S-1-inspired game · reader brief

# Who pays when a lab changes course?

A simulated AI lab can avoid a future compute bill, borrow before credit closes, or pay for the right to decide later. These experiments trace what each choice changes, and where the costs go.

Nicholas ZinnerBeacon BotOctober 2026 · Research brief

Synthetic economics, not an Anthropic forecast. All monetary amounts on this page are simulated USD billions.

[The game](#setup)[One recorded result](#example)[Five questions](#findings)[What it adds up to](#synthesis)

01 / Setup and play

## One model-controlled lab, a scripted economy

The focal lab buys compute service and sells output to customers. A language model chooses its permitted actions; the rival lab and five institutional participants follow prescribed policies. The institutions provide infrastructure, operating inputs, credit, public funding and customer demand. This is not a market of independently thinking AI agents.

Each scenario supplies customer demand. The game’s allocation rules determine how much each lab sells, so the rival can gain sales when the focal lab exits without changing its strategy.

Each history records four half-year turns, from early 2027 through late 2028. Depending on the study, the lab can renew, resize or decline future service, give cancellation notice, buy an option, or choose borrowing and operation. The game then applies its payment and sales rules and saves the resulting balances. Capacity is measured in modeled gigawatts; demand and sales use abstract output units, not measured AI capability. Full service in the example below is 2 modeled gigawatts.

1. 01 · FIRST**Debt comes due**Existing loans have payment priority.
2. 02 · THEN**Credit is drawn**New borrowing depends on access and supply.
3. 03 · BEFORE SALES**Service is paid**The lab funds current service and operation.
4. 04 · LAST**Sales arrive**Customers buy output; unsold output earns nothing.

**Later sales cannot pay an earlier invoice.** A future cancellation changes service on its effective date; it does not erase the bill due today.

Units, decisions and what a “history” means

Service capacity is expressed in modeled gigawatts, while production and customer demand use the game’s output units. These are stipulated quantities, not measurements of an actual data center or AI capability. Full lab service in the examples is 2 modeled gigawatts, with a minimum service bill of 20 and operating expense of 2 per half-year. Paying for capacity, operating it and selling its output are separate events.

A history is one recorded sequence of model choices and rule-imposed consequences, not a forecast or an independent population sample at each turn. The main successor studies contain six information histories, five notice/credit histories and six priced-option histories. They remain separate studies. Notice comparisons fix borrowing and requested operation; the credit comparisons let those choices vary.

The focal model was configured as `gpt-6-astra-900k` and observed as `gpt-6-astra` on the provider wire. This does not establish a separate “900k” model identity.

[Information source report](evidence/information-timing-results-v2/RESULTS.md) · [Notice and credit source report](evidence/remaining-contract-modes-results-v3/RESULTS.md)

02 / What a result looks like

## A bill arrives before the money to pay it

In the fixed-credit history, the lab waited until the final half-year to borrow. At the pre-sales cutoff it had 17.5 available, but needed 22 to pay for full service and operation.

Final half-year · fixed-credit history · synthetic USD billions

Cash before new borrowing**17.5**

Service 20 + operation 2**22 needed**

17.5 cash + 4.5 new loan = 22 available

The model requested 4.5. Credit was available, the loan arrived, and the lab funded full operation before receiving sales.

This is a recorded choice followed by a payment rule. The 4.5 loan solves the immediate shortfall; it is not sales revenue and still has to be repaid.

Follow this history to the endpoint

Across four turns, the focal lab received 70 in sales, paid 80 for service, 8 for operation and 1.5 in interest. Its business contribution after paid interest was therefore −19.5. It nevertheless finished with 10 cash and 4.5 principal owed, having started with capital and borrowed along the way.

Subtracting that principal and 0.225 in future coupons leaves 5.275. This debt-only accounting exercise assumes no further operations, sales, finance, support or asset sales. It is not a played future or a liquidation value.

[Source report: credit comparison and financial outcomes](evidence/remaining-contract-modes-results-v3/RESULTS.md)

03 / Questions → observed results → interpretation

## What changed the lab’s choices?

These are questions used to explain the accepted results, not retrospectively invented experimental hypotheses. The small, unpaired histories do not estimate how reliably models or companies would behave.

### Did information arrive soon enough to change a commitment?

ObservedBoth histories told binding future demand declined both future service commitments. Both histories shown only prior sales bought half service twice. The warning histories split: one matched half service; the other renewed, then cut to a quarter.

InterpretationBinding information helped the lab avoid costs before committing. Its contribution after paid interest improved by about 8.67 against the prior-sales histories, while its own sales fell and combined sales across both labs stayed unchanged. Exit surrendered sales to the rival; it was not continued profitable operation.

Warning limits and who received less

The six histories used the same authored demand path. Each information arm had two unpaired histories. Warning accuracy was specified as 0.75 conditional on the actual demand class; without a calibrated class prior, this was not a 75% probability that a particular warning was true. One warning history matched the prior-sales outcome and the other was worse. That cannot establish warning reliability.

The two labs’ combined improvement of 22 matched 20 less infrastructure-project revenue and 2 less operating-input supplier receipts. Total customer sales were unchanged. The infrastructure project’s costs continued.

[Source report: decision window, contrasts and counterparties](evidence/information-timing-results-v2/RESULTS.md)

### When does cancellation actually stop the bill?

ObservedThe notice agent canceled in late 2027 for service ending in late 2028. It still paid the current and intervening bills of 20 each. Only the final 20 bill and its service rights disappeared.

**Late 2027**Give notice  
Still pay 20

**Early 2028**Notice pending  
Still pay 20

**Late 2028**Exit effective  
Pay 0; no service

Cancellation affects a dated future obligation, not every payment after the decision.

InterpretationThe focal lab’s financial balance improved by 12: it saved 20 in service and 2 in operation, but gave up 10 in sales. The rival gained those sales. Suppliers lost receipts while infrastructure costs continued.

Comparison and counterparty accounting

The fixed and notice histories both required borrowing 5 and requesting full operation. Notice changed the final service payment from 20 to zero. Across both labs the gain was 22, matched by 20 less project cash and 2 less operating-input supplier cash. This was redistribution, not measured aggregate welfare creation. Debt remained payable; installed assets were not refunded or assigned a resale value.

[Source report: the notice comparison](evidence/remaining-contract-modes-results-v3/RESULTS.md)

### Does expensive credit behave like unavailable credit?

ObservedUnder both fixed rates and higher rates, the lab waited to borrow 4.5 in the last turn. Under repricing plus threatened withdrawal, it borrowed 5 in each of the two preceding turns instead.

InterpretationAccess changed borrowing timing in these histories; price alone did not. The higher rate still mattered financially, adding 0.135 in future coupons, rather than extra interest paid during the observed window.

Borrowing by turn and the rival’s shortfall

New focal borrowing · synthetic USD billions; columns run from early 2027 to late 2028

| Credit regime | 2027 H1 | 2027 H2 | 2028 H1 | 2028 H2 |
| --- | --- | --- | --- | --- |
| Fixed rate | 0 | 0 | 0 | 4.5 |
| Price only | 0 | 0 | 0 | 4.5 |
| Price + withdrawal | 0 | 5 | 5 | 0 |

New annual rates rose from 5% to 8% after weak completed revenue; existing notes kept their terms. The focal lab operated fully in all three histories. When access closed, the scripted rival operated only 1.8 instead of 2, despite paying its full 20 service minimum. Lower use did not automatically discount the contractual invoice.

These are single unpaired histories per regime, not evidence that agents generally ignore interest rates. A separate model-free mechanism test restored access after qualifying revenue; it is not an extra behavioral history.

[Source report: credit regimes and limits](evidence/remaining-contract-modes-results-v3/RESULTS.md)

### How much would the lab pay to decide later?

ObservedSix histories offered an option before final demand was revealed. At fee 5 both bought; at fee 6 one bought; at fee 7 neither bought. An owner could later keep full service, halve it or decline it. No half-service exercise occurred.

One eventual high-demand and one eventual low-demand history at each fee · fees in synthetic USD billions

| Fee | High-demand history | Low-demand history |
| --- | --- | --- |
| 5 | Bought → full service | Bought → no final service |
| 6 | Bought → full service | Rejected → fixed full service |
| 7 | Rejected → fixed full service | Rejected → fixed full service |

InterpretationThe fee-5 low-demand exit saved a net 7 against funded fixed-service arithmetic. High-demand buyers kept full service and paid the premium. The price-6 split is not evidence of advance knowledge: economic observations matched within each pair, but prior model-generated plans differed, so complete purchase inputs were unpaired.

Calculated value, cash and business losses

The disclosed prior gave high and low final demand equal probability. Under the funded rules and fixed rival policy, full-service final contribution was 18 in high demand and −12 in low demand. Choosing the best option outcome yields expected contribution 9 before the fee, versus 3 under fixed service. The calculated break-even fee is 6; net expected gains at fees 5, 6 and 7 are +1, 0 and −1. This is a conditional menu calculation, not another simulated history or a fitted willingness-to-pay curve.

The observed low-demand exit saves 20 + 2 − 10 − 5 = 7 after its premium. Its rounded terminal cash is 24, yet cumulative contribution after premium and paid interest is −13.125. All low-demand priced histories lost money on that business measure. Every lab finished owing 10 principal and 0.375 in future, unaccrued coupons; current unpaid claims were zero.

The corrected priced economy includes equal opening capitalization from a finite Treasury transfer. Neither that transfer nor borrowing is business revenue. It is a new matched synthetic economy, not a same-economy continuation of an earlier pilot.

[Source report: decisions, conditional benchmark and finance](evidence/priced-flexibility-results-v2/RESULTS.md)

### Did the model know sales arrived last?

ObservedAn earlier brief omitted the payment-before-sales requirement. After it was stated explicitly, all six control histories borrowed ahead and none lost final service. In the earlier comparison, only one of six borrowed ahead and five lost service.

InterpretationThe earlier apparent planning failure cannot be separated from that missing rule. This consequential before/after correction is not randomized proof of one isolated cause. Keeping both labs operating into weak demand still did not make the business viable.

What the disclosure control can establish

The comparison used the same model identity at a different run time. Histories within cells repeated the same paths and are not independent population samples. The revised brief stated settlement order; plans then repeated that order, while the short borrowing rationales cited withdrawal risk. Reported rationales reveal stated intent, not hidden reasoning or which consideration caused the choice.

The disclosure study’s acceptance included read-only inspection of historical databases. The later successor outcome reviews instead reconstructed saved evidence without fresh database extraction. Those scopes should not be conflated.

[Disclosure source report](evidence/disclosure-control-results-v1/SUMMARY.md) · [Acceptance and limitations](evidence/disclosure-control-results-v1/PARENT-ACCEPTANCE.md)

04 / Synthesis

## Flexibility changed exposure, not demand

The important differences appeared before a payment or commitment became unavoidable. Information could support exit, notice could release a later bill, and borrowing ahead could preserve service when credit closed. A costly option could make exit available without making it worth buying at every price.

Two distinctions hold the results together. **Liquidity is the ability to pay on time; profitability is whether sales cover business costs.** Starting capital and loans can keep a loss-making lab operating. And a lab’s improvement can come from **someone else receiving less**: the rival inherits sales, suppliers lose receipts, and infrastructure costs remain.

These experiments illustrate those mechanisms under authored rules. They do not estimate Anthropic’s finances, failure probability, contract values or a market-wide equilibrium.

Methods, current completion and evidence limits

The scoped S-1 successor program is complete. Current final completion and study acceptance records supersede older “pending” labels preserved in source reports. Commissioning delay/payment timing and supplier recontracting/reassignment were explicitly deferred, not executed. The results assume neither resale gains nor that reuse is impossible. Historical predecessor holds are unchanged.

The main studies contain six information histories / 24 model-controlled turns, five notice-and-credit histories / 20 turns, and six priced histories / 24 turns. These are not one pooled sample. Provider-free tests, scripted controls, failed attempts and arithmetic benchmarks are not additional independent model observations.

Separate reviewers in the same model family reconstructed saved financial evidence and inspected authentic historical database/cold-reconstruction receipts. The latest outcome reviews did not perform fresh database extraction. Evidence remains in producer custody; this is not external empirical replication or independent evidence custody.

Failed fixtures remain in the record. In an early priced fixture, requested loans received no funding because lender supply was zero, and both labs lost final service. Later financing controls were corrected prospectively and verified before the accepted priced histories. That adverse fixture was not rewritten as success. Other development repairs included contractual-minimum accounting and validation of notice and option receipts.

This brief uses saved accepted reports only. It adds no simulations or behavioral observations and offers no financial advice. Browser checks test presentation, not reader comprehension.

Source reports and acceptance records included with this brief

These are labeled sanitized public report copies, not a complete raw-data reproduction bundle. Historical pending wording records earlier states; the final completion summary governs current scope. Private paths, identifiers and delivery metadata are redacted or omitted. Keep the evidence/ folder beside the offline HTML. [Public scope and redactions](PUBLIC-SCOPE.md) · [Offline HTML](index.html) ·

* [Final experiment readout](evidence/final-readout/S1-EXPERIMENT-READOUT.md)
* [Final completion](evidence/FINAL-COMPLETION.json) · [Deferred extensions](evidence/FINAL-CONDITIONAL-SCOPE-DISPOSITIONS.json)
* [Information timing](evidence/information-timing-results-v2/RESULTS.md) · [Acceptance](evidence/information-timing-results-v2/PARENT-ACCEPTANCE.json)
* [Notice and credit](evidence/remaining-contract-modes-results-v3/RESULTS.md) · [Acceptance](evidence/remaining-contract-modes-results-v3/PARENT-ACCEPTANCE.json)
* [Priced flexibility](evidence/priced-flexibility-results-v2/RESULTS.md) · [Acceptance](evidence/priced-flexibility-results-v2/PARENT-ACCEPTANCE.json)
* [Disclosure control](evidence/disclosure-control-results-v1/SUMMARY.md) · [Acceptance](evidence/disclosure-control-results-v1/PARENT-ACCEPTANCE.md)
* [Claim-to-source map](evidence-map.json) · [Full reader text](prose.md)

Research brief · accepted synthetic results · public report-level evidence