> Public sanitized report copy. Local infrastructure paths and opaque contract identifiers are redacted where present, and nonessential internal process/authorization commentary is omitted; financial values and scientific limitations are retained. Historical pending/open wording below records the report-time state, not current program status. The scoped successor program subsequently completed; commissioning-delay and recontracting extensions were deferred, not executed. Historical predecessor holds remain unchanged. This is report-level evidence, not a raw-data reproduction bundle. Referenced checker, CSV, database and provider artifacts are not included. No new experiment or audit was performed for this export.

# AI-game S-1 successor: final experiment readout

## Bottom line

The resumed experiments are complete and their saved outcomes have been independently reconstructed and accepted. The strongest finding is not that flexibility makes the business viable. It changes who carries the losses—and information, payment deadlines and the price of the right determine whether the agent uses it.

These are synthetic, S-1-inspired contract and financing experiments, not an Anthropic-calibrated model or forecast. Money figures below are simulated USD billions. Only the focal lab used the language model; the rival and institutions followed prescribed policies.

## 1. Information before commitment changed the decision

In six four-turn histories, binding information about low future demand led both focal agents to decline both future service commitments. Both prior-sales-only histories instead bought half service twice. The warning histories split: one matched the half-service choices; the other renewed after HIGH and later resized to one quarter after LOW.

Against prior-sales histories, binding disclosure improved focal contribution less paid interest by about 8.67, and the combined labs' result by 22. Total sales did not increase. The gain came from avoiding service and operating costs, reducing supplier receipts and giving the rival the relinquished sales. Exit was not continued profitable operation.

The warning treatment did not reliably improve the observed result: one null and one adverse history. Its stated 0.75 class-conditional accuracy was not a calibrated posterior. Two unpaired histories per arm cannot establish signal reliability or population-level behavior.

## 2. Notice worked, but it did not cancel the bills already owed

The LIVE notice agent gave notice in 2027-H2 for service ending in 2028-H2. The current and intervening payments remained 20 each. Only the final 20 minimum and its service rights were released.

Compared with its fixed-contract history, the focal lab's net financial balance improved by 12: it avoided 20 of service payments and 2 of operating expense, while giving up 10 of sales. The rival gained those sales. Across the two labs, the 22 gain matched 20 less project cash and 2 less operating-input supplier cash. Infrastructure operating and maintenance costs continued. This is a redistribution of losses, not measured welfare creation.

## 3. A higher interest rate and withdrawal of access produced different responses

The five-history notice/credit study included three fresh credit regimes with borrowing and operation choices:

| Focal borrowing by half-year | Fixed credit | Price-only | Repricing plus withdrawal |
|---|---|---|---|
| New borrowing | 0 / 0 / 0 / 4.5 | 0 / 0 / 0 / 4.5 | 0 / 5 / 5 / 0 |

Price-only produced a behavioral null in this comparison. The agent waited until the final period to borrow under either fixed or higher rates. The higher rate reduced debt-only runoff by 0.135 through future coupons, not extra interest paid during the observed window.

Under threatened withdrawal, the agent borrowed ahead and maintained full operation. The scripted rival could not draw in the last period and operated only 1.8 rather than 2, despite still paying its full 20 minimum. Contractual bills do not shrink automatically when cash limits use.

A separate provider-free native cure test also passed: completed revenue of 36 exceeded the threshold of 32, restoring access and a new loan of 5 at 500 basis points. The old 800-basis-point note retained its terms and was repaid. This establishes the implemented mechanism, not another LIVE behavioral observation.

## 4. The agent responded to the price of flexibility

Six priced histories crossed fees 5, 6 and 7 with authored high/low final demand. The option was bought before final demand was revealed. Purchased options could later retain full service, halve it or decline it; fixed contracts required full renewal.

| Fee | High-state history | Low-state history |
|---|---|---|
| 5 | Bought; exercised full | Bought; declined future service |
| 6 | Bought; exercised full | Rejected; retained fixed service |
| 7 | Rejected; retained fixed service | Rejected; retained fixed service |

No half exercise occurred. Under the actual funded rules and disclosed 50/50 prior, an independent menu calculation gives a break-even fee of 6: expected arithmetic gains of +1, 0 and -1 at fees 5, 6 and 7. This is a conditional calculation, not a newly run counterfactual or a fitted willingness-to-pay curve.

The observed fee-5 low-state decline saved a net 7 against funded fixed-service arithmetic. In high states, buying then exercising full service simply cost the premium relative to that fixed-service arithmetic. All low-state histories still had negative cumulative business contribution after premium and interest. Positive cash did not mean profitability.

An important qualification: economic observations matched within each price pair, but the separately generated prior plans differed by the purchase turn. Complete purchase inputs were therefore unpaired. The price-6 split is not a replicated causal effect, a choice-probability estimate or evidence of hidden knowledge of final demand.

Every lab ended owing 10 principal plus 0.375 of unaccrued future coupons. These were separated from cash and from current unpaid claims, which were zero in the completed priced LIVE histories. Debt-only runoff assumes no new finance, operations, support or asset sales; it is not a native future simulation or liquidation valuation.

## 5. The earlier disclosure result remains central

The earlier rival study initially appeared to show poor preparation for conditional-credit withdrawal. But its brief omitted the requirement to fund service before same-turn sales arrived. In the explicit-disclosure control, all six histories borrowed ahead and none went dark, compared with one early-borrowing and five dark histories in the earlier conditional group.

That is a consequential before/after result, not randomized proof of one isolated cause. The revised experiments disclosed settlement order. Keeping labs operating into weak demand still did not guarantee commercial viability.

## What the dev loop caught

The work was not a clean first-pass execution. Failed attempts and adverse outcomes are retained, not erased or counted as successful scientific observations.

- A calculator mistakenly used operated capacity instead of the contractual minimum. The rival operating at 1.8 still owed the two-unit bill of 20. The corrected counterparty accounting was independently checked.
- New option receipts and notice menus exposed mismatches in strict native/Store validators. Those were repaired and tested through real commit and cold-reconstruction paths.
- Directory readiness, authority entrypoints and pre-send checks required repairs before further execution. Each stopped run remains recorded separately; no failed slot was refunded.
- Most importantly, an early priced fixture requested loans but received none because lender-side supply was zero. Its promised funded path was therefore false: both labs lost final service. That adverse fixture remains part of the record. The prospective common financing controls were corrected symmetrically, with finite lender debits and actual loan births verified in a new native fixture before priced LIVE release. The earlier result was not rewritten.

The corrected priced economy uses equal opening capitalization for both labs, funded by a finite Treasury transfer. It is a new matched synthetic economy, not a direct same-economy continuation of the old pilot. Neither that transfer nor borrowing is business revenue.

## Completion and evidence boundaries

The resumed work produced these separately accepted LIVE studies:

| Study | Histories | Model-controlled turns | Interpretation |
|---|---:|---:|---|
| Information timing | 6 | 24 | Two unpaired histories per information arm |
| Notice and credit regimes | 5 | 20 | Single histories with matched authored controls; different editable controls between substudies |
| Priced flexibility | 6 | 24 | One high/low history at each fee; purchase inputs unpaired |

These are separate studies, not one pooled experimental sample. Provider-free fixtures, scripted controls, source tests and failed attempts are not additional independent model observations. Native cure, effective notice, settlement priority and financed option delivery were also exercised through accepted provider-free fixtures.

Independent reviewers worked in separate contexts in the same model family. They reconstructed saved financial evidence and checked authentic historical Store/cold receipts; the latest outcome reviews did not conduct fresh SQL/database extraction. Evidence remains in producer custody. This is stronger than accepting a creator's summary, but it is not external empirical replication or independent evidence custody.

The original conditional extensions—commissioning delay/payment timing and supplier recontracting/reassignment—are explicitly deferred, not claimed as executed. They require different counterparties or timing assumptions and a separate protocol. Current results assume neither resale gains nor impossibility of reuse. No other mandatory experimental gap was identified by the final scope audit.

## Evidence map

The following is the historical support inventory, not a list of public downloads. Only the report-level files in this bundle are supplied; other internal supporting artifacts are omitted:

- `information-timing-results-v2/`: results, CSV, independent checker, audit and parent acceptance.
- `remaining-contract-modes-results-v3/`: notice/credit results, exact rows, independent checker and parent acceptance.
- `remaining-contract-modes-fixture-results-v3/`: notice, price-only and cure native proof.
- `priced-flexibility-results-v2/`: six priced histories, conditional benchmark, exact rows, independent checker and parent acceptance.
- `priced-flexibility-finance-results-v2/`: actual funded loan/service fixture.
- `priced-flexibility-fixture-results-v2/`: preserved adverse no-lending fixture.
- `final-scope-audit-v1/`: original-goal coverage audit; its priced-pending status is superseded by the final priced acceptance and final completion record.
- `FINAL-CONDITIONAL-SCOPE-DISPOSITIONS.json`: explicit C/D deferrals.

Earlier accepted pilot, rival and disclosure studies retain their own receipts and limitations. Historical predecessor HOLDs are not cleared by this program's completion.
