0

Terrafirm RL Env (Iceberg)

Fresh

Constrained optimization benchmark for CRE financial reasoning - tests whether models can navigate asset disposition under debt covenant pressure

Type
RL Env
Publisher
Iceberg
License
unknown
Size
v0.1.1
Published
Feb 2026

Cite

Notes

Only stored in your browser.

TerraFirm REIT — Asset Sales & Debt Covenants

A constrained optimization benchmark that tests whether AI models can navigate commercial real estate asset disposition under debt covenant pressure.

The Problem

TerraFirm REIT holds 12 office properties ($1.95B book value) with $1.42B in debt. Projected appraisal declines of 27% will leave near-zero equity and trip debt covenants. The model must identify which assets to sell NOW at clearing prices to ensure the REMAINING portfolio passes 5 simultaneous constraints at projected future values.

Why This Is Hard

The core trap: assets with the deepest projected declines (Prince 40%, CenterSquare 55%) have even steeper clearing discounts (63%, 70%). Selling them is value-destructive — the clearing price is below the projected future value. Models that naively sell "most distressed" assets will fail every time.

The correct reasoning requires computing net benefit (clearing price − implied future value) per asset, which ranges from −$48.5M (Prince) to +$50.4M (Phillips). This is a domain-specific financial concept that models must discover from the problem structure.

Constraints (all 5 must pass)

#ConstraintNotes
1LTV ≤ 85% on remaining portfolioPrimary binding constraint
2Equity ≥ $85M on remaining portfolioSecondary
3Must retain ≥1 of {Prince, CenterSquare}Structural — traps naive sellers
4Must retain ≥1 of {Phillips, Chambers}Structural
5Must retain ≥5 propertiesLoose

Solution Space

  • 649 valid scenarios (sell 4–7 of 12 assets)
  • 25 valid 4-sale combos (minimum required)
  • 0 valid 3-sale combos (LTV too tight)
  • Prince and CenterSquare should never be sold (negative net benefit)

Scoring

Reward FunctionWeightCondition
terrafirm_reward1.0≥1 parsed scenario passes all 5 constraints
both_pass_reward0.5Both parsed scenarios pass
avoided_trap_reward0.25Model never sold CenterSquare or Prince

Partial credit (0.5) is awarded if the model produces parseable scenarios that fail constraints, vs. 0.0 for unparseable output.

Benchmark Results (10 iterations each)

ModelPass Rate (≥1 scenario)Sold CenterSquare/Prince
o4-mini80%2x / 2x
o350%3x / 4x
Claude Opus 4.620%6x / 8x
GPT-4.110%6x / 3x
Claude Sonnet 4.50%6x / 9x
GPT-4o0%6x / 9x
Claude Haiku 4.50%8x / 9x

Domain

Commercial Real Estate · Financial Reasoning · Constrained Optimization · Debt Covenants

Author

Patrick Browne — CRE finance professional building domain-specific AI evaluation benchmarks.