Investigations
Ask a question about one workflow, read cited findings, and request a change proposal that an independent evaluation then tests.
Watch the workflow: Inspect an authored investigation and quotation · Authored dummy data
The video shows the earlier standalone-run layout. The saved Investigation steps below use the current workspace interface.
Open an authored completed-investigation fixture in Tensile and expand one synthetic quotation. Read the mechanism, recommendation and stated evidence limits. No model ran or events were read; the example IDs do not point to stored events. The displayed supported label and two-of-two counts are authored layout data, not verified findings or customer outcomes. No investigation, change proposal or evaluation is requested.
Read the walkthrough transcript
- 00:00 — Open the authored investigation in Tensile. No model ran or events were read.
- 00:05 — Expand the synthetic quote. These IDs do not point to stored events.
An Investigation keeps a bounded question, its skill, accepted decisions and saved runs together. Each run preserves the inputs and source cutoff that produced its findings. A later edit does not rewrite an earlier result.
Retained work in Investigations
After your workspace completes a source migration, Investigations also lists retained studies, finding assessments and saved evidence. The Type column identifies each source. A retained study is source analysis. A finding assessment keeps its exact revision. Neither becomes a new Investigation run.
Use Finding assessments, Saved evidence or Incident intake above the collection to open the corresponding workflow. These links appear only for migrated, permitted sources. Check candidates preserves access to candidate authoring after study navigation moves.
Search covers the records loaded in the collection. Finding assessments has its own paged table and a Current finding workspace link. Open a source workflow to inspect its own pages and filters. Existing source permissions, retention limits and unavailable states still apply.
In a retained study, choose Browse all retained studies for its paged history. Older studies and Newer studies move through 50 metadata records at a time; opening a row keeps that exact study. The compact selector shows recent studies and the selected historical study.
A source migration changes navigation only after its records reconcile. Existing Home, Trace studies, Findings, Review evidence, Comparison and Triage URLs remain usable. A migrated family redirects to its Investigation destination while preserving supported selections and filters. Empty and unavailable study links also open within Investigations. Discovery candidates remains a separate workflow. Families can move separately. Excluded or held workspaces retain their compatibility views. A rollback restores their previous navigation without deleting source records or rewriting history.
Create an Investigation
- Open Investigations and select New investigation.
- Enter a title, question and business expectation.
- Select the workflow, downstream customer and time window. The window can cover at most 31 days.
- Select Create investigation. This saves the skill. It does not call a model.
- In Inputs, use Questions and scope to add separate questions within the same customer, workflow and time window.
The business expectation states what should happen. If you need an ordered expected path, enter a numbered list with consecutive steps. A single statement remains a single expected condition; the worker does not invent intermediate steps.
You need an owner or admin role to change an Investigation. The shared demo is read-only. Running also needs a configured worker, an Investigation engine profile, a named model connection, current source permission and a workspace spending limit. A saved key alone does not authorize customer data to leave Tensile.
In Inputs → Model source access, a customer approver can authorize an exact source, provider and model with an authorization reference. The permission covers that source's approved scope across Investigations in this workspace. Revoke it there when needed. A source that is no longer permitted cannot be used for a new call.
Edit the skill and run it
Open View skill. The breadcrumb changes to the skill. Its tabs contain Skill, Inputs, Runs, Decisions and Learnings.
Use the edit icon beside the skill to save a new version or cancel the edit. The refresh icon says Rerun on hover. Choose an engine profile and named model connection before running. An unpriced or unsupported model cannot run.
Open a row in Runs to inspect its saved inputs, state and usage. Cancel stops new calls; a request already in flight can still finish and incur a charge. A request with an unknown charge is held for reconciliation and is not automatically sent again.
Read findings, trajectories and evidence
The result tabs use the selected saved run and question:
- Summary shows the returned findings, recommendation and limits.
- Trajectory compares the saved expectation in green with captured behavior in red. Circle and square markers distinguish the paths without relying on color. Each observed path belongs to one execution. Red does not by itself prove a policy breach. Open a step to inspect its exact source; press Escape to close its details.
- Evidence contains the selected result's cited records. Open a record to inspect its quote and exact identities. The findings report uses this same saved result.
- Versions compares two saved results and their inputs. Earlier results remain unchanged after corrections or reruns.
A partial run can contain findings for one question while another question remains unavailable. That does not establish coverage for the unavailable question. Source revocation can withhold a saved result, report or export even when its run metadata remains visible.
Chat and accepted decisions
Open the chat icon to ask about the selected run and question. Choose a named connection and model. Answers must cite the permitted saved excerpts. They do not establish complete coverage or causation from event order.
A chat proposal does not change the skill or decisions automatically. Accept a specific choice to record its author, reason and question scope. Decisions keeps these choices in one table. An accepted expectation affects later runs; it does not alter a saved result.
Use Confirm on a finding to create or reopen its native Evaluation draft. The draft shows the originating question and expectation. These are authored context, not an executable evaluator definition or approval. Confirm does not activate a Check or deploy a change.
Schedules, merge and learning
The assignee control also contains the recurrence settings. Enable a schedule only after its model, source and spending permissions are configured. Each occurrence has a receipt. A blocked occurrence remains visible and cannot claim a completed run.
A merge preserves both histories. With skill merge off, the target keeps its skill. With it on, a model prepares a preview; acceptance saves the new target version and archives the source together. A failed or uncertain preview does not archive either Investigation.
Learnings separates observations from one run and candidate lessons for the workspace. Recording a correction does not automatically modify every Investigation skill. A periodic harness can propose lessons. Automatic promotion requires an approved, versioned policy and its validation evidence. New runs pin exact promoted lessons; rollback does not rewrite earlier runs.
An owner can also authorize an operator to admit selected excerpts from a retained conversation export. These are retrospective evidence, not a new customer execution. Each excerpt preserves its archive hash, conversation identity, original line, speaker and time. The collector can turn an observed correction into a candidate; repeated imports or excerpts from the same conversation count as one origin. Human scope and publication choices remain attributed instructions rather than additional incidents.
The general source-contract policy cannot promote retrospective candidates. A separate owner-approved causation pilot can admit one exact lesson to one named Investigation. It requires a frozen comparison set, actual matched model answers, source and chronology checks, and a verified activation record. Its limit is three new skill revisions within seven days of the first admitted revision. It cannot change the accepted question, business expectation, scope or human decisions.
The pilot starts in a conditional state. Failed or uncertain attempts stop adoption. Later answers and corrections stay out of earlier inputs. Revoked sources, changed owner access, expiry and rollback prevent future lesson pins. Earlier skill revisions and their history remain intact. A comparison through an explicitly selected local subscription is recorded as that transport; it does not prove that a production API connection works. An unchanged correct answer is non-regression evidence, not proof of improvement.
The decisions.md export contains accepted human choices. The learnings.md export contains run observations and workspace candidates with their states. These files have different authority and are not interchangeable.
Availability and limits
A saved Investigation remains usable when execution is not configured. Source-role references require their supported guarded reader; recorded traces cannot silently replace a repository, document or case source. An unavailable result stays unavailable until the required source or configuration is supplied.
The standalone engine workflow remains available separately. A change proposal and a held-out workflow evaluation each require their own configured adapter and source. A passed evaluation does not apply or deploy the proposed change.
Start from a customer incident through Incident intake when your workspace migrates Triage. Existing Triage links remain usable. Inspect single sessions in User Stories.