feat: session 15 assignment - code review policy, proofs & evidence

[manishmcsa01-cmd]

Session 15 Assignment — Model Routing, Agent Economics & Observability

Author: Manish Kapoor (manishmcsa01-cmd)
Workload Domain: Automated Code Review & Security Audit (proofs/tasks/code_review.jsonl)

---

Key Evidence & Findings Summary

1. Floor Reproduction & Unit Tests

• glc_v4 test suite: 436 passed, 20 skipped
• S15Code test suite: 277 passed, 0 failed
• All core proof scripts executed cleanly against live gateway (GLC_BASE_URL=http://127.0.0.1:8111):
  • p2_budget_holds: PASS (5/5 checks)
  • p3_denial_of_wallet: PASS (6/6 checks)
  • p4_trace_export: PASS (11/11 checks) — Trace ID: 4db4392e22b2440f468bdb4813b7c9b0
  • p7_cross_model_ladder: PASS (10/10 checks) — 3 distinct providers/models verified.

2. Custom Workload & Economics

• Capability Ladder:
  • Economy: groq / openai/gpt-oss-120b ($0.15 / $0.75 per Mtok)
  • Standard: gemini / gemini-3.1-flash-lite ($0.25 / $1.50 per Mtok)
  • Frontier: groq / llama-3.3-70b-versatile ($1.50 / $4.50 per Mtok)
• Disjoint Judge Panel: gemini-2.5-flash and meta/llama-3.1-8b-instruct.
• Cost-per-Call Fallacy: Strategy B (Always Economy) lowered cost per call by 63.1%, but resulted in a 10.7% HIGHER cost per resolved task ($0.00008432 vs $0.00007616) due to retries on complex code review tasks.
• Winner: Strategy C (Budget-Aware Cascade) achieved an 87.5% resolution rate while cutting cost per resolved task by 41.7% compared to Strategy A.
• Break-Even Resolution Rate: Derived at 36.9% from measured price spread.

3. Adversarial Budget Attack (proofs/adversarial_budget.py)

• ALL 4 SCENARIOS PASSED ✓:
  • Scenario A (Runaway Loop): BudgetRefused triggered and spend capped.
  • Scenario B (Unaffordable Tier): Micro-budget ($0.000001) refused prior to provider call.
  • Scenario C (Cascade Climb): Escalation capped by budget ceiling.
  • Scenario D (Metering Verification): Transport calls match ledger charges rendering exact agreement.

---

Full benchmark tables, proof JSON outputs, case studies, and reproduction commands are documented in README.md.

[manishmcsa01-cmd]

Session 15 Assignment — Model Routing, Agent Economics & Observability

Author: Manish Kapoor (manishmcsa01-cmd)
Workload Domain: Automated Code Review & Security Audit (proofs/tasks/code_review.jsonl)

---

Key Evidence & Findings Summary

1. Floor Reproduction & Unit Tests

• glc_v4 test suite: 436 passed, 20 skipped
• S15Code test suite: 277 passed, 0 failed
• All core proof scripts executed cleanly against live gateway (GLC_BASE_URL=http://127.0.0.1:8111):
  • p2_budget_holds: PASS (5/5 checks)
  • p3_denial_of_wallet: PASS (6/6 checks)
  • p4_trace_export: PASS (11/11 checks) — Trace ID: 4db4392e22b2440f468bdb4813b7c9b0
  • p7_cross_model_ladder: PASS (10/10 checks) — 3 distinct providers/models verified.

2. Custom Workload & Economics

• Capability Ladder:
  • Economy: groq / openai/gpt-oss-120b ($0.15 / $0.75 per Mtok)
  • Standard: gemini / gemini-3.1-flash-lite ($0.25 / $1.50 per Mtok)
  • Frontier: groq / llama-3.3-70b-versatile ($1.50 / $4.50 per Mtok)
• Disjoint Judge Panel: gemini-2.5-flash and meta/llama-3.1-8b-instruct.
• Cost-per-Call Fallacy: Strategy B (Always Economy) lowered cost per call by 63.1%, but resulted in a 10.7% HIGHER cost per resolved task ($0.00008432 vs $0.00007616) due to retries on complex code review tasks.
• Winner: Strategy C (Budget-Aware Cascade) achieved an 87.5% resolution rate while cutting cost per resolved task by 41.7% compared to Strategy A.
• Break-Even Resolution Rate: Derived at 36.9% from measured price spread.

3. Adversarial Budget Attack (proofs/adversarial_budget.py)

• ALL 4 SCENARIOS PASSED ✓:
  • Scenario A (Runaway Loop): BudgetRefused triggered and spend capped.
  • Scenario B (Unaffordable Tier): Micro-budget ($0.000001) refused prior to provider call.
  • Scenario C (Cascade Climb): Escalation capped by budget ceiling.
  • Scenario D (Metering Verification): Transport calls match ledger charges exactly.

---

Full benchmark tables, proof JSON outputs, case studies, and reproduction commands are documented in README.md.