# Stage 120 — State-dependent annual floor grants

## Result

Across 54 paired configurations, **45 adaptive paths pass** the current-need funding and annual-consumption gates, versus **18 fixed-schedule paths**. All paths complete the horizon.

| Production mode | Adaptive passes / 27 | Fixed passes / 27 | Interpretation |
| --- | ---: | ---: | --- |
| External output | 18 | 0 | Adaptive grants close the actual annual gap at 10% and 20% dividend ceilings. |
| Capital feedback | 27 | 18 | The matched control already meets the floors; adaptive requests and levies are zero. |

The comparison is not a pure estimate of schedule targeting: the adaptive rule also changes the treatment of unfunded requests and the owner's behavioral assessment base, as specified below. Counts are scenario outcomes, not probabilities or independent empirical observations.

## Current-state rule

For each recipient at an annual step, let T be the nominal consumption floor, S current spendable income before the grant, and R available cash plus equity-redemption resources. Under the inherited 90% income-spending rule, the requested grant is:

**B = max(0, T / 0.90 − S, T − R).**

The first constraint makes desired consumption reach the target; the second makes resources sufficient to purchase it. Conditional on that current state, a fully paid B is the smallest nonnegative grant satisfying both constraints. It is not a lifetime optimum or a guarantee about future states.

T equals the declared floor times matched annual-control year-1 real consumption, cohort mass and current price. The reference stays fixed. The policy does not lower its target after a bad outcome, and it is not anchored to a pre-automation observation.

The owners' actual levy is limited by total current requests, the current dividend ceiling and available owner cash. Receipts are allocated proportionally to recipient requests. A zero request produces zero levy and zero grant.

## Funding and contractual distinction

The adaptive instrument is a **current-year funding-limited grant**, not an unconditional benefit entitlement. An unmet request remains visible as current unmet need. It does not create tax arrears or a future grant debt. The following year recalculates its own gap; missed consumption is not counted again as an automatic claim.

By contrast, the fixed comparator preserves the stage-119 assessed-tax and transfer-arrears rules. Its owner response q applies to the full accrued assessment. In adaptive mode q applies to the actual assessed and paid levy. When a request is fully funded these amounts coincide, but failed cases can follow different behavior and obligations. No claim is made that all differences are caused by targeting alone.

A case cannot pass merely because no liability was created: the funding gate separately requires every current request to be paid, and the consumption gate tests actual annual consumption.

## Example and feedback

At a 90% floor, q=0.5 and a 10% dividend ceiling in the external-output mode, the adaptive path reaches 90% in both recipient cohorts. Its twenty-year tax sum is about 1.4182 nominal model units below the fixed schedule, and terminal capital is about 0.18% higher than that fixed comparator. These are undiscounted accounting comparisons, not welfare gains.

In the capital-feedback mode the adaptive path exactly reproduces the no-policy control: no requests, no owner levy and no extra grants. This avoids imposing a positive payment merely because an old static schedule prescribed one. It does not establish that grants are unnecessary under other productivity, ownership or fiscal assumptions.

The engine recomputes portfolios, corporate investment, cash, capital, public benefits and future dividends using the inherited annual mechanisms. Employment and prices remain external; the capital-only elasticity of 0.3 remains a declared assumption. Investment is still residual output after consumption and existing government goods, so policy effects depend on that closure.

## Declared grid and scope

There are two production modes, three floors (90%, 95%, 100%), three owner responses (0, 0.5, 1) and three dividend ceilings (5%, 10%, 20%). Each adaptive case is paired with a freshly computed fixed-schedule case and the mode-specific no-policy control.

The nine adaptive external-output cases with a 5% ceiling fail the combined gates and remain in the archive. No outside funding or new credit is inserted to make them pass.

This remains an annual experiment. It does not verify monthly transfer punctuality, estimate behavioral coefficients, identify an optimal policy or establish full macroeconomic equilibrium.

## Validation

Seven tests verify current-state grant sufficiency and minimality for income and resource constraints, zero-overlay recovery, zero policy when no need exists, cash/cap/unfunded-request identities, failure with zero funding despite unmet need, invalid inputs and archive reproduction.

No original simulator files were changed. The new variant also makes its reported cash-surplus field include the policy's cash receipts/payments; the acceptance checks use actual state balances and consolidated money identities.

```sh
python scripts/adaptive_floor_grant_v120.py --output data/processed/adaptive_floor_grant_v120.json
python scripts/test_adaptive_floor_grant_v120.py
```

[Paired annual paths, requests and unmet need](data/processed/adaptive_floor_grant_v120.json) preserve all cases, both comparator rules, references and hashes. Repository research, web and frontend CI remain the merge gate.

## Next research step

Test whether these results survive changes to the capital-response elasticity and the rule that allocates output between consumption and investment. Keep current-need, entitlement and timing outcomes separate. A rule that restores a floor under one residual-investment closure does not establish that the same financing base and productive capacity survive under alternative closures.