Beyond Ledgers: Financial Obligations as Convergent State Machines

Convergent Obligation Systems

Financial systems frequently operate in environments where the complete sequence of payment events is not observable. Payment transactions may be delayed, partially reported, or entirely absent when systems rely on external servicing platforms. In these environments, the internal representation of financial obligations must remain consistent even when the underlying event stream is incomplete.

Traditional financial system architectures assume that accurate system state requires a complete event history. Under this assumption, the absence of a payment event prevents the system from determining whether an obligation has been satisfied. However, many operational systems receive only indirect financial signals, such as balance snapshots or aggregated adjustments.

The model presented here enables deterministic financial state computation under such conditions by treating the financial obligation itself as the primary state entity. Payment events and balance observations are interpreted as signals that update the obligation state through deterministic transition rules.

This piece formalises the concept of convergent obligation systems, in which the internal state of the obligation converges toward the correct financial state as additional observations are processed, and states the conditions under which it does not.

1. Obligation State as the Primary System Variable

Let the financial obligation between two counterparties be represented by the system state:

$S_{t} = (A_{t}, I_{t}, B_{t}, M_{t}, F_{t})$

Where:

$A_{t}$ is the arrangement state at time $t$
$I_{t}$ is the set of instalment states
$B_{t}$ is the outstanding balance
$M_{t}$ represents monitoring conditions
$F_{t}$ is the set of diagnostic flags

$F_{t}$ is new. Without it the model has nowhere to put an observation it cannot interpret, and the only remaining options are to guess or to discard, both of which break the guarantees below.

The state $S_{t}$ fully describes the operational condition of the obligation at time $t$ .

Financial observations are represented as $O_{t}$ , which may include payment events, balance snapshots, operational actions, and monitoring evaluations.

The system evolves according to:

$S_{t + 1} = f (S_{t}, O_{t}, c (Δ B_{t}))$

where $f$ is the deterministic transition function and $c$ the classification function defined in Section 4. The prior formulation was $S_{t + 1} = f (S_{t}, O_{t})$ , which applies observations without interpreting them. Section 4 explains why that is not merely incomplete but incorrect.

2. Determinism of the Obligation State Machine

The transition function must be deterministic. The same observation applied to the same system state always produces the same resulting state:

$f (S, O) = S^{'}$

and for any duplicate observation:

$f (f (S, O), O) = f (S, O)$

This ensures idempotent event processing, which is essential for financial systems that may receive duplicate signals or reprocessed transactions.

Idempotency is not automatic. It holds for balance snapshots, because a repeated snapshot yields $Δ B = 0$ and produces no allocation. It does not hold for discrete payment events, which must be deduplicated on identifier before reaching $f$ . An implementation that treats an observation as an increment to a running total rather than as an absolute measurement loses the property entirely.

3. State Convergence Under Partial Observability

Where the true sequence of payment events is not observed, the system receives periodic balance measurements. Let $B_{t}$ be the observed balance at time $t$ . The system computes:

$Δ B_{t} = B_{t - 1} - B_{t}$

This delta represents net financial movement between observations.

Where, and only where, the delta is classified as payment, the reconciliation algorithm allocates it sequentially across unpaid instalments in due date order:

$a l l oc a t e d_{k} = min (Δ B_{t}, u n p ai d_{k})$

After allocation:

$p ai d_{k} = p ai d_{k} + a l l oc a t e d_{k}$ $u n p ai d_{k} = u n p ai d_{k} - a l l oc a t e d_{k}$ $Δ B_{t} = Δ B_{t} - a l l oc a t e d_{k}$

Repeated until the delta is fully allocated or no unpaid instalment remains. Where delta remains after the schedule is satisfied, it is recorded as surplus state and flagged. The prior version stated that excess allocation must be applied to subsequent instalments without saying what happens when there are none, which leaves the terminating condition of the loop undefined.

The qualifying clause at the head of this section is the substantive correction. In the prior version the algorithm was applied to $Δ B_{t}$ without a sign test. Where the balance rises, $Δ B_{t}$ is negative, and $min (Δ B_{t}, u n p ai d_{k})$ returns that negative value. Allocation then decreases $p ai d_{k}$ , which violates both the allocation integrity invariant and monotonic progression declared in the same document.

An interest posting would therefore un-pay instalments the customer had satisfied. The model as previously stated did not merely omit adverse observations; it processed them incorrectly through a function that was never intended to receive them.

4. Classification of Observed Movement

Balance movement must be interpreted before it is applied.

$c (Δ B_{t}) \in {p a y m e n t, n e u t r a l, e x o g e n o u s A d j u s t m e n t, r e v er s a l, an o ma l y}$

Condition	Classification	Handling
$Δ B_{t} > 0$	payment	Allocate per Section 3
$Δ B_{t} = 0$	neutral	No change
$Δ B_{t} < 0$ , matching adjustment record	exogenousAdjustment	Accept balance, leave allocation unchanged
$Δ B_{t} < 0$ , matching reversal record	reversal	Controlled reverse allocation
$Δ B_{t} < 0$ , no matching record	anomaly	Accept balance, leave allocation unchanged, raise flag

Matching requires that the record's effective date falls in $(t - 1, t]$ and that record amounts sum to $∣Δ B_{t} ∣$ exactly. Attribution by inspection is operator judgement and would remove determinism at the point the classification exists to supply it. Reversal is never inferred from the sign of the delta; it requires an explicit record.

Anomaly is the safe default. It alters only $B_{t}$ and $F_{t}$ , so every invariant in Section 5 holds trivially under it. This is deliberate: an observation the system cannot explain must be visible, not absorbed.

5. Invariants and Guarantees

The prior version presented seven items as invariants. Four are invariants, being conditions that hold at all times. Three are guarantees, being properties of the system's behaviour over time. Conflating them matters, because an invariant can be checked on every transition and a guarantee cannot.

Invariants

Financial conservation.

$B_{0} + \sum a d j u s t m e n t s = \sum p a y m e n t s - \sum r e v er s a l s + r e mainin g B a l an ce$

The prior version gave this as $B_{0} = P + R$ . That identity fails on the first interest posting, because adjustments increase the obligation without any payment, and it fails again on the first reversal, because a reversed payment remains counted in $P$ . Both terms are required and payments must be counted net.

This is the most useful invariant in the model. Its residual is diagnostic: where the identity does not close, the gap equals the value of movement the system has observed but not classified. Section 8 demonstrates this.

Instalment allocation integrity.

$0 \leq p ai d_{k} \leq am o u n t_{k}$

Preserved on the upper bound because $a l l oc a t e d_{k}$ is capped at the shortfall, and on the lower bound because reverse allocation is capped at $p ai d_{k}$ .

Single state occupancy.

$\forall t, \exists! s \in S_{A} : A_{t} = s$

The prior version wrote this over $S$ and $s t a t e (O, t)$ , but $S_{t}$ is a tuple rather than a member of a state set, so the assertion as written did not typecheck against its own definition in Section 1. The property applies to the arrangement component $A_{t}$ over the arrangement state set $S_{A}$ , and separately to each instalment over the instalment state set.

Forward progression except under classified reversal.

$P e n d in g \to D u e \to P ai d$

$P e n d in g \to D u e \to M i sse d \to P ai d$

Backward movement occurs only where $c = r e v er s a l$ , and each instance is recorded with its justifying record. The prior version permitted reverse transitions "under explicitly defined reconciliation rules" without defining any, and separately treated Missed as forward terminal. Those two positions together make recovery from a missed instalment impossible, which removes the behaviour that repayment arrangements exist to provide.

Guarantees

Termination. Every obligation reaches a terminal state: Completed, Failed, or Cancelled. Allocation either satisfies the schedule, giving Completed, or elapsed time exceeds the configured failure threshold, giving Failed. Cancellation is available from any non-terminal state.

Termination is conditional on monitoring running. Where observation has ceased and monitoring is suspended to avoid escalating on absence of evidence, an obligation does not terminate. The prior version asserted termination safety without stating any termination condition, and this conditional is the honest form of it.

Idempotent processing.

$δ (δ (s, e), e) = δ (s, e)$

Subject to the deduplication requirement in Section 2.

Deterministic replay. Given initial state, configuration, and the ordered sequence of observations with their classifications, the full state history is reconstructable. Ordering is part of the requirement. The prior version listed inputs, parameters, and signals without ordering, and an unordered set of observations does not determine a unique history under a non-commutative transition function.

6. Convergence Property

Let $S_{t}$ be the estimated system state and $S_{t r u e}$ the true financial state. Then:

$t \to \infty lim S_{t} = S_{t r u e}$

provided that:

observations are eventually consistent, meaning the observed balance eventually reflects the true balance
adverse movement is correctly classified against records
the invariants of Section 5 are preserved

The second condition was absent from the prior statement and does not follow from the first. Classification depends on an adjustment record feed the system does not control. Where that feed is incomplete, movement classifies as anomaly, allocation ceases to advance, and the system does not converge.

What it does instead is accumulate flags. This is the intended behaviour rather than a shortcoming. A model that continued allocating under unexplained movement would diverge silently, and silent divergence in a financial system is materially worse than a visible refusal to proceed. The convergence result is real but it is conditional, and any implementation should be able to state which of its accounts currently satisfy the conditions.

7. Obligation Systems as Convergent Deterministic Processes

This model differs from traditional financial system architectures. Instead of treating the ledger as the primary representation of truth, the obligation state machine becomes the central system component. Payment events, balance snapshots, and operational actions are interpreted as signals that update obligation state through deterministic transitions.

Financial obligation systems should be designed as deterministic state machines whose internal state converges toward the correct financial condition as additional observations are processed, and which declare non-convergence rather than approximating it.

Under this architecture:

financial ledgers become derived views
balance records become signals
obligation state becomes the source of truth

8. Case Illustration: Arrears Arrangement on an External Servicing Platform

The following traces a single account through eight weeks. It is representative of a common integration: a lender holds the arrangement, a third party services the account, and the lender receives weekly arrears balance snapshots plus an adjustment file that is reliable most of the time.

Setup. Arrears of £1,200, repayable over six monthly instalments of £200. Failure threshold thirty days past the oldest unsatisfied instalment. Opening balance $B_{0} = 1200$ .

Week	Observed balance	$Δ B_{t}$	Classification	Basis	Effect
1	1,200	n/a	n/a	Opening	All instalments Pending
2	1,000	+200	payment	Positive movement	Instalment 1 Paid, arrangement Active
3	1,000	0	neutral	No movement	No change
4	1,035	−35	exogenousAdjustment	Fee record for £35 dated in window	Balance accepted at 1,035. Allocation untouched
5	835	+200	payment	Positive movement	Instalment 2 Paid
6	875	−40	anomaly	No record in window	Balance accepted at 875. Allocation untouched. Flag raised
7	675	+200	payment	Positive movement	Instalment 3 Paid. Flag remains open
8	675	0	neutral	Late fee record for £40 arrives	Week 6 delta reclassified. Flag cleared

Three things in this trace are worth drawing out.

Week 4 is where the prior model fails. A £35 fee raises the balance. Under the unclassified algorithm of the previous version, $min (- 35, u n p ai d_{k})$ returns $- 35$ , and instalment 2 loses £35 of allocation it had never received while instalment 1 would be next in line to be stripped on the following adverse observation. The customer paid on time and the system records them as further behind. With classification, the balance moves and the schedule does not.

Week 6 is where the model declines to guess. A £40 increase arrives with no record. It could be a fee whose record is late, a reversal of the week 5 payment, or a correction to an overstated balance. Each implies a different handling and the signal does not distinguish them. The model accepts the balance, leaves allocation alone, and flags. Note that it does not stop processing: week 7 allocates normally with the flag still open, because one unexplained observation must not freeze an account.

Weeks 6 to 8 show the conservation invariant working as a diagnostic. At week 7, with the anomaly unresolved, the known figures are $B_{0} = 1200$ , classified adjustments of £35, payments of £600, and remaining balance £675:

$1200 + 35 = 1235 against 600 + 675 = 1275$

The identity does not close. The residual is £40, which is precisely the value of the unclassified movement. The invariant does not merely detect that something is wrong; it quantifies exactly how much of the account is unexplained. At week 8 the fee record arrives, the week 6 delta is reclassified as exogenousAdjustment, and:

$1200 + 75 = 1275 against 600 + 675 = 1275$

The account has converged.

The case where it does not converge

Change one thing: the week 8 record never arrives. The account continues receiving payments and allocating them correctly, so its instalment states remain right, but the conservation residual stays at £40 indefinitely.

The model does not resolve this and does not pretend to. What it provides is an account carrying an open flag, an exact residual, and a date from which the residual has been outstanding. That is enough for an operational process to investigate, and it is the difference between a system that is wrong and a system that knows the extent of what it cannot explain.

The limitation worth naming is that the model cannot distinguish a deficient adjustment feed from an account with genuinely unexplainable movement. Both present identically as accumulating residuals. Telling them apart requires a population level view across accounts, which sits above this model rather than inside it.

9. Implications for Financial System Architecture

Independence from complete event histories. The system does not require complete transaction histories to maintain consistent state, provided adverse movement is classified against records.

Deterministic auditability. Transitions recorded with their triggering observation and classification provide a reconstructable explanation of behaviour. The week 8 reclassification in the case above is auditable precisely because both the original classification and the correction are retained.

Resilience to partial observability. Systems operating with delayed or incomplete signals converge toward correct state under the stated conditions, and report their own non-convergence otherwise.

Structural reframing. Financial systems should be modelled as state machines governing obligation relationships rather than ledgers processing transactions.

10. Summary

Convergent obligation systems provide a deterministic framework for modelling financial obligations under incomplete observability. By treating obligation state as primary, classifying observed movement before applying it, enforcing invariant preserving transitions, and applying deterministic reconciliation, the system maintains consistent financial state and converges toward correctness over time.

The convergence result holds under three stated conditions, one of which depends on an external record feed. Where that condition fails, the system is designed to declare the gap and quantify it rather than approximate through it. That behaviour, rather than the limit expression itself, is what makes the architecture usable in production.