# AI Alignment Infrastructure Generated: 2026-11-11T09:34:32.667Z Source build: local Canonical docs: https://teammately.ai/docs --- id: intro.correctness-infrastructure title: AI Alignment Infrastructure summary: Connect expert judgment, explicit standards, coverage, and experiments into an enduring capability for improving enterprise AI. kind: concept product_area: introduction status: stable updated: 2026-11-06 canonical: /docs/introduction/correctness-infrastructure --- # AI Alignment Infrastructure Your enterprise’s expertise lives in the way people do their work. A specialist recognizes a missing constraint. A reviewer explains why two similar situations require different decisions. A team knows which tradeoffs it accepts and who can settle a disagreement. AI Alignment Infrastructure carries that knowledge into development and keeps it useful as systems and requirements change. ## Definition Teammately connects the work of experts, product owners, and engineers: - **Lemon Elicit** prepares contributions and follows the reasoning behind expert decisions. - **Rubrics and Policies** express requirements, applicability, and how judgments combine. - **Coverage Engineering** defines which situations the benchmark should represent. - **Lemon Weave** constructs Cases, supporting material, and targeted training examples. - **Trialground** executes saved Harnesses and assesses recorded behavior. - **Lemon Code and Lemon Goal** develop and investigate candidate implementations. These capabilities reinforce one another. A surprising trial can reveal an omitted condition. A specialist’s qualification can change the rubric and the Cases needed to exercise it. ## Decision checkpoint | Question | Begin with | Useful result | | --- | --- | --- | | Which distinction changes the expert’s decision? | Expert Contributions | A reason, condition, or exception | | How should the system be assessed? | Correctness Governance | Applicable rubrics and Regime treatment | | Which situations should be represented? | Coverage Design | Stories, tuples, and intended proportions | | What did the candidate actually do? | Trialground | Responses, captured trajectories, and judgments | | What should development investigate? | Lemon Code, Lemon Goal, or Weights training | A candidate intervention and a way to assess it | ## Make expertise reusable A single expert answer can inform more than one example. Preserving its reasoning lets a team apply the distinction across Cases, investigate exceptions, and return with a focused question when a later result challenges the standard. The Project Agent Brief and Reference Materials supply shared context. Contributions retain the specialist’s judgment. Policies and rubrics express accepted standards. Benchmarks make the situations and criteria explicit. Experiments show how a candidate behaves under those conditions. ## Keep the development decision connected A benchmark score is useful when the team can trace it to the conditions tested and the criteria applied. Inspect which requirements improved, which regressed, and whether the Case population supports the intended conclusion. Engineers can use that finding to change retrieval, instructions, tools, or orchestration, or to prepare material for weights training. Experts can clarify a requirement when the result exposes an unresolved judgment. {% example-demo title="A recommendation that fits the business" %} A merchant explains why a relevant product is a poor substitute when a required accessory is unavailable. Lemon Elicit develops the qualification into a proposed criterion. Coverage Design represents availability and substitution conditions. Trialground compares recommendation Harnesses, and the engineer can see whether a change respects the merchant’s distinction across those Cases. {% /example-demo %} ## Related workflows {% related-card-grid title="Put the connection to work" %} - [Product Quickstart](/docs/quickstart) - [Working with Lemon](/docs/concepts/lemon) - [The Teammately Correctness Lifecycle](/docs/introduction/correctness-lifecycle) {% /related-card-grid %} ## Related reference pages {% related-card-grid title="Find the right scope" %} - [Product Map](/docs/getting-oriented/product-map) - [Object Model](/docs/object-model) - [Product Responsibilities](/docs/introduction/product-boundaries) {% /related-card-grid %} --- id: intro.what-is-teammately title: What is Teammately? summary: Put your enterprise’s expertise to work shaping AI behavior through expert elicitation, shared standards, deliberate benchmarks, and measured improvement. kind: concept product_area: introduction status: stable updated: 2026-11-06 canonical: /docs/introduction/what-is-teammately --- # What is Teammately? Teammately is **AI Alignment Infrastructure** for enterprises. Your experts and engineers use it together to shape AI around your organization’s knowledge, judgment, and standards. An experienced practitioner can explain why a recommendation misses an operational constraint, when an exception changes the answer, or which tradeoff a team would accept. Teammately carries that reasoning into the standards, benchmarks, Harness changes, and training material your team develops. [Lemon](/docs/concepts/lemon), your AI agent for alignment, prepares inquiries, develops coverage, constructs material, and investigates improvements. Experts review the interpretation of their judgment. Engineers inspect proposed changes and measured results before deciding what to adopt. ## Definition **Expert Contributions** uncover the reasons, preferences, and exceptions behind a judgment. **Rubrics and Policies** give that judgment an explicit form, with applicability and a project Regime that defines how results combine. **Coverage Engineering** determines which situations the benchmark should represent and in what proportions. **Lemon Weave** constructs Cases, controlled alternatives, supporting materials, and training examples around those needs. **Trialground** runs saved Harnesses and evaluates recorded behavior against applicable rubrics. **Lemon Code** helps engineers develop a Harness through a working conversation. **Lemon Goal** pursues benchmark objectives through candidate experiments. **Weights training** connects targeted datasets to exports, connected training services, and subsequent assessment. ## Decision checkpoint | Your objective | Start with | What your team develops | | --- | --- | --- | | Understand a consequential judgment | Expert Contributions | Reasoning, conditions, and proposed standards | | Express a shared requirement | Correctness Governance | Policies, rubrics, and scoring treatment | | Represent important situations | Coverage Facets and Coverage Design | Stories, tuples, and a deliberate Case mix | | Test a candidate | Harnesses and Benchmark Evaluations | Recorded behavior and rubric results | | Develop an implementation | Lemon Code or Lemon Goal | Reviewed changes and evaluated candidates | | Teach a missing distinction | Weights training | Targeted datasets and trained-candidate assessment | | Learn from application behavior | Fieldloop | Investigated Threads and follow-up work | ## More impact per human judgment Specialist AI must handle the conditions that matter in your business: source authority, operational constraints, local preferences, costly mistakes, and consequential exceptions. Your team needs to express those distinctions and see whether a change improves the behavior they care about. Teammately keeps the connection visible. An expert’s qualification can refine a rubric. Its applicability can identify a coverage need. A constructed Case can expose a Harness weakness. A trial can test a proposed intervention, and its result can motivate a more focused expert question. ## Begin with the work at hand Bring an expert question, a coverage gap, an existing Harness, production interactions, or a training objective. Project context and reusable Assets support the work across Benchmarks. Each Benchmark connects its Case population, standards, evaluations, and improvement work. Experts establish and review judgment. Project operators approve shared standards and benchmark changes. Engineers decide which implementation and training results to integrate into their systems. {% example-demo title="Grounded enterprise search" %} A specialist explains when a newer agreement overrides an older handbook. Lemon Elicit develops that qualification into a proposed rubric. Coverage Design identifies the conflicting-document situations the benchmark should represent. Lemon Weave constructs Cases, Trialground assesses a retrieval Harness, and Lemon Goal tests a change against the resulting failures. {% /example-demo %} ## Related workflows {% related-card-grid title="Choose your first assignment" %} - [Product Quickstart](/docs/quickstart) - [Working with Lemon](/docs/concepts/lemon) - [Product Map](/docs/getting-oriented/product-map) - [Weights Training](/docs/weights-training) {% /related-card-grid %} ## Related reference pages {% related-card-grid title="Understand the platform" %} - [AI Alignment Infrastructure](/docs/introduction/correctness-infrastructure) - [Key Objects and Relationships](/docs/getting-oriented/key-objects-and-relationships) - [Product Responsibilities](/docs/introduction/product-boundaries) {% /related-card-grid %} --- id: intro.correctness-lifecycle title: The Teammately correctness lifecycle summary: Follow expert judgment into standards, coverage, evaluation, and the next development decision. kind: concept product_area: introduction status: stable updated: 2026-11-06 canonical: /docs/introduction/correctness-lifecycle --- # The Teammately Correctness Lifecycle A specialist’s qualification can change how your team develops AI. The lifecycle keeps that reasoning connected as it becomes a criterion, a Case, an experiment, and a finding the team can act on. ## Definition Project foundations provide shared context, standards, and reusable Assets. Each Benchmark gives that work a concrete purpose: which situations to represent, which behavior to assess, and which improvement to investigate. The work can begin at any useful point. An imported interaction may expose a question. An expert may explain an exception. An engineer may bring an existing Harness and a behavior to improve. ## Decision checkpoint | What you learn | Next destination | What stays connected | | --- | --- | --- | | A condition changes the expert’s judgment | Correctness Governance | The contribution, criterion, and applicability | | A situation is missing from the benchmark | Coverage Design | The coverage need, tuple, and Case | | A candidate fails an applicable requirement | Harness improvement | The starting candidate, result, and objective | | A result exposes ambiguous expectations | Expert Contributions | The question and the Cases that raised it | | A distinction needs teaching material | Weights training | Criteria, coverage, examples, and separate assessment | ## Build the shared foundation Write the Project Agent Brief and connect Reference Materials. Define the input and materials that Cases should contain. Use Dimensions, Topics, and Scenarios to describe relevant variation. Lemon can help develop these foundations. Review the proposals in their project context and refine the language around distinctions that matter to your team. ## Turn judgment into something the benchmark can test Request a focused Contribution. Lemon Elicit prepares and adapts the work; experts explain their reasoning and review proposed interpretations. Policies organize shared expectations. Rubrics express observable criteria. Applicability identifies the Cases where a criterion matters, and the Regime defines how judgments contribute to the result. ## Represent and assess the behavior Coverage Design connects the intended situation mix with existing and constructed Cases. Review the Cases and applicable standards, then preserve the selected population in a Snapshot and Benchmark Version. Trialground runs the saved candidate against that benchmark. Inspect responses, captured trajectories, rubric judgments, and completeness before interpreting the aggregate. ## Put the findings to work Use Lemon Code for interactive Harness development, Lemon Goal for benchmark-driven investigation, or External Agents for work in your application’s codebase. Use Weights training to prepare targeted teaching material and assess the resulting candidate separately. {% example-demo title="An exception becomes a development target" %} A specialist explains that an expired agreement still matters when the current contract incorporates it. The team refines applicability and constructs Cases around that condition. A trial shows that retrieval filtering discards the older agreement too early. The engineer can now test a precise intervention and return with a focused question if an exception remains ambiguous. {% /example-demo %} ## Related workflows {% related-card-grid title="Continue the lifecycle" %} - [Product Quickstart](/docs/quickstart) - [Expert Contributions](/docs/expert-contributions) - [Coverage Design](/docs/coverage-design) - [Harness Improvement](/docs/improve) {% /related-card-grid %} ## Related reference pages {% related-card-grid title="Understand the connections" %} - [Object Model](/docs/object-model) - [Rubrics and Policies](/docs/correctness-governance/policies-and-rubrics) - [Weights Training](/docs/weights-training) {% /related-card-grid %} --- id: product-loop title: The Teammately correctness loop summary: See how expert inquiry, standards, coverage, experiments, and improvement inform one another. kind: concept product_area: introduction status: stable updated: 2026-11-06 canonical: /docs/product-loop --- # The Teammately Correctness Loop The next useful development question often comes from the work you just completed. A comparison reveals an exception. A trial exposes a missing condition. An expert’s explanation suggests a better Case. Teammately keeps those discoveries connected so your team can deepen its understanding and improve the system against it. ## Definition Lemon Elicit develops expert reasoning into proposed standards. Rubrics and Policies express what matters and where it applies. Coverage Engineering defines the situations the benchmark should examine. Lemon Weave constructs material around that design. Trialground assesses candidate behavior, and Harness improvement or Weights training puts the findings to work. ## Decision checkpoint | Finding | Follow it into | Useful next question | | --- | --- | --- | | Experts prefer different outputs | A focused Contribution | Which condition or responsibility explains the difference? | | A criterion applies too broadly | Policy or rubric applicability | Which Cases should it govern? | | A consequential combination is missing | Coverage Design | What would a Case need to demonstrate? | | The candidate fails a clear requirement | Harness development | Which intervention could change the behavior? | | A recurring distinction needs teaching | Weights training | Which examples teach it, and how will it be assessed separately? | ## Before and after | Before | After | | --- | --- | | A surprising result | A concrete question about behavior or the standard | | An expert qualification | Reviewed wording and applicability | | A coverage need | A Case with a clear construction purpose | | A proposed intervention | A saved candidate and comparable trial results | ## Read the connection in both directions From a result, inspect the Case, applicable criterion, and candidate behavior. From a standard, inspect its rationale, expert contribution, applicability, and representative Cases. These paths help distinguish an implementation problem from an unclear expectation or missing situation. Each finding can then reach the people and workflow best placed to resolve it. {% example-demo title="A boundary the aggregate concealed" %} A Harness improves overall, but Cases with conflicting documents regress. The team inspects the rubric results and asks a specialist which source should control. The answer sharpens applicability, motivates a new coverage tuple, and gives the engineer a more precise source-selection objective. {% /example-demo %} ## Related workflows {% related-card-grid title="Choose the next piece of work" %} - [Product Quickstart](/docs/quickstart) - [Lemon Elicit](/docs/concepts/correctness-elicitation) - [Coverage Design](/docs/coverage-design) - [Lemon Goal](/docs/concepts/lemon-goal) - [Weights Training](/docs/weights-training) {% /related-card-grid %} ## Related reference pages {% related-card-grid title="Inspect the basis of a result" %} - [Rubrics and Policies](/docs/correctness-governance/policies-and-rubrics) - [Benchmark Evaluations](/docs/benchmark-evaluations) - [Object Model](/docs/object-model) {% /related-card-grid %}