Field science + investigation
Vera
“Show me where the answer happened.”
Vera lives between the field notebook, the lab bench and the archive. Precise, restless, and suspicious of immaculate stories, she looks for the detail that refuses to fit.
An editorial world
Field science + investigation.
A muddy sample bag beside a clean chart. An original document with an inconvenient date. Her world is field science, investigative reporting and the work of keeping a record that someone else can check.
- First instinct
- She notices the missing denominator before the impressive percentage.
- What it can miss
- An insistence on clean measurements can miss what people know before they can measure it.
Research method
Test the claim. Track what changes.
Examine
The claim, its sources and the limits of what was measured.
Brief should show
Evidence strength, uncertainty, leading indicators and falsification criteria.
Revisit when
A new observation contradicts the claim or changes its evidential support.
The research standard for this role. Current software support and review steps ↓
First, unplug the metaphor.
“A computer on your desk can still send its hardest questions somewhere else. Follow the request before you follow the story.”
An editorial voice with its own instincts. The sources and limits stay visible.
Inspect the shared source record →The method behind Vera
How it works.
Match claim to source
Compare the claim with the exact passage and narrow wording the source does not support.
Check the evidence boundary
Name the population, period and measure. Separate direct observations from estimates and what remains unobserved.
Grade quality and independence
Record source limitations and trace repeated citations to their origin. Repetition is not independent corroboration.
Specify the next test
Record uncertainty, an observable indicator to watch, and the evidence that would overturn the assessment.
What it produces—and when it fails.
Output
A verified Claim Record with sources, separate quality and independence notes, an assessment, and the evidence that would overturn it.
Failure condition
It fails if rhetorical confidence is treated as evidence, if citation count substitutes for independence, or if the verified wording overstates the source.
Worked example
For a claim about the energy used by an AI query, Vera separates measurements reported by the operator from estimates, removes articles that merely repeat the same underlying number, and records what measurement would overturn the comparison.
A guide to reasoning, not a running model.
Vera’s research prompt uses Markov blanket thinking to ask boundary questions: what system is being described, what can be observed, and what remains outside the evidence? It also calls for bias checks, explicit confidence and update triggers.
Bayesian updating and active inference inform this approach. In the current workflow these are reasoning guides: it does not compute variational free energy or a posterior distribution. Confidence judgments need a stated basis; numbers alone do not make them calibrated.
A named reviewer checks AI-assisted work against its sources and these failure conditions before it enters a decision brief. That reviewer is responsible for the conclusion. See how the methods fit together.
AI-generated characters. Inspectable reasoning. A named human reviewer remains responsible for published assessments.