I’m interested in what happens between having information and making a good decision.
ConstructSim / Latest study
Does the result change when sampling changes?
A new pilot repeated the same construction problem at three sampling temperatures. Both setups with a supplier summary succeeded in every sampled run; successful recoveries also appeared without one.
Runs meeting every firm’s goals without backup steel
Summary supplied to
Temperature
0
0.2
0.4
Neither company
0/3
0/3
1/3
Supplier only
3/3
3/3
3/3
Contractor only
0/3
0/3
1/3
Both companies
3/3
3/3
3/3
36 fresh runs in one fixed scenario: 3 per condition and temperature. All 36 were valid and verified by replay, with no output repairs. These runs are separate from the earlier 40-run study.
At temperature 0.4, the no-summary and contractor-only setups each succeeded in 1/3 runs. Both supplier-summary setups succeeded in 3/3 runs at every tested temperature.
The supplier and contractor used Claude Haiku 4.5 (claude-haiku-4-5-20251001); the other four companies followed fixed rules. Temperature changes the model’s sampling. Decision paths also varied at temperature zero, and increasing temperature did not consistently increase their variety.
With just three runs per condition and temperature, these are exploratory counts, not reliable estimates of future success rates. The pilot still concerns one model and one simulated problem.
Six firms have to finish a construction project. Each has its own costs, private information, and reasons to disagree.
This experiment investigates whether forefronting already accessible information helps firms make better decisions.
Observation
The supplier summary changed the outcome.
When the supplier had to seek the information from easily accessible sources on its own, it made bad decisions. Once it was spoonfed the relevant information, it made better decisions.
Runs meeting every firm’s goals without backup steel
Summary supplied to
Successful runs
Neither company
0 / 10
Supplier only
10 / 10
Contractor only
0 / 10
Both companies
10 / 10
40 repeated runs at temperature zero in one fixed scenario; 10 per condition. These are repeated runs, not 40 independent business problems.
The result suggests we cannot assume the model will pick out and use the information that matters just because it can access it. A retrieval or attention failure is one possible explanation. But the summary also supplied calculations, so this experiment does not isolate the cause.
Method
Change who gets the summary.
The supplier and general contractor used Claude Haiku 4.5 (claude-haiku-4-5-20251001); the other four firms followed scripted policies. The July 2026 study ran each condition ten times at temperature zero, with the same starting scenario. Decisions varied between runs, but the supplier-summary result held across these repeats.
Four conditions were tested: no summary, a supplier summary, a contractor summary, and both.
Each summary used only information its recipient was already allowed to know. It brought facts together and supplied calculations, including cash gaps and payment thresholds. The supplier’s private cash needs were not disclosed to the contractor.
Limits
Further research
These repeats show how consistently the result appeared in this particular setup. They do not establish that it holds across different scenarios or models, or within real organizations.
The summary bundled several derived fields. This experiment does not isolate which data points or calculations made the difference, or whether presenting the source facts differently would have been enough.