AI Model Evaluation & Safety Authority — Frontier-Model Recovery

The mandate

Twelve days before a financing diligence review, the evaluation lead resigned after the founders overruled a recommendation to withhold a bilingual coding model. Subsequent analysis found benchmark contamination and an agentic tool-use regression that the existing release memo had not disclosed, forcing the board to freeze deployment and remove release authority from the product council.

The interim must be available within fourteen days and will hold the executive safety seat for nine months. A global search for a permanent leader begins after the new evaluation architecture survives its first external replication in month four; the contract includes a thirty-day successor overlap and will not extend if the search slips.

Handover is achieved when every production model has a version-linked evaluation record, leakage-controlled benchmark suite and documented deployment boundary; two independent red teams have closed critical findings; and three consecutive release councils have applied the same evidence threshold without founder exception. The permanent appointee must then chair a complete review and accept the residual-risk register.

The interim may block any candidate release, quarantine suspect evaluation data, allocate sanctioned evaluation compute, commission specialist testing within a ₹4 crore envelope and appoint temporary incident leads. Changes to the board’s risk appetite, external disclosure of vulnerabilities, expenditure above that ceiling and permanent appointments require board approval; the interim may not amend fundraising representations or publish model weights.

Core model architecture, consumer-product pricing and the next capital raise are outside this assignment. The seat may require evidence from those workstreams, but it does not own research-roadmap selection, enterprise contracting or remediation of unrelated software-security debt.

Why this seat is open

The disputed release exposed a safety function whose veto existed on paper but could be reversed without recorded evidence. Neither founder can occupy the role credibly while the board examines the incident and prospective investors test governance. An independent operator is needed to rebuild the release system before a permanent appointment is made.

What you will own

• Ratify a release-evidence standard that links each model’s threat assumptions, capability claims, deployment boundary and residual risk to reproducible evaluation results.
• Authorise, condition or block candidate releases through written decisions that identify failed thresholds, compensating controls and the evidence required for reconsideration.
• Reconstruct the contaminated benchmark lineage and decide which historic performance claims must be withdrawn, restated or retested against protected holdouts.
• Stage adversarial evaluations covering autonomous tool use, multilingual misuse, code execution, data exfiltration, deceptive behaviour and loss of operator control.
• Convene a model-risk council with protected escalation routes, declared dissent, conflict records and an explicit prohibition on undocumented founder overrides.
• Commission two independent replications and reconcile material variance between internal scores, external findings and behaviour observed in enterprise pilots.
• Transfer the evaluation registry, compute plan, unresolved research questions, incident playbooks and first ninety-day agenda through a witnessed successor review.

Candidate qualifications

• Held chief safety, model assurance or enterprise evaluation authority inside a frontier-model laboratory or an advanced AI platform releasing foundation models.
• Personally withheld or constrained a consequential model release and can evidence the technical threshold, executive challenge and subsequent disposition.
• Designed evaluation harnesses spanning capability measurement, adversarial robustness, agentic action, multilingual performance and post-deployment monitoring.
• Investigated benchmark leakage or training-data contamination deeply enough to separate genuine capability movement from invalid measurement.
• Presented model-risk evidence directly to boards, investors or public authorities while protecting sensitive methods and unresolved vulnerabilities.
• Completed a leadership transition in a founder-led environment where safety independence, commercial urgency and scarce evaluation compute were in active tension.

Non-negotiables

• Available within fourteen days for an exclusive assignment, with four working days each week in Bengaluru and monthly travel to Hyderabad.
• No undisclosed investment, employment negotiation or paid relationship with a competing model laboratory or proposed evaluation supplier.
• Will exercise the documented release veto even when delay affects a financing, customer launch or public benchmark commitment.
• Accepts that the option component is illiquid, dilutable, conditional on approvals and separate from the cash day rate.

1. 49 words maximum. State your notice position, earliest Bengaluru start date and any commitment that could prevent exclusive service.
2. 49 words maximum. Describe one model release you personally blocked or constrained, including the failed evidence threshold and eventual outcome.
3. 49 words maximum. How did you prove that a benchmark gain reflected contamination rather than a genuine capability improvement?