You do not need Triplets yet, and we will say so.
AI grounding and evidence verification
Ground your AI answers in evidence you can inspect.
Triplets certifies what your AI can answer, refuses what it cannot, and keeps watch as the evidence changes. Built first for clinical and medical evidence.
Built for regulated industries grappling with the question:
Does a confidently wrong answer cost you more than a fast answer is worth?
Everything below is about you, because your current system optimizes for exactly the wrong side of that trade: it always answers, fluently, whether or not the evidence supports it.
The failure has six faces. You will recognize yours.
The confidently wrong answer is not one problem. It shows up six distinct ways, and most organizations live with several. Each one below is something a team somewhere says every week.
FACE 01The silent conflict
"Our sources don't always agree, and the AI doesn't tell us when they don't. It just answers."
Your corpus disagrees with itself by design: studies conflict, authorities diverge, documents contradict. A retrieval system averages the disagreement away or serves whichever fragment ranks highest. Triplets makes disagreement visible instead: every question resolves to supported, refused, or conflicted, and conflicted is an answer, not a failure.
FACE 02The re-answered question
"We ask the same questions every week and redo the work every time. Nobody knows if last month's answer still holds."
Your team's questions are not queries; they are standing decisions. A question asked twice deserves to be monitored, not re-answered from scratch. Triplets keeps questions alive as monitored decisions and tells you when, and only when, an answer's support changes.
FACE 03The stale answer
"Something changed upstream, and we found out from a customer, not from our system."
A regulation moved, a policy updated, a formulary changed, and every answer downstream went stale silently. Triplets recompiles continuously against the change, surfaces exactly which supported answers just became unsupported, and gets to you before the rejection letter does.
FACE 04The silent supersession
"The right answer and the obsolete answer both live in our system, and the AI can't tell which is which."
Every document management system holds documents alongside the versions that replaced them, with equal retrieval weight. Which one answered your last question? Triplets treats version history as structure, not metadata: every answer names the exact version it came from, and a document that conflicts with its own replacement surfaces as conflicted.
FACE 05The undefended answer
"Our clients don't just want the answer. They want to defend it, and right now that defense is our reputation and a PDF."
If you answer on behalf of someone else, the deliverable is not the answer; it is the ability to show your work. Triplets emits a reproducible certificate: same corpus, same questions, same partition, every time, re-runnable by your client's own reviewers.
FACE 06The audit you cannot replay
"If an auditor asks why our system said X eighteen months ago, reconstructing that is a project, and the reconstruction might not match."
Someday someone will ask why your AI said what it said. Triplets makes that a lookup, not an investigation: recompile against the evidence exactly as it stood on that date and get the same partition back.
Recognize two or more?
Bring us your hardest oneContinuous Trust Certification is the category. Triplets is the compiler.
A certified AI system is not one that scored well on a benchmark. It is one with a reproducible record of what it can answer from the evidence, what it must refuse, what is in conflict, and whether all of that is still true today. A certificate that is not kept green is a photograph of a moving object, which is why "continuous" is in the name.
Triplets is the compiler behind the category. Documents in. Monitored decisions out. Everything in between reproducible.
What every answer looks like afterward.
Every question your organization asks resolves into one of three states, and each state is useful.
Supported
Answered from your compiled evidence, traceable to the exact facts and versions that support it.
Refused
Declined explicitly, in writing, because the evidence cannot support an answer. Refusal is designed behavior, not a failure: every refusal is a logged corpus-expansion decision, and a system that declines to guess is worth more than one that always answers.
Conflicted
Your sources disagree, and the disagreement is surfaced with a verdict on which source governs and why, instead of being averaged away.
The partition is a map of your evidence boundary, not a grade. Nobody fails it.
Proof you can run yourself.
Five tests define the category. Run them against any system, including ours:
- 01
Ask a question the corpus cannot support and see if it refuses or improvises.
- 02
Ask about a refuted claim and see if the answer carries its state.
- 03
Correct a bad source and ask again next month.
- 04
Feed it two sources that disagree.
- 05
Ask for the same verdict twice and hand both to your auditor.
Most systems fail all five, fluently. See the five tests in detail.
The numbers behind the promise.
48 hours
to your first certificate via the Trust Baseline.
3 states
on every claim: supported, refused, conflicted.
88-94%
pass plus partial in design-partner-stage evaluations across four domains; the number that matters is measured on your corpus.
18
provisional patent families, Patent Pending.
Built on the Enigma Vault, which is PCI DSS Level 1 and SOC 2 Type II
Three ways to start, in ascending order of commitment.
One session
The supersession compile.
The fastest proof in the framework. We compile your own corpus, no external data needed, and show you where documents conflict with the versions that replaced them. Most organizations see a conflicted partition in the first session.
Request a supersession compileOne corpus
The staleness sweep.
Point us at one corpus that depends on external rules, rates, regulations, formularies, and we show you which of your standing answers already went stale.
Request a staleness sweep48 hours
The Trust Baseline.
The full first compilation: a representative slice of your corpus, your hardest questions, your first fingerprinted certificate in 48 hours, with no integration work on your side. Nobody fails a Baseline; you either learn your boundary or confirm it.
Start the Trust BaselineWhat the certificate proves. And what it does not.
It proves support within your compiled evidence, reproducibly. It does not prove truth, medical correctness, or regulatory compliance, and we will say that before your lawyer does. In clinical collateral, certificate vocabulary aligns with the GRADE-style evidence hierarchy your reviewers already use, so it reads as an extension of their discipline.
Bring us the decisions you cannot afford to get wrong.
How Triplets turns documents into traceable answers.
Documents in. Monitored decisions out. Everything in between is reproducible.
How Triplets compiles evidence for AI answers.
Per-query retrieval
Most systems fetch fragments at query time and hope the model behaves. A per-query system can re-answer a question. It cannot tell you what changed since last month, or why.
Compiled state
Triplets compiles first and answers from the compiled state, the way source code is compiled before it runs, because only a compiled, persistent state can be checked, versioned, and compared over time.
Seven capabilities. One machine.
Everything Triplets does reduces to seven capabilities working together. Every claim we make traces to one of them, and every one is present at every tier, because the architecture is never tier-gated.
01
The compiler.
Your corpus, trials, contracts, policies, specs, filings, is ingested into a structured model of entities, relationships, and verified values, with every claim resolved to a canonical fact carrying its provenance.
02
The tri-state partition.
Every question resolves to supported, refused, or conflicted within the compiled evidence. Disagreement between sources becomes a visible state with a governing verdict, not an averaging accident. Refusal becomes a designed, logged outcome, not a model mood.
03
Continuous recompilation and state-transition tracking.
The compiled state is compared continuously as evidence changes. When a supported answer becomes unsupported, or a refusal becomes answerable, that transition is the event, and you are alerted when, and only when, a decision's state changes. Not a feed. A state change.
04
Question persistence.
Questions live on as monitored decisions rather than dying as queries. The question your team asked in March is still being watched in September, which is what turns "redo the analysis" into "check the monitor."
05
Corpus versioning.
Every version of every document is structural, not metadata. The system cannot confuse a document with the version that replaced it, and you can recompile against the corpus exactly as it stood on any past date.
06
The provenance graph.
Every answer names its sources, versions, and the path from evidence to verdict. "Why did it say that" is answered by the answer itself.
07
The reproducible certificate.
Same corpus, same questions, same partition, every time, fingerprinted so a third party, your client, your auditor, your regulator, can re-run it and get identical verdicts.
Follow one question through the machine.
Compile.
Your corpus becomes canonical facts with provenance and state.
Decide.
Your standing questions are answered against the compiled state and recorded: what supports each verdict, what conflicts, where the boundary sits.
Partition.
Each answer lands as supported, refused, or conflicted, and the certificate reports all three honestly.
Monitor.
New evidence lands, a rule changes, a document is superseded. Affected decisions are re-evaluated automatically; monitoring is nearly free once the compiled asset exists.
Alert.
You hear from the system only when a decision's support changes. Silence means the certificate is still green.
Answered by the strongest method that applies.
Deterministic lookup.
Exact facts answered from structured, verified values. No generation, no room to improvise.
Relationship traversal.
Questions spanning entities and documents answered by walking verified relationships: which label covers which indication, which clause governs which agreement.
Constrained generation.
Open questions answered by a model bound to the compiled evidence. Verified facts override the model, never the reverse.
Past the boundary, refusal.
When no method can support an answer, Triplets declines, explains why, and logs the gap. Refusal is not the system giving up. It is the system declining to guess, in writing.
Track source conflicts and superseded documents.
Real corpora contradict themselves: a label and a later study, a contract and its amendment, a spec and its revision. A typical system silently serves whichever fragment ranked highest, which means your answer depends on a relevance score nobody audits. Triplets surfaces the conflict, determines which source governs and why, records the verdict, and shows its work. When a document conflicts with its own replacement, that surfaces too, which is why the fastest proof we offer is a compile on your own corpus: no external data, just your documents and their versions, and a conflicted partition most organizations did not know they had.
Monitor changes in evidence support.
A refusal is structural: that evidence is not compiled. Want the question answerable? Add the evidence and recompile. Scope conversations become corpus conversations, and the certificate grows with the corpus instead of aging with it.
A layer, not a rip-and-replace.
Triplets is substrate-agnostic. Your vector databases, graphs, and pipelines stay where they are; Triplets re-ingests the corpus into its own structured stores and governs answering at query time. Adopting it is not a migration, and in embedded deployments it certifies the corpus your existing tools answer from.
Triplets
Governs answering at query time
Your vector databases
Stay where they are
Your graphs
Stay where they are
Your pipelines
Stay where they are
You cannot prompt your way to a boundary.
Four things that do not work.
"Just prompt it: never hallucinate."
A prompt is a request, not a constraint; the model has no ground truth to check itself against.
"Tell it to refuse when unsure."
Confidence does not track correctness, so you trade coverage for fabrication and get both wrong.
"Fine-tune it on our domain."
Training lowers error rates; it cannot set a floor, cite provenance, or survive a model swap.
"Set temperature to zero."
Deterministic decoding makes output repeatable, not correct.
Correctness has to be enforced by a system outside the model, one that knows what is verified, what is not, and where the boundary sits. That system is the compiler.
See the machine run on your own evidence.
The Trust Baseline compiles a representative slice of your corpus and issues your first certificate in 48 hours, no integration work on your side. Or start smaller: a supersession compile or staleness sweep proves the mechanism in a single session.
When AI answers depend on evidence that changes.
AI evidence verification in practice: six ways the confidently wrong answer shows up, ten industries where it costs the most, and the entry engagement that proves the fix in each. First market: clinical and medical evidence, deliberately.
The one-question filter.
Does a confidently wrong answer cost you more than a fast answer is worth?
An apology
If a wrong answer costs you an apology, you are not a customer yet, and we will tell you so.
A consequence
If it costs you a reportable event, a rejected filing, a nonconformance, or a client's trust, find your pattern below.
Most organizations that buy recognize themselves in more than one.
The silent conflict
Your sources disagree, and the AI answers anyway.
"Our sources don't always agree, and the AI doesn't tell us when they don't. It just answers."
Evidence bodies conflict by design: studies diverge, authorities disagree, guidance contradicts guidance. A system that averages the disagreement away has quietly made an editorial decision nobody reviewed. Triplets makes the disagreement itself visible: every question resolves to supported, refused, or conflicted, with a verdict on which source governs. The conflicted state is the demo centerpiece, because it shows you decisions you did not know were contested.
- Where it leads
- Clinical and medical evidence, where making disagreement visible in GRADE-aligned vocabulary is native to how reviewers already think; legal research across conflicting authorities; regulatory filings; multi-source catalogs.
The re-answered question
Your team re-answers the same questions to check nothing changed.
"We ask the same questions every week and redo the work every time. Nobody knows if last month's answer still holds."
Research, credit, underwriting, and procurement teams run on cycles: the questions barely change, the answers do, and the re-checking is unbudgeted work that exists only because nothing is watching. Triplets keeps questions alive as monitored decisions. Last month's supported answer is still being watched; your team checks the monitor instead of redoing the analysis.
- Where it leads
- Financial services research, credit review, and model-adjacent workflows; underwriting; procurement. The sell here is recognition, not education, because this is already how these teams think about their work.
The stale answer
The rules changed, and your AI kept giving the old answer.
"Something changed upstream, and we found out from a customer or a rejection, not from our system."
When an external authority owns the ground truth, regulations, filed rates, payer policies, formularies, standards, every downstream answer goes stale on someone else's schedule. Insurance is the sharpest case: advisory bureaus produce on the order of twelve thousand discrete changes a year, and in most states a filing an insurer takes no action on applies automatically. The enforcement record shows the cost: carriers cited for calculating on superseded rates, and in one case a single undetected rating error ran three years and ended in more than $18 million of restitution, because detection, not judgment, was the failure. Triplets sells the protected state: continuous recompilation against the change stream, and an alert that a supported answer became unsupported before it costs you.
- Where it leads
- Insurance underwriting and filings, compliance and regulatory change management, payer policy, standards-driven manufacturing. The staleness sweep is the proof: point it at one corpus and see which standing answers already went stale.
The silent supersession
The current version and the obsolete version both answer questions.
"The right answer and the obsolete answer both live in our system, and the AI can't tell which one is which."
Every document management system holds amended contracts next to originals, revised SOPs next to superseded ones, replaced SKUs next to current ones, all with equal retrieval weight. Which version answered your last question? Triplets treats version history as structure: every answer names the exact version it came from, and a document that conflicts with its own replacement surfaces as conflicted.
- Where it leads
- Legal precedent and playbook corpora, manufacturing SOPs and specs where answering from a superseded revision is a nonconformance, pharma protocol and SOP control, e-commerce catalogs answering customers from replaced SKUs. This is also the fastest proof in the framework: a supersession compile runs on your own corpus with no external data and shows a conflicted partition in the first session.
The undefended answer
Your clients want you to show your work.
"Our clients don't just want the answer. They want to defend it to their stakeholders, and right now that defense is our reputation and a PDF."
If you answer on behalf of someone else, a sponsor, a payer, a client, the deliverable is not the answer; it is the reproducible artifact behind it. Triplets emits a certificate: same corpus, same questions, same partition, every time, re-runnable by the client's own reviewers. And refusal strengthens the position rather than weakening it, because a partner who sometimes says the evidence does not support a claim is more credible to the sponsor, not less.
- Where it leads
- CRO and HEOR firms delivering evidence work to sponsors, medical affairs answering payers and HCPs, consultancies of every kind whose product is defensible judgment. When a payer challenges your dossier, the certificate is what you hand them.
The audit you cannot replay
Someday, someone will ask why your system said that.
"If a regulator or auditor asks why our system said X eighteen months ago, reconstructing that today is a project, and the reconstruction might not match."
Audit retrospection is adversarial: the question arrives after the fact, about a state that no longer exists. Triplets makes it a lookup. Corpus versioning means you recompile against the evidence exactly as it stood on that date and get the same partition, and the provenance graph means every answer carries its own evidence trail.
- Where it leads
- Pharmacovigilance and regulatory affairs under inspection, financial services model risk, quality systems facing audits, market conduct exams in insurance. The buyer is whoever owns the risk, and model AI guidance from regulators now names exactly these properties: reproducibility, traceability, version control, data lineage.
Where each industry starts.
One pattern leads per industry, chosen by the failure your team has already lived through. The rest arrive in the second conversation.
| Industry | The failure you have already seen | Start with |
|---|---|---|
| Clinical and medical evidence | Sources that disagree, answered anyway | Trust Baseline with staleness sweep |
| HEOR and medical affairs | A challenged dossier defended by reputation | Certified evidence corpus |
| Pharmaceutical | An inspector asking about a past answer | Baseline on SOPs or safety docs |
| Legal | A superseded clause that got cited | Supersession compile on your precedents |
| Financial services | The same review redone every cycle | Baseline on one recurring decision workflow |
| Insurance | A rate change discovered by rejection | Staleness sweep on one product line |
| Compliance and risk | Regulatory drift found manually | Staleness sweep on the obligation register |
| Manufacturing and quality | An answer from a superseded revision | Supersession compile on one line's documents |
| Healthtech | Governance reviews stalling your deals | Design-partner engagement on your evidence layer |
| E-commerce catalog | A replaced SKU answering customers | Supersession compile on catalog and policy |
A note on the last row: catalog accuracy passes the filter only when a wrong answer genuinely costs more than a fast one. If it does not in your case, we will say so before you spend a cycle.
Where Triplets does not belong.
The boundary
A product whose premise is refusing to overclaim should be precise about its own boundary. Wrong tool for: narrative and creative content, where there is no verifiable fact to constrain against; math and code generation, which need different correctness machinery; real-time streams and very short or non-text corpora. And the qualifier cuts both ways: conflict density, change rate, and consequence of error are what make a fit, not data volume. A small corpus that changes weekly is a better candidate than a huge static archive.
Why certification renews
The certificate was true the day it was issued. The evidence moved the day after. Keep the asset alive, or keep a photograph: the renewal logic is structural, not commercial, and it is the reason "continuous" is in the category name.
The fine print that is actually load-bearing. Patent claims filed across 18 provisional families, Patent Pending. Triplets is built on the Enigma Vault, which is PCI DSS Level 1 and SOC 2 Type II. The certificate proves support within your compiled evidence, reproducibly; it does not claim truth, and we will say that before your lawyer does.
Bring us the decisions you cannot afford to get wrong.
Start with the proof that fits your pattern: a supersession compile in one session, a staleness sweep on one corpus, or the full Trust Baseline with your first certificate in 48 hours. No integration work on your side, and the report is yours either way.
Your questions about Triplets, answered.
Evidence and the Trust Certificate
What is Triplets?
Triplets is the compiler behind Continuous Trust Certification, built by Enigma Vault. It compiles your evidence into traceable, certified knowledge, monitors whether standing answers remain supported as evidence changes, refuses questions the evidence cannot support, reconciles conflicts, and produces a reproducible certificate showing what is supported, what is not, and why.
What is Continuous Trust Certification?
A category of assurance for AI systems and evidence bases. Instead of a one-time benchmark or audit opinion, it is a fingerprinted, re-derivable certificate over compiled evidence, kept green by continuous monitoring, with an alert whenever a standing decision's support changes. Five buyer-runnable tests define it: refusal, state, memory, conflict, and audit.
What does the Trust Certificate prove?
Support within your compiled evidence, reproducibly. It does not prove truth, medical correctness, or regulatory compliance, and we state that plainly. It shows what your evidence supports, what the system refuses and why, and where sources conflict, in a form a third party can re-run and verify independently.
Why can't we just prompt the model not to hallucinate?
Because fabrication is the mechanism, not disobedience. A prompt is a request, not a constraint, and the model has no ground truth to check itself against. Confidence does not track correctness, fine-tuning cannot set a floor of zero, and temperature zero makes output repeatable rather than correct. A boundary has to be enforced by a system outside the model that knows what is verified. That system is the compiler.
How does Triplets answer a question?
Through a hierarchy of methods over the compiled state, strongest first. Deterministic lookup answers exact facts from verified values with no generation. Relationship traversal answers questions that span entities and documents. Constrained generation handles open questions with the model bound to the compiled evidence. Past the boundary, Triplets refuses, explains why, and logs the gap.
What are the accuracy numbers?
In design-partner-stage evaluations across four domains, Triplets reached 88 to 94 percent pass plus partial. Those figures carry that qualifier deliberately, and the Baseline measures the numbers that actually matter on your own corpus. We do not publish accuracy claims without a refusal boundary attached, because an accuracy number without a boundary means confidently wrong at a known rate.
Is the certificate aligned with clinical evidence standards?
Yes, deliberately. In clinical use, certificate vocabulary aligns with the evidence-hierarchy language reviewers already work in, GRADE-style, and the certificate format is designed for methodologist review. It should read as an extension of your evidence discipline, not a rival to it.
Is the technology protected?
Yes. Patent claims have been filed across 18 provisional families, and the product is Patent Pending. Triplets is built on the Enigma Vault, which is PCI DSS Level 1 and SOC 2 Type II.
Changes, conflicts and version history
What happens when the evidence changes?
The compiled state is compared continuously. When a label updates, a finding is superseded, or a source is corrected, Triplets re-evaluates the standing decisions it touches and alerts you when, and only when, a decision's support changes. That is the difference between a certificate and a photograph of one.
How does Triplets handle conflicting sources?
It surfaces the conflict rather than smoothing it over. When two sources disagree, Triplets records the disagreement, determines which source governs and why, and shows its work in the certificate. Most systems silently serve whichever chunk ranked highest.
How is this different from an assurance audit?
An audit is an opinion, and it is true for exactly one day. The Baseline produces a fingerprinted compiled state that is re-derivable by a third party, re-runnable by your regulator, and kept green as the evidence changes. One is a signature. The other is surveillance.
Won't the model vendors add this?
Longer context windows make rereading cheaper. They do not decide source precedence, preserve claim states, or refuse on your evidentiary standard. And a model vendor's promise to its market is coverage, so a certificate of refusal issued by the vendor being judged would not be credible. Certification has to be neutral.
Does Triplets replace our vector database, graph, or retrieval stack?
No. Triplets is substrate-agnostic. It re-ingests your corpus into its own structured model and governs answering at query time, alongside whatever you already run. Adopting it is not a migration, and in embedded deployments it certifies the corpus your existing tools answer from.
Where is Triplets not the right tool?
Narrative and creative content, math and code generation, real-time streams, and very short or non-text corpora. Each needs different correctness machinery. If your corpus is a poor match, the Baseline shows it before you commit further.
Getting started and pricing
What is the Trust Baseline?
The first compilation of your corpus, delivered in 48 hours. You pick a representative slice and your hardest questions. Triplets compiles it, runs your questions side by side against your current stack, and issues your first certificate: supported, refused, conflicted, fingerprinted. It creates the compiled asset that continuous certification keeps alive, and it is scoped as a fixed-price engagement; request a proposal from the pricing page.
Can we fail the Baseline?
No. Nobody fails a Baseline. Its output is a partition, supported, refused, and conflicted, which is a map of your evidence boundary, not a grade. Every refused question is a corpus-expansion decision: add the evidence, recompile, and it becomes answerable.
Our platform team could build this. Why shouldn't they?
They can start. Most internal knowledge-graph efforts stall before production on entity resolution, conflict handling, and quality at scale, which are exactly the steps Triplets automates. There is also a structural issue: a trust certificate issued by the organization it certifies is not credible to an external auditor. Neutrality is part of the product. The Baseline runs on your corpus either way, and the report is yours.
Who is Triplets for?
Organizations that cannot afford confident-wrong synthesis or stale decisions. First: clinical and medical evidence. Medical-information teams, health-system AI governance committees, digital-health product leads, medical affairs, HEOR, and regulatory teams at pharma and CROs, and clinical AI vendors who need an embedded certification layer.
How is it priced?
Triplets engagements are scoped to your corpus, entity volume, and engagement model: the 48-hour Trust Baseline, a certified evidence corpus for a therapeutic area, or embedded certification under your brand. Because scope drives the number, pricing starts with a conversation rather than a rate card. Tell us which decisions you cannot afford to get wrong and we will come back with a scoped proposal, typically within two business days. Talk to us about pricing.