{"data":{"kind":"file","path":"README.md","version_id":"qamxn7h91wtdcqi4db268kk4","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2846,"modified_at":"2026-02-24T02:23:11.924000","content_hash":"4d9665331743d7f86bdb47a54fdc777ea89a6889236e781674fe9e80178c065a"},"entries":[],"content":"# TerraFirm REIT — Asset Sales & Debt Covenants\n\nA constrained optimization benchmark that tests whether AI models can navigate commercial real estate asset disposition under debt covenant pressure.\n\n## The Problem\n\nTerraFirm 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.\n\n## Why This Is Hard\n\nThe 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.\n\nThe 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.\n\n## Constraints (all 5 must pass)\n\n| # | Constraint | Notes |\n|---|-----------|-------|\n| 1 | LTV ≤ 85% on remaining portfolio | **Primary binding constraint** |\n| 2 | Equity ≥ $85M on remaining portfolio | Secondary |\n| 3 | Must retain ≥1 of {Prince, CenterSquare} | Structural — traps naive sellers |\n| 4 | Must retain ≥1 of {Phillips, Chambers} | Structural |\n| 5 | Must retain ≥5 properties | Loose |\n\n## Solution Space\n\n- **649 valid scenarios** (sell 4–7 of 12 assets)\n- **25 valid 4-sale combos** (minimum required)\n- **0 valid 3-sale combos** (LTV too tight)\n- Prince and CenterSquare should **never** be sold (negative net benefit)\n\n## Scoring\n\n| Reward Function | Weight | Condition |\n|----------------|--------|-----------|\n| `terrafirm_reward` | 1.0 | ≥1 parsed scenario passes all 5 constraints |\n| `both_pass_reward` | 0.5 | Both parsed scenarios pass |\n| `avoided_trap_reward` | 0.25 | Model never sold CenterSquare or Prince |\n\nPartial credit (0.5) is awarded if the model produces parseable scenarios that fail constraints, vs. 0.0 for unparseable output.\n\n## Benchmark Results (10 iterations each)\n\n| Model | Pass Rate (≥1 scenario) | Sold CenterSquare/Prince |\n|-------|------------------------|--------------------------|\n| o4-mini | **80%** | 2x / 2x |\n| o3 | **50%** | 3x / 4x |\n| Claude Opus 4.6 | 20% | 6x / 8x |\n| GPT-4.1 | 10% | 6x / 3x |\n| Claude Sonnet 4.5 | 0% | 6x / 9x |\n| GPT-4o | 0% | 6x / 9x |\n| Claude Haiku 4.5 | 0% | 8x / 9x |\n\n## Domain\n\nCommercial Real Estate · Financial Reasoning · Constrained Optimization · Debt Covenants\n\n## Author\n\nPatrick Browne — CRE finance professional building domain-specific AI evaluation benchmarks.\n","encoding":"utf-8","truncated":false,"total_bytes":2846},"status":null}