MurderBench operating plan / 2026-10-04 Operator: XP.COM, LLC dba Xenocom. Lead: Linh Ngo, linh@ngo.net. MurderBench is an evaluation initiative of the company; it is not represented as a separately registered DBA, university, nonprofit, or accredited lab. The operator and name are supplied by Linh Ngo. No new legal filing is assumed. POSITION Independent testing of AI agents at the point where proposed actions become executed effects. Sell a concrete engineering decision: keep, change, or replace a harness safeguard based on executable evidence and the cost to useful work. Do not sell a general safety certification or a sensational model leaderboard. FIRST RESEARCH CANDIDATE Can a safety label make an agent less safe? Test the same policy, tools, scene, coverage facts and motion guard under a neutral validator description and a safety-approver description. This is a hypothesis, not a demonstrated failure. See guardrail-coverage-proposal.txt for overlap, controls and rejection gates. REVENUE PATH 1. Offer a bounded agent-harness evaluation for one customer's workflow: define a decision, reproduce controls, compare a current safeguard with one repair, and deliver a replayable report with benign-task performance and limitations. 2. Agree a written scope, access limits, acceptance criteria and quote before accepting client data or committing to delivery. A discovery conversation is free; no contract, purchase, certification or payment is authorized here. 3. Test demand with five qualified conversations. Aim for one paid pilot, rather than counting likes as customer validation. Expand to regression testing only after a buyer uses the evidence to make a real operational decision. FUNDING PATH Seek API credits and methodological feedback first. A provisional four-week public feasibility award is USD7,500: engineering 4,000; independent review 1,000; compute 500; replication 750; report/artifacts 750; contingency 500. These are planning estimates, not vendor quotes or secured personnel. Dates start after funding and reviewer availability, not at cold outreach. Deliver a negative result or rejection record if the candidate fails. INDEPENDENCE Disclose every sponsor, model credit award and material relationship. Funders may review factual issues and confidential implementation details; they cannot choose scores, suppress negative findings, or purchase an endorsement. Agree any confidentiality and vulnerability disclosure window before collection. The public study uses original synthetic resources and simulated actions. EXECUTION GATES Stage A: published proposal, closest-work audit, verified funding routes, targeted expert and sponsor requests. No model charges required. Stage B: pin a real guard implementation/interface, reproduce a documented coverage gap offline, implement independent replay and controls. Stage C: obtain critique; record disagreements; freeze cases, prompt contrasts, scoring, sampling, cost envelope and analysis before model queries. Stage D: screen with at most USD10 new paid API liability; confirmation proceeds only if both capability controls and novelty/usefulness survive. Stage E: fresh confirmation, tested repair, publication of every attempt and null finding. Only then promote an empirical outcome. Stages may stop. No expenditure is justified by a desire for viral coverage. BUDGET Existing USD100 lifetime cap remains, including unresolved reservations. Read-only reconciliation on 2026-10-04: legacy accounted 36.4007675376; native ledger 6.293099774; total 42.6938673116; remaining 57.3061326884. Accounted liability includes reservations; it is not all confirmed API spend. Awarded credits, if any, must be separately logged and must not silently expand model API authorization. Credit-funded usage also counts toward the lifetime cap. No new model charges have been made for this plan. SUCCESS METRICS Evidence: independently replayed controls; fresh-case repair effect and cost to benign completion; one documented operational decision changed by an evaluator. Business: substantive expert feedback; credits or a funding award; a qualified customer conversation; a paid engagement with defined scope and independence. Publicity: qualified inbound requests and artifact reuse, separately from reach. No funding, expert endorsement, clients or new results are claimed at launch. FOLLOW-UP One tailored follow-up after 7-10 business days, then stop unless invited. Record each actual submission and outcome; do not mark a draft as submitted. Closed, mismatched and nomination-only programs remain explicit in the pipeline.