Fide AI

Our approach to RL environments

RL environments should come with evidence.

Reinforcement learning environments now shape what AI agents learn to do. Fide builds, extends and reviews them for labs, and each comes with the evidence a team needs before training: how the reward was checked against expert reviewers, which shortcuts were tried against it and what was kept sealed for testing.

Our point of view

Environment quality, not environment count, decides what a model learns.

High reward should mean the task was done.
Agents find the shortcuts a grader misses: editing tests, reading answers, claiming work they did not do. An environment that pays for a shortcut trains the shortcut, at the full cost of the compute.
Review has to keep pace with scale.
Public hubs now list thousands of environments, and few publish how their rewards were checked. A team should see that evidence before it trains on an environment.
Finishing the task is not the whole job.
Agents deployed into real work have to accept a stop or a correction, stay within the permissions they were given and report honestly. Few environments put people with authority into the task.
Deference should never beat doing the job.
Rewarding caution on its own teaches agents to refuse. Useful work has to earn more than blanket refusal, or an environment trains timidity instead of judgment.

Where we are going

The judgment agents need most lives outside the lab.

Universities, medical and legal bodies, research programs and faith communities have spent decades defining good judgment and human flourishing in their fields. Little of that work reaches a reward function. We are building toward environments made with the institutions that hold this judgment, so labs can train and evaluate agents on the work that matters most to people.

Expert frameworks, made trainable
We translate a framework into tasks, proxies and rewards, and calibrate the grader against the experts who hold it. The institution’s judgment sets the bar, not ours.
Many frameworks, side by side
Where serious institutions disagree about what doing well means, we keep their frameworks separate by default rather than blend them, so a lab can see how a model performs under each.
Held-out sets that stay sealed
Evaluation sets are kept out of every training release, so results keep their meaning. Independent results for models trained on our environments come from an outside evaluator or host.

The standard we hold ourselves to

Every environment ships with a validation report.

The report shows how the reward was checked, so a team can inspect the evidence before training on it. An environment is released only when its report allows it. Before then, it can be shared as a preview under agreement, with every gate it has not yet met listed in its report.

Reward validity
How often the grader agrees with qualified reviewers, with false-pass and false-fail rates for each behavior the reward scores. A behavior whose grading stays below its agreement floor earns no reward.
Exploit testing
Degenerate policies, replayed reward hacks and prompt-injection probes run against every release. An unexplained exploit blocks release.
Sealed held-out tests
Test families created before a buyer has access, kept apart from training families and marked with canaries. Held-out sets are never sold for training.
Difficulty
Pass rates by model against a stated target band, with tasks retired or fixed when they saturate or break.
Transfer and baselines
Held-out families and changed conditions, compared with baselines that involve no training on the environment.
A reproducible record
The report regenerates from its inputs. Where it is signed, the signature shows the report has not changed. The methods and evidence, not the signature, support the findings.

What a report looks like

Illustrative · invented values

Environment
Invoice-approval assistant (fictional)
Status
Preview: 2 of 6 release gates not yet met
Label
Development validation, prepared by the builder

Reward validity

BehaviorAgreement with reviewersFloorFalse passFalse failRewarded
Completes the approval correctly92%85%3%5%Yes
Stays within spending authority88%85%4%8%Yes
Stops when a manager intervenes79%85%6%15%No, below floor
Reports what it actually did90%85%2%8%Yes

Exploit testing

14 degenerate policies scored below the reference solution. 3 replayed reward hacks scored zero. The learned exploiter found 1 shortcut, explained and fixed before this version.

Held-out tests

Training and held-out families are disjoint and carry canaries. Sealed sets were fixed before any buyer had access.

Outside evaluation

Fide’s own measurements of models trained on this environment are development validation. Independent results on the capabilities it targets need an outside evaluator.

Release gates

  • Every rewarded behavior at its agreement floorNot met
  • Shortcuts score below the reference solutionMet
  • No unexplained exploitMet
  • Difficulty within the target bandMet
  • Training rights completeMet
  • Outside review requestedNeeds a person
Illustrative excerpt. The environment is fictional and every value is invented; these are not results. A full report also covers difficulty by model, transfer and baselines, monitorability, cost and the complete run record.

What our environments teach

Agents you can deploy, not only agents that finish.

Fide environments are designed to put people in the task: users with authority, decisions reserved for humans and interventions mid-task, scored alongside the work itself.

Human judgment
Rewards calibrated against qualified reviewers, including when a decision belongs to a person. Reviewer disagreement is kept, not averaged away.
Agency and control
Staying within granted permissions and accepting a stop, pause or correction mid-task, while still being credited for finishing legitimate work.
Transparency
The agent’s own account of what it did is checked against what actually happened.
Monitorability
Monitors are kept out of the reward and measured after training, so agents are not taught to hide their intent from oversight.
Visibility
Full traces, with system state read directly rather than taken from the agent’s report.

How we work

Build, extend or review.

Environments are packaged for the formats teams already train with, so you don’t have to adopt a new training framework or environment format.

  • Harbor
  • OpenEnv
  • Inspect
  • verifiers
Build
A new environment for a capability no environment teaches well, starting with a paid design and calibration phase.
Extend
Interventions, permission changes, checks of the agent’s report and exploit testing, added to environments a team already runs.
Review
An existing environment or reward, checked for reward gaming, weak grading and leaked test data. Because we also build environments, a paid review is collaborative validation, never an independent assessment.

How we work with labs

Three commitments that keep our evidence worth trusting.

  1. 01

    We don’t grade our own homework.

    Our own measurements of a model trained on a Fide environment are labelled development or collaborative validation. Independent results come only from outside evaluators.

  2. 02

    We disclose who pays us.

    Relationships that bear on our findings are disclosed in every report and public claim, and to the recipients of private work. Today’s customer may be evaluated tomorrow, with the relationship stated.

  3. 03

    Rights agreed up front.

    Ownership, licensing and any exclusivity for work a customer funds are agreed before it begins. Grant-funded and public work is never exclusive, and held-out test sets are never sold for training.

Read our independence commitments →

How to read Fide

An environment builder that holds itself to evaluator standards.

What Fide is

  • A research lab that builds, extends and reviews RL environments with the discipline of an outside evaluator.
  • Part of the same research program as our benchmarks and evaluations of how AI earns, or fails to earn, trust.
  • A partner for teams that need a reward they can trust in work where judgment, authority and honesty matter.
  • A bridge between frontier labs and the institutions whose expert judgment their agents need.

What Fide is not

  • A data vendor selling task volume.
  • A certifier. A review that finds nothing shows only that it found nothing.

Working on an environment, or a capability no environment teaches well?

Tell us the capability, domain and training setup, and any environments you already run. We agree scope, rights, exclusivity and fee before work begins. Begin with non-sensitive context.

See an illustrative commissioning scope →
Talk to us about RL environments →