tcp/ap · the agentic protocol

Agentic AI you can trust.

Reproducible & auditable decisions

Every decision is logged, SHA-256 fingerprinted and auditable.

micp · model inference control protocol

Governed conversations.

Chatbots fail two ways: they agree with everything, feeding delusion and dependency, or refuse and shut down when someone needs them most, and they drift there turn by turn. MICP is an alignment control plane: it follows the whole trajectory and aligns every reply with your declared policy before the chatbot user sees it, without ending the conversation. Same situation, same outcome: nothing that breaks your policy reaches the user, and every reply that does is signed, logged and auditable.

Demos

1 live · 1 soon

Companion chatbot

Sycophancy is an actively researched failure of chatbots: they agree with vulnerable users until they reinforce self-harm or share their delusions. Safety training pushes the other way, into over-refusal: the chatbot shuts the conversation down when someone needs it most.

Detect and block sycophancy early, while balancing against over-refusals, in real time.

→

Customer service chatbot

In 2024 a Canadian tribunal held Air Canada liable for a refund policy its chatbot made up; a car dealer's chatbot was talked into agreeing to sell a new SUV for $1.

Offer only what the policy allows, the same for the same facts, whatever the customer types.

soon

tcp/ap · trusted cognition protocol / agentic protocol

Verifiable decisions.

AI agents fail quietly: the same case gets different decisions, and nobody can show whether a decision was right. The hard part isn't adding guardrails that block bad actions; it's ensuring, at runtime, that every decision follows your rules, and proving it afterwards. Same facts, same decision: TCP/AP makes agents deterministic, and every decision is logged, fingerprinted and verifiable.

Demos

4 live · 9 soon

Clinical scribe

In a 2026 study of AI scribe notes, physicians found omissions in 18% and invented content in 11.5%; 5.3% had errors posing serious or imminent risk (UC Davis, JMIR Medical Informatics).

Prevent invented facts and flag uncertainties before the final journal note is produced.

→

Travel agent

In TravelPlanner, a 2024 benchmark of AI travel agents, GPT-4 Turbo with tools completed only 0.6% of trip plans that met every constraint (ICML 2024).

Make sure the traveller gives every detail needed, then book exactly what they asked for.

→

HR Agent

In a 2024 University of Washington study, AI resume screeners favoured white-associated names 85% of the time and female-associated names only 11% (AIES 2024).

Screen every CV the same way and only by your hiring policy, never on name, gender, age or nationality, with the reason on record.

→

NDA Agent

General-purpose chatbots hallucinate on 58–82% of legal queries, and even AI legal research tools built for lawyers made up information in 17–33% of answers (Stanford, 2024).

Never invent a position: review every clause against your playbook, draft new NDAs only from your own wording, and release nothing without a lawyer’s approval.

→

SDR Agent

Sales agents invent facts about the prospect, and email people who should not be emailed.

Trace every claim in the email to a source, and send nothing your rules do not allow.

soon

Claims triage

Claim bots pay what they should not, or bounce what they should pay.

Check every claim against the policy terms, and pay only what the policy covers.

soon

Invoice approval

Invoices get paid with wrong amounts, or to a bank account that quietly changed.

Match invoice, order and delivery, and pay only what was ordered and delivered.

soon

Loan pre-approval

Lending models decline without a reason, and treat the same applicant differently.

Apply your lending rules the same way every time, and explain every decline.

soon

KYC onboarding

Onboarding checks miss sanctioned names and let incomplete files through.

Check every document and name against your onboarding rules, the same way for every customer.

soon

Clinical trial eligibility

Screening tools guess at criteria the patient record does not answer.

Check each criterion against the evidence, and never guess what the record doesn't say.

soon

Content moderation

Moderation bots remove fine posts and miss harmful ones, with no reason given.

Check each post against your written policy, the same way every time, with the reason on record.

soon

Vendor check

New suppliers get approved without the checks the company requires.

Check sanctions, certificates and spend limits before anyone can buy.

soon

Sensai

Coding agents drift from the task's rules and may say “done” when it isn't.

Hold the agent to your rules, and check every “done”.

soon