Sirjan Singh
Back homeINTERACTIVE LAB / DETERMINISTIC SIMULATIONS

Failure modes,
made inspectable.

Four focused browser demonstrations. Educational, deterministic, and explicitly not connected to production systems.

01

Cassandra supervision

Step through a trace as an agent failure is detected, patched, and replayed.

Read case study
LAB_01 / AGENT_SUPERVISIONDETERMINISTIC DEMO
01Agent runRUNNING12ms
02Tool callPASS418ms
02

SurviveCity reward hacking

Adjust reward weights and see how a policy can optimize the score while missing the objective.

Read case study
LAB_02 / REWARD_HACKINGEDUCATIONAL SIMULATION
REWARD63
SURVIVAL73
BEHAVIORBALANCED POLICY
03

Vendable guard

Compare a negotiable prompt rule with deterministic entitlement enforcement.

Read case study
LAB_03 / NEGOTIATION_GUARDFICTIONAL TRANSACTION
PROMPT RULE

Could you make an exception?

NEGOTIATING…
STATUTORY GUARD

Published entitlement remains outside the model context.

MEDIAN HELD / 100%
04

Contextual retrieval

Inspect the lexical, semantic, merge, and reranking stages of hybrid retrieval.

Read case study
LAB_04 / RAG_RETRIEVALDETERMINISTIC SAMPLE DATA

Exact symbols and lexical matches

01tool_gate.ts · 0.94KEEP
02approval-policy.ts · 0.88CANDIDATE
03run-controller.ts · 0.73CANDIDATE