Assistant Gym Improvement
End-to-end system for taking an exact assistant on an exact job from baseline → failure diagnosis → intervention → fresh retest → evidence.
The methods, tools and capabilities behind Assistant Improvement, Interaction Support and Outcome Review. This is a reference to current and planned capabilities, not 37 separate products.
Explore the 37 capability families and their current release stages. Some support more than one product. The labels below distinguish bounded tools, development work and planned capabilities.
37 capability families
End-to-end system for taking an exact assistant on an exact job from baseline → failure diagnosis → intervention → fresh retest → evidence.
Give the assistant realistic experience and feedback before consequences are real.
Determine whether an exact assistant configuration meets explicit requirements for a job.
Replace "the assistant said so" with inspectable evidence about a claim, state or action.
Make a bounded capability decision against a defined standard.
Communicate a demonstrated bounded capability in a form another party can understand and inspect.
Reconstruct what happened, compare it to governing instructions and authority, and turn findings into repair.
Make an organization understandable and usable by assistants arriving from customers, partners or outside parties.
Let a human or assistant describe the problem in ordinary language and route them to the right part of the Gym.
Give an assistant durable, task-relevant access to current instructions, sources, resources, owners and supersession relationships.
Preserve operating guidance that proved useful so the lesson survives the conversation and reaches future work.
Use observed failures and strengths to select the next intervention or practice that matters for that assistant.
Match a demonstrated weakness with the most relevant human, assistant, expert or structured training resource.
Select the right participant to create a useful collaborative, adversarial, negotiating or diagnostic interaction.
Resolve bounded factual disagreement with evidence: Supported, Refuted or Unverifiable.
Determine whether participants followed the rules, mandates and permissions governing an interaction.
Create a shared stateful environment in which humans and assistants interact under defined roles, information, tools and rules.
Improve the ability of several participants to divide work, preserve context and finish one joint task.
Improve advocacy under conflicting incentives while preserving the owner's mandate and authority limits.
Train and test practical work where participants broadly want the same outcome.
Test behavior across sequences where later decisions depend on earlier interactions.
Define what passing means before assessment begins: outcomes, critical failures, evidence and validity.
Bind every result to the exact configuration that produced it.
Package a bounded result, its scope and supporting evidence so another party can inspect it.
Prevent old evidence from silently becoming a claim about a changed assistant or changed job.
Reassess capability after material changes to model, configuration, tools, standard or environment.
Bring the owner's rules, current evidence, permissions, budgets and operating limits to the point of action.
Answer the immediate question: is this assistant allowed to take this action now, with this source and permission state?
Give the owner a low-friction diagnostic of where deeper testing, training, memory, supervision or audit may help.
Maintain the work history, weaknesses, playbooks and evidence around the assistant the user already has as it takes on more jobs.
Build a complete work package for a domain: workflows, scenarios, tools, standards, evidence and playbooks.
Train and test insurance purchasing, coverage, claims, evidence and negotiation work.
Train and test item selection, constraints, price/number checking and authorized purchasing workflows.
Train and test travel selection, schedules, constraints, policies and changes.
Train and test precise product/part identification where near-matches are costly.
Train and test structured legal intake and disciplined escalation within the assistant's role.
Train and test support workflows where actual system resolution matters more than conversational fluency.