Collty Intelligence Core

It is Collty's proprietary team and project intelligence architecture. It connects the permission-aware work graph, semantic retrieval, domain algorithms, evaluation, model routing, provenance and controlled learning. External reasoning models are replaceable components inside this system, not the owner of its evidence or decision policy.

No. It is a domain intelligence system built around Collty's own evidence model and algorithms. Specialized models contribute interpretation and synthesis where they add value; retrieval, ranking, permissions, validation, evaluation and activation controls remain inside Collty.

Only authorized, task-relevant professional evidence is considered: competencies, roles, experience, cases, industries, languages, availability, capacity, delivery history, verified outcomes and collaboration signals. Team-level analysis also examines coverage, complementarity, workload, resilience and commercial context when the project requires it.

The assessment is designed around professional evidence and derived work signals, not sensitive personal attributes. Credentials, payment details, encryption material and unrelated workspace content are excluded. Context is minimized and filtered before any provider request.

Collty does not send customer workspace content to model providers for training shared foundation models. Provider processing is used to deliver the requested feature under the applicable service terms, while Collty keeps its work graph, evaluation data and learning controls within its own platform boundary.

The Core builds a task-specific evidence packet instead of exposing the work graph directly. Allowlisted structures, field filtering, redaction and evidence identifiers reduce the context to what the current task requires. Deeper protected traces are retained only when needed for evaluation, security or investigation.

Generated conclusions must resolve to permitted source objects and grounded evidence identifiers. Domain graders detect unsupported members, invented evidence and missing mandatory capabilities, preventing fluent model output from being accepted without source-backed support.

A production output can be captured as evidence, but it never becomes the expected answer automatically. Only a separately authored, expert-reviewed and versioned golden reference can be approved for comparison. Live case and run counts are operational governance metrics rather than static marketing claims.

Collty does not hard-code a changing production count into this public page. The exact approved-case count, domain coverage and evaluation-run history are live governance metrics available to authorized reviewers. Captured production outputs are never counted as approved cases until an expert authors and confirms the golden reference.

Domain graders examine request relevance, mandatory capability coverage, unsupported-member and invented-evidence failures, role allocation, availability, capacity, collaboration compatibility, resilience and human usefulness. The system evaluates both each professional and the team as a working whole.

Architect uses a separate evaluation contract for outcome phases, task completeness, dependencies, realistic duration, sprint cadence, ownership, confirmed specialists, evidence grounding and commercial consistency. A fluent-looking plan cannot pass solely because its text is well formed.

Yes. A candidate calibration is versioned against the active baseline and evaluated on a deterministic holdout set excluded from fitting. The active baseline remains unchanged until the candidate passes every promotion gate.

Promotion requires both minimum absolute and relative improvement plus a positive paired 95% confidence interval on weighted holdout evidence. Offline success does not authorize promotion by itself: the candidate must also survive controlled production observation before a person can activate it.

Displayed recommendations are stored as privacy-safe exposures. Feedback is linked to the exact option and rank; empirical observation propensity is smoothed and converted into bounded inverse-propensity weight so frequently exposed positions do not dominate calibration. This mitigates observed ranking bias, but it is not presented as complete causal identification of every unobserved outcome.

Yes. Under the current versioned contract, a 1-5 rating is mapped with (rating - 1) / 4, a 0-100 delivery score is divided by 100, and binary completion or recommendation outcomes remain 0 or 1. Raw values, normalized values and the contract version remain distinguishable for audit and controlled migration.

No. Time-sensitive signals use an explicit half-life contract. Recent evidence receives more weight while older evidence decays gradually instead of disappearing abruptly. The decay policy is versioned with the signal definition.

Collty combines direct delivery evidence with conservative shrinkage-based adjustment for project and team conditions. Regularization prevents unstable extreme scores and separates individual signals from shared project context. This improves attribution discipline, but Collty does not describe it as a causal estimate of a person's isolated effect.

No. Feedback first becomes a scoped, normalized learning signal. Signals are aggregated into a candidate calibration, tested on holdout evidence, observed safely and reviewed by a person. No single rating, project or customer can directly rewrite production weights.

Shadow evaluation compares candidate behavior without replacing the user-visible result. Model shadow mode can run a replaceable provider behind the same minimized context and evaluation contract, allowing Collty to compare quality, latency and cost without silently changing the primary workflow.

It must pass versioned private evaluation, deterministic holdout comparison and statistical promotion criteria. It can then enter a small sticky canary allocation, where live exposures and outcomes are measured. Final promotion requires explicit human approval.

Yes. Promotion preserves the previous active version. An atomic rollback restores that version without retraining, rewriting source evidence or waiting for a new deployment.

High-value runs retain provenance for the task, evidence boundary, context and grader versions, algorithm version, provider, result status and subsequent human decision. Operational records favor fingerprints and identifiers; protected payloads are encrypted when deeper review is necessary.

Provider calls are attributed to the responsible account and feature, including measured input, output, cached and reasoning usage when the provider supplies it. Retries and shadow runs are recorded as separate execution events, while deduplication prevents the same logical request from being charged or executed twice.

The user-facing result does not wait for calibration. Exposure and learning events are written through a durable asynchronous outbox, active policy is cached, and calibration runs in the background with cooldown and single-flight protection. Governance adds small bounded metadata writes rather than another AI pass to every request.

Collty monitors domain quality, critical grader failures, latency, provider errors, retries, token usage, cost, outbox health and drift across versions. These signals are evaluated together because a higher-scoring system is not an improvement if it becomes unreliable, materially slower or disproportionately expensive.

Activation is permission-controlled and human-approved. Candidate ranking weights cannot promote themselves, and consequential project actions remain subject to the product's role and approval boundaries. Database policies and service-only governance operations enforce the same boundary below the interface.

Collty assesses each AI-enabled workflow separately by intended purpose. Team Assembly, Signals applicant ranking, Team Intelligence and task allocation are conservatively governed as high-risk systems under the employment and work-management area of EU AI Act Annex III. Human approval remains an essential control, but it does not by itself remove that classification. Planning, canvas and conversational assistance have separate limited-risk assessments and may not silently inherit personnel-decision authority.

For a displayed match, Collty can summarize the task-relevant evidence that supported inclusion, such as capabilities, role coverage, relevant cases, industry experience, availability, capacity and verified collaboration or delivery signals. Collty presents alternative team options rather than making the final choice. The user can inspect each available profile, compare professional experience and competencies, remove or replace people and choose the composition. The explanation does not disclose another person's private information, internal weights, security controls or proprietary ranking formula.

Yes. Collty can identify missing or weak categories of professional evidence and suggest how to make a profile clearer, for example by adding specific capabilities, cases, role evidence, industry experience, availability or current capacity. This is profile-quality guidance, not a promise of inclusion or selection. Team recommendations remain specific to each project request, and the customer or project owner makes the final choice.

No. An ordinary search does not publish its hidden candidate pool to people who were not shown, and Collty does not notify every profile considered by semantic retrieval. In Signals, an applicant can correct their own profile evidence or report an AI-ranking error, while the client can manually approve, restore or remove candidates. This preserves meaningful system correction without exposing another user's search, candidate pool or private evidence.

No. Rating a result is a learning signal used in versioned evaluation and calibration. A system-error or profile-evidence correction request opens a separate authenticated case tied to the user's own account and the relevant output. The two records have different purposes and one never silently substitutes for the other.

No. Collty is not an employer, recruiter or tribunal and does not overturn a client's independent commercial choice. A professional can correct their own profile evidence or report that an AI ranking used inaccurate or unsupported evidence. Collty can investigate and correct the system or source record, while the client remains responsible for selecting the team.

AI-assisted interfaces identify their AI role at the point of interaction, and people-affecting recommendations remain under the customer's human control. A versioned inventory, Annex III assessments, risk and data-governance contracts, quality-management controls, technical-file evidence, post-market monitoring, incident records and role-specific AI literacy records support the operational governance expected by the EU AI Act. This describes Collty's implemented compliance architecture, not a claim of regulator approval or formal certification.

Yes. The architecture is designed in alignment with the NIST AI RMF and NIST Generative AI Profile principles for lifecycle governance, risk mapping, measurement, documentation, human oversight and continuous production monitoring. It also follows ISO/IEC 42001 and ISO/IEC 23894 principles for traceability, AI management, risk controls and continual improvement. This describes architectural and operational alignment with those requirements; it does not constitute formal ISO certification or NIST endorsement.