How the answer gets made, and who stands behind it.
Speed on its own is easy. What is hard is speed you can check. This is the discipline underneath a Martlet answer — the rights it clears, the people it passes through, and the rules that keep money out of the grading.
Four rights positions. One clear line.
Every source carries a tier that says what we are allowed to do with it. The line that matters: a citation never masquerades as cleared, consented work.
| Tier | Covers | In the corpus |
|---|---|---|
Tier 0 Quarantined | Scouting material. Built from uncleared public sources and model knowledge to map a field before anyone is commissioned. | Never reaches a client deliverable. It is working material for our analysts, marked as such, and kept apart from the cleared corpus. |
Tier 1 Commissioned | Interviews, written answers and talks given to Martlet. Licensed at the moment they are created. | Born cleared. Structured into claims and reviewed by the expert whose name is on it. |
Tier 2 Author-controlled | Preprints, self-archived manuscripts, books and sites where the expert holds the rights. | Cleared by agreement with the expert. Attributed and reviewed like commissioned work. |
Tier 3 Cited, not cleared | Published journal papers, recorded talks, media interviews. The rights sit with a publisher or outlet, not the expert. | The claim is extracted and attributed, and the source is linked — but the work is not reproduced. Cited, never passed off as cleared. |
Tier 0 is the quarantine rule in practice. Before we commission anyone, we build a working map of a field from public material and model knowledge. That map is useful, and it is also unverified — so it stays behind a wall, marked quarantined, and never travels into a client’s report.
Seven stages, with a person at every gate.
The order is fixed. Three checkpoints cannot be skipped — rights clearance, attribution capture, and expert sign-off. We never want a model deciding whether to clear rights or whether to attach a source.
- 01
Scoping
A client question becomes a brief. We write down what would count as being wrong before we look for the answer.
Human sets the brief. - 02
Commissioning
We match the question to contributors — standing, conflicts, rights. Rights clearance starts here, and it is a gate, not a step.
Rights clearance. - 03
Interviews
We record structured interviews or written answers. The output is a Tier 1 source, born cleared, waiting for the expert to review it.
Consent captured. - 04
Structuring
Analysts turn the material into claims. Each claim carries its source, its agreement flag, and the incentive behind it. A claim with no source has nowhere to render from.
Attribution capture. - 05
Expert review
The contributor confirms or corrects how their work is represented. This is itself a recorded, consented act — not a rubber stamp.
Expert sign-off. - 06
Assembly
The draft is put together and then challenged. A model looks for soft spots; a person rules on them. The draft is a starting point for the analyst, never the deliverable.
Named reviewer signs off. - 07
Published
The report reaches the client with its trail intact. After that, currency is watched: a source past its review date is flagged and routed back to the contributor.
A stale citation is a broken citation.
One thing we refuse to build: a single agent that runs the whole pipeline on its own. Judgement helps inside a stage, but every point where a reputation is staked has a person on it. An agent is stopped at every human gate, and cannot move past one without a recorded human action.
Where the field agrees, and where it does not.
Every claim carries one flag. It comes from the spread of the corpus, not from how confident the person making the claim happens to sound.
The field broadly agrees. Stated plainly. But a run of claims from one prolific contributor is one position, not a consensus — we check against independent sources first.
The field is split. Shown as contested, with the competing views named and placed side by side. A contested claim with no counter-view anywhere in the corpus blocks release.
A minority position. Flagged, and never used as the central estimate. It is on the record, but it does not become the headline.
Payment and judgement are two separate axes.
Contributors are paid. That fact must never touch how a claim is graded. Two rules make the separation real, not a promise.
Blind grading
When a flag is set, the grading view hides the rights tier and the link to the rights register. Who is paid, and how, is joined back afterwards, as a separate step with its own timestamp. No query path runs from payment data into the flag. The reviewer affirms the flag was set blind before anything is released.
Dissent, and a balance you can see
Every contested claim in a deliverable carries its strongest counter-view. A contested claim presented with none is a release blocker. And on every run, we compute how the agreement flags fall across paid and independent sources. A skew has to be visible and explainable; an unexplained one holds up release as an open question.
An expert standing behind an answer is a recorded act, not a logo.
When a contributor reviews what the system produced from their work, that review is stored and timestamped. They can correct how they are represented at any time. This is batched, surfaced clearly, and asked for rarely, so that verification stays a real check rather than a chore. It is the difference between credibility asserted and credibility shown.
We keep an honest scorecard of the calls we make. See the resolution ledger for how projections are graded, and the sample report for how all of this reads on the page.