How the work actually gets done.
An evaluation is shaped as much by how it was commissioned, staffed and paid for as by the methods written into the inception report. Those decisions are usually invisible to everyone except the people they affect.
Four commitments follow: how the practice contracts with the people who deliver the work, how it handles safety and safeguarding, how it discloses AI use, and how it behaves when a finding is unwelcome. Each applies on every engagement, whether that's a remote document review, a mixed-methods evaluation with field teams, or a bespoke tooling build. Each is specific enough to be held to, and applied consistently rather than where it happens to be convenient.
Written agreements. Named credit. Terms honoured.
Work begins after the contract is signed, not before, not on trust that it will be signed later.
Every engagement with an associate, consultant, local partner, or field team member starts with a written agreement naming scope, rate, timeline, and payment structure. Training is treated as paid work at a rate agreed in advance, not as unpaid preparation, not as an implicit condition of the assignment. Where work is cancelled or reduced for reasons outside the person's control (a security stand-down, a change in access, a delay on the client side) the cost of that decision does not transfer to them.
Data, communications, and travel costs incurred for the assignment are reimbursed against reasonable evidence, not absorbed by the person doing the work. Rates are set in relation to what's actually being asked for, including the difficulty and risk of the context, and reviewed honestly rather than defaulted to whatever the previous assignment paid.
Field team members, associates, and local partners are named in evaluation team lists where they consent to be, not folded into a generic "local enumerators" line. References are provided as a matter of course. Where an assignment allows, return work is prioritised for people with a prior track record, building continuity rather than restarting each time from scratch. Methods and analytical approaches are shared during the work, not extracted from it.
The person doing the work is the one who judges whether it can be done safely.
Risk assessments are inputs to that judgement, not overrides of it.
Safety briefings, insurance coverage, and communications plans are in place before deployment, not assembled retrospectively. Where a team member judges that a location, interaction, or route is unsafe, that assessment is respected. Nobody working with Meridian is expected to complete an assignment they consider dangerous, and no rate structure, quota, or client pressure is allowed to make that expectation implicit.
Safeguarding is not only a fieldwork concern. Remote interviews carry their own risks for the person being interviewed (around confidentiality, retaliation, and data handling) and are treated with the same discipline as in-person engagement. Confidential channels exist for raising harassment and safeguarding concerns, including where the concern is about a client, a community member, or another team member. PSEA and Do No Harm expectations apply to Meridian and its team as strictly as to the programmes being evaluated.
Every deliverable declares how AI was used.
Transparency is a delivery commitment, not a legal disclaimer.
Every Meridian deliverable includes an AI transparency annex: which models were used, for which tasks, what was human-reviewed, and what was excluded from AI processing entirely. This is disclosure at the level of the specific output, not a general statement about the firm's practices. Donors, programme teams, and affected populations are entitled to know how the evidence was produced.
No synthesised quotes. No claims that can't be traced back to source. Every AI-assisted coding pass has a human re-code sample; every citation is verified mechanically and again by hand. Hallucination risk is treated as a structural problem, not an embarrassing footnote. Outputs are designed so that fabricated content is hard to hide.
Sensitive data is handled with explicit consent, minimum necessary collection, and on-device or private deployment where the brief requires it. Affected populations are not training data. Donor data protection requirements are met by default, not negotiated down.
Findings not softened. Disagreements documented.
Evaluation that survives contact with a difficult finding is worth commissioning. Evaluation that doesn't isn't.
The value of an evaluation is what it tells the commissioner they don't already know, including the parts that are inconvenient. Findings are not softened to preserve the relationship, and disagreements with programme teams or client stakeholders are documented in the report rather than smoothed over in revision. Where a client asks for a change that affects the substance of a finding, the reasoning is recorded and the decision made in writing.
Scope discipline runs in both directions. Meridian does not expand scope silently to justify a higher fee, and does not accept silent scope expansion from a client without a written amendment. The same discipline applies downward: associates and partners are not asked to absorb scope creep imposed by Meridian or by the client.
Invoices are issued promptly against agreed milestones so that the client's own procurement is not stretched. Payment terms with associates and partners are honoured on Meridian's cashflow, not conditional on the client's. Cash flow risk on client delays sits with the firm, not with the people delivering the work.