How we govern the AI that does your work.

Hedge · consulthedge.ai · version 7 · effective 8 August 2026

Published by: Hedge Labs for Project Management L.L.C., trading as Hedge AI, operating the Hedge AI platform at consulthedge.ai and serving customers worldwide. Contact us at support@consulthedge.ai.

At a glance

Hedge AI sells finished work produced by AI systems. That sentence carries obligations. This policy states them — publicly, so they can be checked against how the platform actually behaves. It applies to every customer, in every country.

• Every piece of work is checked by a system separate from the one that produced it, against an enforced quality floor, before you see it.
• A human always holds the consequential actions. The system cannot spend money, publish, push to a main branch, or change its own rules without a person's approval.
• Every task is on the record — what was asked, what was used, what was produced, what it cost — in a log nobody can alter (the content in it stays deletable, as the Privacy Policy explains; the minimal event record remains).
• Your work does not teach our models unless you explicitly say yes. Anonymous records of how the platform ran, and material stripped of anything identifying anyone, do.
• Deciding is excluded on purpose. The platform prepares, produces and checks; the judgement and the decision stay with you.

1 · Purpose and scope

This policy sets the rules under which AI systems are selected, deployed, operated, monitored and retired across the Hedge AI platform. It applies to every AI system in the product, everyone at Hedge who operates one, and every third-party model we route work to — worldwide, without a weaker version for any market.

Its purpose is to make responsible AI a property of the system rather than of anyone's good intentions: to protect customer data, to prevent unchecked output from reaching a customer, to keep a human accountable for consequential actions, and to keep our use of AI lawful as the rules evolve.

It is one of three public documents that work together: the Terms of Service (the contract), the Privacy Policy (personal data), and this policy (how the AI itself is governed).

2 · Our principles

1. Human accountability. AI produces; people remain responsible. Consequential actions — spending money, publishing publicly, changing the rules an AI specialist works under, pushing to a main branch — always stop for a human decision. The system cannot authorise itself.
2. Independent verification. The system that writes is never the system that marks. Every delivered task is reviewed and scored by a separate system against an enforced quality floor; work below the floor is repaired and re-checked, not shipped.
3. Transparency about what was produced. Every task records what was asked, what was used, what was produced and what it cost, in an append-only log that cannot be altered — the content within it remains subject to your deletion rights (Privacy Policy, Section 9); the minimal event record remains. Work delivered by Hedge AI is AI-produced by design — that is the product, stated openly, never disguised as human work.
4. Data protection by design. Customer data is separated at the point it is read or written. Connecting a source never widens who can see a file. Consent to use data to improve the service is explicit, off by default, and revocable.
5. Fairness. AI systems inherit the bias of what they were trained on. Output is independently reviewed before delivery, our prohibited-use rules (Section 6) ban discriminatory applications outright, and material that fails our quality bar is never used to teach anything — a bad input cannot quietly become a lesson.
6. Safety and robustness. Execution runs in isolated, hard-capped environments that hold no platform credentials and cannot outspend their budget. The worst case is bounded by architecture, not by promises.
7. Lawful use. Every model is licence-checked for commercial use before it enters the engine. Our obligations under the data-protection and AI laws of the markets we serve are treated as a floor, not an aspiration.

3 · How Hedge AI produces work — what we tell you, and what we keep confidential

What we disclose. Work is produced by a coordinated team of AI models: a production team drafts, a judge merges, an independent checker scores the result out of 100 against the enforced floor of the effort level you chose (floors run from 70 to 95), a fixer repairs what falls short, and a re-checker seals it. An additional audit layer re-examines finished work outside the production path. For code, tests are the judge: the work runs, and it passes or it does not.

What we keep confidential. Which models write, which check, in what order, against which recipe. You buy an outcome standard, not a model; the engine re-selects models continuously and is vendor-neutral by design. The score on your work is always a score the pipeline actually produced — presenting an invented score is a prohibited act under this policy.

4 · Human oversight

• Guarded actions. Spending money, publishing to the outside world, altering an AI specialist's operating rules, and pushing to a protected branch require an explicit human approval, every time. Approval rights are per-action and revocable.
• Spend controls. Every AI system runs under a spend cap that warns before it stops. A runaway system exhausts its own budget inside a logged workspace — nothing more.
• Reversibility. An AI specialist's configuration can be rolled back in one step. Sandboxes are destroyed after each run.
• The excluded class. Deciding, and owning the decision, is deliberately not built. The platform will draft the recommendation and show its reasoning — the call, and accountability for it, stay with the customer.

5 · Classification and risk management

Every AI system in the platform is classified by what it can affect, and controls follow the classification:

Risks are assessed before a system ships and re-assessed when it changes. Accuracy and reliability are tested continuously through the scoring pipeline on live work, not at a single point in time. Where a customer's use case is itself regulated (legal, financial, audit, medical, engineering), the Terms of Service require qualified human review on the customer's side before the work is relied on — the platform's checking raises quality; it does not replace professional responsibility.

6 · Acceptable and prohibited use

Acceptable use is producing, analysing, drafting, coding, reviewing and reporting on material the customer is entitled to use, inside the customer's own workspace, with the outcome independently checked before delivery.

Prohibited without exception — for us and for anyone using the platform, in every country:

• Using customer data to train shared models without explicit, freely given, revocable consent.
• Allowing an AI system to authorise its own spending, publication or rule change.
• Delivering unchecked output to a customer as finished work.
• Presenting a score the pipeline did not actually produce.
• Using AI in ways that, by objective or by effect, discriminate against people, manipulate or deceive them through subliminal or purposefully deceptive techniques, or exploit vulnerability arising from age, disability, or social or economic situation.
• Using AI to categorise people by protected characteristics — including inferring race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation from biometric data — or to score people socially.
• Using AI to assess the risk that a person will commit a criminal offence based on profiling or personality traits.
• Using AI to build or expand facial-recognition databases through untargeted scraping of facial images from the internet or CCTV.
• Using AI to infer people's emotions in workplaces or educational institutions, except for medical or safety reasons.
• Using AI for real-time remote biometric identification of people in publicly accessible spaces.
• Using AI to create or spread non-consensual intimate imagery or child sexual abuse material.
• Using the platform to produce content that is unlawful.
• Introducing a model whose licence does not permit commercial use.

Transparency of AI-produced work. Work delivered by this platform is AI-produced and sold as such. Where the law that applies to you or your audience requires AI involvement to be disclosed, you must not present the work as human-authored — see also Section 12.

Violations by customers are handled under the Terms of Service. Violations inside Hedge are handled under Section 11.

7 · Data used by AI systems

• AI systems read customer content to perform the task asked — grounding work in the customer's own material is the product working as described, not a secondary use of data.
• Training on identifiable customer work is opt-in only. The "improve the service" consent is a separate, off-by-default, revocable switch (Privacy Policy, Section 4). Without it, a customer's work teaches nothing. Where we offer credits for switching it on, the platform is unchanged for anyone who declines, and nothing is clawed back from anyone who later switches it off.
• What we learn from without asking, and why it is different. Operational records of how the platform ran — effort levels, scores, retries, timings, costs, failure patterns — describe the machine, not the customer. Separately, material that has been irreversibly stripped of everything identifying a customer, their business or any person is no longer about anybody, and we use it to train, tune and evaluate. The engineering standard is the control here, not the wording: de-identification is verified before material enters a training set, and material that fails verification is treated as customer content, not as anonymous.
• The quality gate protects learning. Material that fails the quality bar on the way in is stored for reference only and never used to teach anything.
• Third-party model providers process customer data solely on our instructions to perform tasks; we do not permit them to use customer content to train their models.

8 · Third-party models

Models enter the engine only after a documented licence check confirming commercial use is permitted, and only with a recorded approval carrying a name (Section 9). Providers are bound to process data on instruction only. Because the engine is model-agnostic, any provider can be replaced without changing what customers were promised — the quality floor is Hedge's commitment, not a vendor's.

9 · Governance — who is accountable

• The Chief Executive Officer holds final accountability for AI governance: which models enter the engine, the quality standards they are held to, and the design of every guarded action. No model is added, swapped or removed without a recorded approval carrying a name.
• The platform's own quality machinery — an independent auditor, reviewer and advisor layer — is structurally separate from the systems that produce work and cannot be overridden by them.
• Every AI system in the product has a named operational owner responsible for its behaviour, monitoring and incidents.
• As the organisation grows, governance duties will be delegated to named roles under this same policy; accountability at the top does not move.

10 · Monitoring, audit and incidents

• Output is scored on every task. An independent audit layer re-examines finished work outside the production path; its cost is built into the price of every tier, not sold as an extra.
• Spend is monitored per system against caps that warn before they stop.
• Findings that recur become a change to the system — a rule, a recipe fix, a new test — not a note in a file.
• Incident handling: when an AI system produces something harmful, materially wrong, biased, or handles data in a way it should not, the system involved is suspended, the cause is found and fixed, and affected customers are informed of what happened and what changed.

11 · People, training and consequences

Everyone who operates an AI system at Hedge is responsible for knowing this policy and the limits it sets, and for keeping current as it changes. Understanding what these systems cannot be trusted to do is a job requirement, not an optional briefing.

Breaches of this policy are treated as seriously as any other breach of trust: the system involved is suspended, the cause is found, affected customers are informed, and where a person is responsible, ordinary disciplinary and legal consequences follow.

12 · Your responsibilities as a customer

The platform's checking raises the standard of the work; it does not transfer your professional obligations to us. When you use Hedge AI:

• review output before you rely on it, publish it or file it;
• keep a qualified human in charge where your field's rules require one;
• disclose AI involvement where the law or your professional rules require disclosure;
• and use the platform only within the acceptable-use rules of Section 6 and the Terms of Service.

13 · Reporting a concern

If you believe an AI system on this platform has produced something harmful, biased or materially wrong, or has handled data in a way it should not, tell us at support@consulthedge.ai. Reports are investigated, reporters are protected from retaliation, and where a customer is affected we say so.

14 · The rules we watch

AI regulation is moving, and it moves at different speeds in different countries. We track the laws and frameworks relevant to every market we operate and sell in — including the European Union's AI regulation, the data-protection laws of the markets we serve, and international AI risk-management standards. We treat new obligations as a floor to build to, not a ceiling to argue with, and we apply the strictest applicable standard across the platform rather than running a different product per country. Where a rule requires a change to the platform, the change is made in the system, not in the wording.

15 · Review

AI moves faster than policy. This document is reviewed at least annually, and immediately when something material changes — a new class of system, a new legal obligation, or an incident that teaches us something. The Chief Executive Officer owns that review. The date at the top is the last time it happened.

Hedge Labs for Project Management L.L.C., trading as Hedge AI · support@consulthedge.ai