How we show our work
Defensibility you can check.
Almost every vendor calls their work "traceable" and "auditable." Few can show what that means in practice. These four techniques are what it means in ours, and each one comes from an engagement we've delivered.
A claim without a receipt is a marketing statement. When a reviewer asks "how do you know," the answer shouldn't be "the analyst said so."
Every quantitative claim on a program analysis site is backed by a structured ledger entry colocated with the component that displays it. The entry captures the exact claim text, a confidence level (HIGH, MODERATE, or LOW), and a reproducible source, usually a SQL query against the warehouse. Before release, a verification pass reruns the ledgered numbers against the database. Entries that can't be rerun are marked pending in the ledger, and we correct any that don't match.
claim_id: china_optomechanics_share
claim: "China accounts for 36.9% of optomechanics publication share."
confidence: HIGH
sources:
- type: database_query
database: supabase
table: program_citations
query: |
SELECT country, COUNT(DISTINCT citing_doi) AS papers
FROM program_citations
WHERE research_community = 'optomechanics'
GROUP BY country
ORDER BY papers DESC;
expected_results:
china_share: 0.369AI output looks the same whether it came from a verified source or a fabrication. A researcher reading an AI summary has no way to tell what the archive actually says and what the model guessed.
Every citation in the archive console is a direct quote from a source PDF, with its page range and a link to the original document. When an answer reasons past its sources, it says so in prose, and the reasoning carries no citation number. Figures carry the same distinction. A mark that rests on a cited passage is drawn solid. A mark the model inferred is drawn hollow with a dashed outline. A careful reader can always tell what the archive says from what the model extrapolated.
An AI that's eager to give the answer is a terrible teaching tool. It's also a terrible decision-support tool in any setting where the human has to own the call.
The Georgetown platform runs a delegation contract written directly into the code. The AI cannot verify assumptions. It cannot define success criteria. It cannot advance the team past a phase gate. Those gates are PL/pgSQL stored procedures — the database itself enforces the constraint, not a hope that the model will behave. The contract is small. The effect is large. It changes the whole shape of what the product is for.
export const DELEGATION_CONTRACT = {
// The AI observes and assists. The human decides.
canVerifyAssumptions: false,
canDefineSuccessCriteria: false,
canAdvancePhaseGate: false,
// Gates are enforced at the database layer, not in application code,
// so no agent can call its way past them.
gateEnforcement: 'plpgsql_stored_procedure',
} as const;A model can cite a source that doesn't say what it claims. And an assessment that says "likely" never finds out whether it was right.
Every statement in a Deepfield assessment either cites a source passage with a verbatim quote or is marked as the system's own inference. Each quote is checked mechanically against the run's frozen source snapshot. A quote missing from the passage it cites goes back to the model, and if it still fails, the run fails. Each key judgment is backed by dated, checkable propositions. Their forecasts go into a hash-chained ledger before the answer can be known, and they're scored when the date arrives. Our analysts review the key judgments by hand against the sources before the assessment goes out.
key_judgment KJ-2: "Competitor X ships a rival system within the horizon"
probability: likely (analyst judgment, not scored) confidence: moderate
rationale:
"X has begun pilot production"
cites passage P-118 quote: "pilot production began in March"
check: quote found word for word in the frozen snapshot
"X is likely to prioritize this line over its legacy product"
judgment: true (marked as inference, no citation)
backed by proposition PR-7:
"X announces general availability before 2027-06-30"
settled by: X's press releases knowable_at: 2027-06-30
forecast: 0.64 (median of 5 draws), committed to the ledger, seq 4182
resolution: pending score: when it resolves"Defensible" is either a claim or a property of the work. In ours, it's a property of the work.
Initiate Contact
Bring us a decision you have to make and defend.
Tell us about the decision you're trying to improve. We'll schedule a briefing with our principals to understand your environment and see whether the fit is right.
Schedule a briefing