Deep where it counts.
One system, three configurations. The agents differ, the jobs differ, the boundary differs — the record layer underneath is the same one, and it ships with all of them. They are staged, not parallel: aviation security leads, and each panel below says exactly where it stands. We are looking for one design partner per industry. What follows is the design and the limit, not client results.
Analyst and field research on agent systems at large; predictions are labeled as such. None of it is a Khelion result — ours go to the Lab when they are real.
Manifest for aviation security
Operator compliance · deployment codename AirsideScreening is already a human-supervised detection system, and the machine is the part that works. What fails is everything around it: background checks that must be reviewed continuously or repeated every twelve months for every pass holder, recertification windows that drift, competency that degrades unnoticed between audits, a security programme that has to absorb every new regulation, an inspection file assembled in panic. Manifest holds that perimeter and never touches the screening decision.
Competency drift
Reads the Threat Image Projection results the operator already has to keep. Week 33: 1,284 projections across six lanes; the night shift on lane C falls from 87.1% to 79.8% over six weeks — a −2.4σ drift that monthly averages hide. A targeted plan goes to the supervisor. It never writes into the TIP system, and a TIP score is never a sanction.
Credentials and passes
Holds the whole population of pass holders against the rule that applies: background check reviewed continuously or repeated inside twelve months, pass valid five years at most, withdrawal immediate on failure or non-return. Shows what lapses next and prepares the file. A person with authority decides about a person.
Training and recertification
Tracks every recertification window and recurrent-training date against rest and rotation limits, and flags the week-38 clash seventeen days before it becomes a finding. It delivers no approved training and certifies nobody — that belongs to the authority.
Regulation into your programme
Each new implementing regulation is compared clause by clause against your own security programme: what changed, who owns it, by when, and the draft amendment. Built to work without classified material — that stays with the cleared human.
Inspection and corrective actions
An evidence pack that is always current, findings tracked to their corrective action plan and closure, every claim linked to the record behind it. It refuses to produce evidence it cannot source.
Occurrence reporting becomes mandatory
Commission Regulation (EU) 2026/247 of 2 February 2026 replaces Annex II of Regulation (EC) 300/2008. From 1 January 2028 every Member State must run a process for reporting, classifying, processing, storing, protecting, analysing and aggregating aviation security occurrences, with a confidential channel and a common classification. It creates a job most operators do not have yet.
To: Head of Screening Ops Re: Lane C night cohort — retraining proposal (W34) Detection on partially occluded categories fell from 87.1% to 79.8% over six weeks (n=1,284). Two 45-min refreshers proposed, off-peak, W34. Sources: led_48188 · led_48194 · led_48201 [ Draft — requires review and signature ]
Manifest for medical devices
The regulatory file, produced by agents · codename ContinuumA device manufacturer owes the authorities a stack of documents that never stops: periodic safety update reports, post-market surveillance reports, clinical follow-up literature reviews, trend analyses. The work is multi-source assembly with citation discipline — read everything, attribute every claim, write it so a notified body cannot knock it down. It is exactly what an agent system does well and exactly what a human has to sign. So we sell the signed file, not software for your quality system.
PSUR and PMS reports
One device family, one report, one fixed price. Agents pull complaint data, vigilance records, registry and literature sources, and draft to the structure the notified body expects. Every sentence carries its source. A regulatory professional reviews and signs. We are contracting that signature rather than pretending to hold it: today it comes from a named partner, disclosed to you before you sign anything.
Article 88 trend analysis, defensible
The regulation makes you predefine the method and the observation period. Most companies do it in an unvalidated spreadsheet, and notified bodies attack exactly there: undocumented competence, unvalidated calculation, method misapplied. We deliver the method dossier — chosen method justified, thresholds argued, calculation reproducible and versioned, refreshed each quarter.
PMCF literature review
The search protocol, the screening log, the appraisal, the write-up. Reproducible: run the same protocol next year and the delta is visible instead of being re-litigated from memory.
Complaint backlog, cleared
A one-off engagement on a backlog that has grown past manual triage: normalise the intake, de-duplicate, code, cluster and order the queue with every decision traceable. The algorithm orders the queue. It never decides what belongs in it.
We are deliberately not selling you software
Putting a tool inside your quality system triggers software validation on your budget, a supplier quality agreement, an audit right, and a procurement grid that a young company fails on financial-stability criteria alone. AI-native vendors already sit in that seat and we are not going to out-muscle them there. A signed deliverable is a purchase order — and it puts our work where it can be judged: on the file itself.
# written before the data, not after series: complaints / 10k exposure-days, by lot method: CUSUM + Bayesian change point · 2 of 2 agree period: quarterly, justified in writing competence: named reviewer, qualification on file validation: calculation versioned and reproducible flagged: lot L-2291 · day-3 removals +38% not_flagged: 6 apparent spikes — matched to a sales campaign, not to the device
Manifest for insurance operations
The document back office · codename BinderA wholesale broker receives statements, membership files and claim documents from hundreds of partner offices, in as many formats as there are senders. Somebody re-keys them. That re-keying is the cost line — not underwriting, not pricing, not advice. Manifest turns incoming documents into checked, structured data in the system of record, and stops at every point where a human has to decide.
Statements, one schema out
Forty formats in, one checked file out, every morning. Amounts reconciled against the contracts, missing lines named rather than guessed, discrepancies raised with the partner office that sent them. The unit is a file processed, and you can count it.
Membership and enrolment intake
Incoming enrolment files read, required documents checked off, what is missing requested automatically with the right reference. The handler opens a complete file instead of building one.
Claim correspondence
Inbound mail sorted by what it actually asks for, attached to the right file, prepared for the handler. Anything touching a refusal, a liability question or a vulnerable customer stops and waits for a person.
Delegated-authority reporting
If you manage business on someone else's paper, you owe them a report in their format on their schedule. Supervisors have said plainly that these reports arrive inconsistently. We produce yours from the work itself, not from a quarter-end scramble.
Two things we no longer sell here
We had planned a third-party register for European operational-resilience rules and an AI inventory. The register does not apply to insurance intermediaries below 250 staff and €50m — most brokers are legally out of scope, and selling it would have meant selling an obligation they do not have. The AI inventory matters, but from December 2027 and to insurers, not to brokers today. Both dropped.
# measured on your own team, before we start today: 2 people · 3 h each morning · re-keying errors: caught at settlement, not at intake after: one reviewed file at 06:50, exceptions only we_charge: on files processed, not on seats # if the count does not move, neither does the price
Most of the work worth automating is specific to one company and appears in no catalogue. We take one process you already measure and engineer the system around it. That is the bespoke route, and it is at the bottom of this page.
Your own process.
Most of the work that exhausts a team is specific to that team: a reconciliation nobody outside the company would recognise, a dossier assembled from six systems, a watch someone runs by hand every Monday. No vendor sells that. We build it, with Manifest underneath from day one.
Two weeks inside the process
Not a workshop. We watch the real work: what gets opened, what gets copied, where the judgement happens, what the person is afraid of getting wrong. Those two weeks are the paid scoping.
The warrant before the agent
What it may do, what it may never do, when it stops and asks. Agreed in writing at this point, not discovered in production.
Built on your own past cases
Including the awkward ones. If it cannot beat your current process on those, we say so and stop.
You keep the parts
Warrant, evaluation set and log are yours from day one. Handover of the configuration is a fixed fee, agreed at signature.
How it is priced
A paid scoping fortnight, deducted in full from the build. The build is quoted after scoping, in writing — and the quote holds. Manifest is included from day one: we do not ship an agent we cannot account for.
# what the week produced process: "weekly supplier-risk review" owner: "Head of Procurement" today: 6.5 h/week, 2 people, 3 systems judged_on: suppliers flagged before an incident today's_score: 4 of 11 caught in the last 12 months autonomy: act_on_approval human_gate: any contact outside the company kill_rule: below 8 of 11 at month 3, we stop and refund the last month verdict: build it # the other verdict exists
Sometimes the answer is a spreadsheet
Some weeks end with “you do not need an agent here — here is the two-day fix instead”. You keep the written brief either way. Charging for the week is what lets us say it.
Programme panels illustrate the design, the boundary and the reporting format. No client has run these yet; they are not client results. Real numbers go in the Lab once they are real.