Step 2 of 4 · Stages 02–03
Evaluate & Remediate
See how models answer about you today, then resolve every claim your sources cannot back.
Senso reads what the models say about you, splits each answer into its individual factual claims, and checks every claim against your context layer. A claim with no supporting document, or one that contradicts a source, becomes a gap for you to decide.
- Choose questions to trackeach one is run against every model
- See how models answer todaythe baseline you compare against later
- Read the gapswhat your context could not back up
- Fix it at the sourcein place, never a second copy
Produces a list of unsupported claims, written back into the context layer.
How to run it
| Step | API call | Why it matters |
|---|---|---|
| Choose questions to track | POST/org | Registers a question to monitor across models. Track the ones that decide a purchase rather than the ones with the most volume — volume and consequence are different things, and only one of them costs you a customer. |
| See how models answer today | GET/org | Returns how each model currently answers your tracked questions. Read this before you write anything: content with no gap behind it is content nobody asked for. |
| Read the gaps | GET/org | Returns the gap queue. Searches fill it as they run: a query through the API, the MCP server or the CLI that finds nothing becomes an API search gap, weak on the first miss and open once the same question misses again. Checking every claim in a model answer against your sources is what will turn answers into gaps too. That part is not built yet. |
| Fix it at the source | PUT/org | Updates a source in place and cuts a new version; the old text stays readable. Adding a second document instead leaves two versions of the same fact, which is the conflict you were trying to remove. If nothing you own backs the claim at all, that is a new source and belongs in step 1. |
Ask your agent
Compare the answers with our sources. Show me any conflicting or missing details.
Stage 02 — Evaluate
Does the answer match your truth?
An answer can sound right and still miss the facts that matter. These are the questions your buyers ask agents, by funnel stage. Someone gets cited in those answers. Pick your industry and see whether it is you.
Sign in and improve narrative control across your industry.
Representative rows, not live data. Yours come from your own tracked questions.
Not wired up yet
There is no public industry endpoint, and every organization route needs a key, so the rows above are representative rather than live. Your real rows come from your own tracked questions once you have an organization.
Stage 03 — Remediate
- 1Let evaluation extract and check claims.Each claim from each answer is compared with your context. It runs in the background as things change, not while you wait.
- 2Read the gaps, highest consequence first.GET /org/gaps. Six kinds from claim evaluation, below, plus questions a search could not answer. Each names where it came from.
- 3Decide the material ones.Supply what is missing, confirm which source governs, or reject the correction. Record the decision with POST /org/gaps/{gapId}/resolutions.Requires a human
- 4Write the resolution back into your sources.The previous version stays readable. Nothing is overwritten silently.
Prioritization happens inside Remediate. There is no separate stage called Prioritize.
Status
Claims evaluation is not built yet. What exists today are utility endpoints, not a running system, and none of it is in the published API. The intent is event-driven — evaluation fires as content and answers change, rather than on a nightly batch. The gap queue itself does ship: searches through the API, the MCP server or the CLI that find nothing are filed as API search gaps, and the queue is read and resolved through /org/gaps.
Why it is not instant
Checking every claim against your sources takes time, so results arrive after the fact rather than while you wait. Ingestion, search and answers stay available throughout.
A finished ingest does not mean your claims have been checked. If you are reporting state, say ingestion is done and evaluation is queued.
What a gap can be
Contradicted
An approved source says otherwise.
Unsupported
Nothing in your context backs it.
Missing evidence
Plausible, but nothing was found.
Stale
Correct once, past its freshness window.
Ambiguous authority
Two sources could govern; unclear which.
Source conflict
Both approved, and they disagree.
One more kind is produced today without any evaluation: an unanswered question. A query that finds nothing is filed as an API search gap. Send X-Senso-Signals: off on test and synthetic queries to keep them out.
Gap
A problem your context could not back up, waiting on a decision. Gaps are surfaced, never fixed on their own.
Full definition →Human gates
- Two approved sources disagree, or it is unclear which governs
- The claim affects price, policy, eligibility, coverage, compliance, safety or terms
- The fix would change what a published page means
An agent can summarize the evidence and recommend. It cannot rewrite a source of record or sign off for someone else.
How to reach it
/org/promptsOpen in API reference →Create a question to track. Typed awareness, consideration, evaluation or decision./org/analytics/mentionsOpen in API reference →Mention rate: the share of answered runs that named you./org/analytics/citationsOpen in API reference →Citation Rate, Coverage and Share, per tier and per source./org/gapsOpen in API reference →The gap queue. Add statuses=weak to include questions that missed only once./org/gaps/{gapId}/resolutionsOpen in API reference →Record what was done about a gap: content added, a ruling, or a dismissal./org/kb/nodes/{id}/rawOpen in API reference →Correct a source. Cuts a new version; the old text stays readable./org/product-lines/{id}Open in API reference →Update a price, term or eligibility rule at its source of record.Every metric here has an exact denominator. Analytics has the math.
Rules for this step
- Surface conflicts. Do not pick a winner and edit the source to match.
- A claim that cannot be supported stays unresolved. It does not become true by default.
- Model answers are observations. They rank what to work on; they never overwrite a source.
- Ask the same questions each run, or the comparison means nothing.
Done when
- You know which questions you lose, and to whom
- Every gap names its type and where it came from
- A person decided each material conflict, and it is recorded with the claim
- Nothing was corrected silently