- Control and speed
- How it fails
- The bank has to be able to read it
Middleware across banks
- Example
- Synapse
- Why choose it
- Fast to launch with several banks
- How it fails
- One failure freezes users at many banks
Synapse is the case to know. In May 2024 its partner banks froze about $219 million "due to financial ledger reconciliation issues", and the bankruptcy trustee later reported that a gap between what the banks held and what Synapse's ledger said existed before the bankruptcy; estimates ran from $65 million to $95 million. The ledger sat on MongoDB and AWS, and the trustee lost access to systems when the staff who held the credentials were let go. The CFPB later paid $46.2 million from its penalty fund to the people affected.
Pass-through deposit insurance didn't help, because it covers the bank failing and Synapse wasn't a bank. Money held in an agent's name is insured as if each owner held it, provided the bank's records disclose the arrangement and each owner's share can be worked out from the bank's or the agent's records. If the ledger is wrong or unreachable, the money sits frozen at a solvent bank.
In September 2024 the FDIC proposed a rule for these custodial accounts: per-owner records in a standard file, daily reconciliation, direct access to records an outside company keeps even if it goes bankrupt, and a yearly certification by a senior officer, at about $250 million across 600 to 1,100 banks in the first year. It has no date on the regulatory agenda, and I found no final rule as of October 6, 2026. Banks ask for most of it by contract anyway; a law firm's August 2026 summary lists real-time, auditable ledger reconciliation among sponsor banks' standard demands.
My test for any program: if the ledger vendor went dark tonight, could the bank pay every cardholder tomorrow from data it already holds?
Cloud security (CNAPP): posture, workloads and permissions
There's no single system of record.
| Record | System of record |
|---|---|
| What's actually configured | The cloud's own API |
| What was meant to be configured | The infrastructure-as-code repository |
| Risk across clouds, code and runtime | The security platform's graph, minutes to a day behind |
| Who's fixing it | The ticketing system |
| What happened during an attack | The cloud's audit logs (AWS CloudTrail, Microsoft Entra, Google Cloud audit logs) |
| Every technology contract of an EU financial firm | Its DORA register of information |
| What security costs | The cloud bill, when bought natively or through the marketplace |
Drift is the gap between the repository and the cloud; staleness is the gap between the cloud and the graph. Retention is short by default: AWS Security Hub's older posture product deletes findings after 90 days active or 30 days archived, while PCI, SOC 2 and FedRAMP auditors usually want longer, so evidence gets exported to storage or a SIEM (the security team's log analysis system). And because Security Hub Extended and marketplace purchases land on the AWS bill, finance and FinOps teams see security spend alongside the CISO.
Identity and trust: KYB, KYC and ongoing monitoring
Every fact has an owner, and the platform owns fewer than its dashboard suggests.
| Fact | System of record |
|---|---|
| The company exists, and its status | The registry; a monopoly in each state or country |
| Tax ID belongs to the name | IRS or CRA; not public |
| Who owns it (US) | Nobody public: the bank's or platform's own due-diligence record |
| Who owns it (Canada) | The company's own register; for federal companies, a copy at Corporations Canada |
| Who owns it (UK) | The PSC register at Companies House |
| Who owns it (EU) | National ownership registers |
| What you checked, when, and the result | Your KYB and KYC record |
| Who reviewed an alert, and why they closed it | Your case-management system |
| Document images and biometric templates | Usually the verification vendor, on its own retention schedule |
Your KYB record is a derived record. For each field it should say where the value came from, when you pulled it and whether it matched. That provenance is what sponsor banks and auditors ask for, and when an examiner asks why you approved someone, the source and date of each verified fact is your answer.
The case-management system is the record a regulator reads: who saw what, when, and why they closed it. Keep verification records five years in the US and Canada. A biometric law like Illinois's wants templates destroyed on a schedule, so the usual answer is to keep the results and delete the templates.
For banking-as-a-service the ledger is a record of its own. When Synapse, a middleware company between fintechs and banks, collapsed in 2024, its records and the banks' didn't reconcile, and end users were left short by tens of millions of dollars.
Login and access for apps: authentication and authorization
For enterprise users the app owns fewer facts than its database suggests.
| Fact | Source of truth |
|---|---|
| Whether the user still works there | The customer's directory, fed by HR |
| Their role | The directory's groups or the app; choose deliberately |
| What they may do on each record | The app's permission store |
| Whether a session is live | The identity provider or the app's session store |
| Who did what, when | The app's audit log |
Google's Zanzibar paper treats stored relationships ("Ana is an editor of board 42") as the ledger of access, with a consistency token so a permission removed a second ago isn't missed by a stale read; Google reported over 2 trillion stored permissions and millions of checks a second.
The audit log is the record investigators read. Okta missed file downloads in its 2023 support-system breach for about 14 days because two ways of viewing a file wrote different log events. I'd treat sign-in events (success, failure, factor used, recovery, admin resets) as a record customers can read and export.
Models and inference: what AI runs on
| Fact | System of record |
|---|---|
| What one call used | The provider's usage record per request: input, cache write, cache read, output, tier, region |
| What the company spent | The provider's usage and cost API, the gateway's log and the cloud invoice, which rarely agree |
| How much headroom is left | Rate-limit headers on every response |
| Whether the provider was up | Its status page, which the provider writes and which can lag |
| What a lab has contracted to buy | Backlog at the clouds; take-or-pay schedules in filings |
On cloud marketplaces Anthropic converts usage into consumption units of $0.01 and reports them to AWS or Azure, so a discount shows up as fewer units. A private lab's "run-rate" is a recent month times twelve. Your own gateway log is the only record of errors from your side, and it's the evidence an SLA claim needs.
Observability: metrics, logs and traces
There's no single system of record.
| Record | System of record |
|---|---|
| What happened in production (what was indexed) | The observability backend, for as long as it keeps it |
| The full log history | Often an object-storage bucket the customer owns |
| The incident timeline | The incident tool |
| Who owns each service | A service catalog, if one exists |
| What you'll be billed | The vendor's usage meter |
| Security and audit evidence | The SIEM or the log archive |
The billing ledger is the one I'd watch. The vendor's meter decides the bill, and customers rarely have an independent count. Datadog bills custom metrics on the monthly average of hourly counts, and hosts on the 99th percentile of hourly counts, so the top 1% of hours is forgiven and a week-long spike is billed. A pipeline gives the customer its own meter of what was sent, which helps in a billing dispute.
For security and compliance teams, logs are the record: PCI requires a year of audit logs, SOC 2 auditors want evidence, and a breach investigation depends on what was kept. For the reliability side the same logs are disposable signal. The two retention philosophies collide whenever one platform serves both teams.
Phone numbers, messaging and voice
Your inventory database is a copy. Nearly every fact a customer cares about is owned by someone else.
| Fact | System of record | Run by |
|---|---|---|
| Which carrier holds a number block | LERG (the national routing guide) | iconectiv |
| Where a ported number routes | NPAC | iconectiv, porting administrator since May 2018 |
| Number assignments, reassigned numbers | NANPA, Pooling Administrator, Reassigned Numbers Database | SomosGov (current contract option runs to Nov 30, 2026) |
| Toll-free ownership and text-enablement | TFNRegistry, TSS Registry | Somos |
| Text-enablement of hosted long codes | NetNumber's registry (per providers) | NetNumber |
| 10DLC brands, campaigns, vet scores | TCR, plus each carrier's provisioning | TCR, DCAs, carriers |
| Short code leases | US Short Code Registry; txt.ca in Canada | CTIA; Canadian Telecommunications Association |
| Provider robocall certifications | Robocall Mitigation Database (RMD) | FCC |
| 911 caller location | ALI or, in NG911, LIS (the databases 911 centers query) | 911 service providers |
| WhatsApp assets, limits, charges | WhatsApp Manager, webhooks, pricing analytics | Meta |
| RCS agents, launch state, billable events | Google's console and daily billing reports | |
| Alpha senders abroad | India's DLT ledgers, Singapore's SMS Sender ID Registry (SSIR), ACMA's register, CNMC's alias registry | Operators or regulators |
The one record that's yours is consent. It's the sender's own system of record, and in a lawsuit or a carrier audit it's the only defence you have. CTIA's short code handbook expects consent records kept six months after an opt-out. The reassigned-number safe harbor only protects callers who can prove they checked the Reassigned Numbers Database, so keep those query logs too.
Voice AI: real-time voice agents
Four records describe one call, and they disagree by design.
| Fact | System of record |
|---|---|
| Minutes billed for the call | The carrier's and communications platform's call detail records |
| What the agent did, turn by turn | The voice platform's call log |
| What was said | The transcript and the recording |
| What changed in the business | The tool-call log in the business's own systems |
| Whether the caller agreed to be called | The deployer's consent records |
Two details matter. In a speech-to-speech model the transcript is a side product of a separate recognizer, so an audit or eval built on transcripts can miss what the model actually "heard". And after an interruption, what the model thinks it said and what the caller heard diverge unless the history was cut back.
Consent records are the defence in every TCPA case, and the platforms say plainly that they don't keep them for you: Retell's terms say it "does not obtain consent on behalf of Customer". Billing is its own reconciliation job, since five vendors can meter one call in carrier minutes, stream time, audio tokens, characters and voice minutes billed by the second.
Field Guides are learning notes, not legal or compliance advice. Rules and fees change; check the cited primary sources before you act on anything here.