Skip to main content

Anchored Reasoning

Status: Reference Date: 2026-05-09 Tier: Premium · Patent-pending

What it is​

Anchored reasoning is the premium tier of Agience that anchors every AI answer to verified evidence in your encrypted memory and signals an explicit gap when the evidence doesn't warrant a conclusion.

In plain language: it's the layer that stops your AI from making things up.

The platform-tier of Agience gives your AI an encrypted, source-attributed substrate to work with — that's the data layer. Anchored reasoning (patent-pending) extends those guarantees into the AI's answer surface itself.

What it does​

When an MCP-capable agent (Claude Desktop, VS Code, Cursor, your own) queries Agience with Anchored reasoning enabled:

  1. The query routes to encrypted hybrid search across the principal's authorized light-cone — the same encryption guarantees as the platform tier apply.
  2. The reasoning engine selects evidence anchors from the returned artifacts using a constrained-attention mechanism that operates on evidence density rather than retrieval similarity alone.
  3. Generation is conditioned on retrieved, provenance-tagged anchors at every step. An impasse signal fires when the registry of available evidence does not warrant continuation.
  4. The returned answer carries source attribution per claim and an explicit gap when evidence is insufficient or contradictory.

What it prevents​

Hallucination. Generation is anchored to retrieved evidence at every step, not just at the input prompt. The model cannot wander into invented territory because the gate fires before it produces tokens the evidence does not support.

Confident wrong answers. When evidence is partial or contradictory, the system surfaces the gap explicitly rather than producing a smooth but wrong synthesis.

Untraceable claims. Every claim in an Anchored answer points back to the artifact that produced it. The provenance chain is preserved at the answer level, not just at the artifact level.

When to use it​

Anchored reasoning is the right tier when:

  • Your AI agents operate on data where being wrong has a cost (regulated content, customer commitments, compliance evidence, legal research, medical context).
  • You need traceable answers — every claim in an AI output must resolve to a source artifact that a human can review.
  • Your team has been burned by an AI hallucination before. The premium upgrade lands the moment the first wrong AI summary causes pain.
  • You're deploying AI on regulated data and need both encryption (the platform tier) and answer-level provenance (Anchored reasoning).

It's overkill when:

  • You're prototyping; the open-source platform tier is the right starting point.
  • The cost of a hallucinated answer is low (creative writing, brainstorming, ideation).
  • Your AI agents only need the encrypted memory layer, not answer-level reasoning over it.

How it works with the rest of the platform​

Anchored reasoning depends on the platform's three structural properties:

Encrypted hybrid search. The reasoning runs against your encrypted substrate; we never see your content or queries. (Same encryption guarantees as the platform tier.)

Built-in provenance. Anchored reasoning needs evidence-attributed artifacts to anchor against. Workspace artifacts that have crossed the commit boundary into a Collection are the canonical evidence base.

MCP-native interface. Anchored reasoning is exposed through the same MCP tool surface as the platform's regular search. Any MCP-capable agent can opt in.

What you'll see in your AI client​

With Anchored reasoning enabled, AI answers come back with two things the platform tier doesn't provide:

  1. Per-claim source attribution. Each statement in the answer points to the artifact (or specific evidence span within an artifact) that produced it. Click through; see exactly where the claim came from.
  2. Explicit-gap signaling. When the evidence doesn't warrant a conclusion, the answer says so. You'll see phrases like "the committed evidence does not include information about X" rather than a confident fabrication.

Pricing​

Anchored reasoning is included in the Anchored tier:

  • Anchored — $199 / seat / month
  • Includes everything in Team
  • Adds: Anchored reasoning, source-linked answer provenance, explicit-gap signaling, dedicated key management, priority SLA, single-tenant available

For pricing details, see Pricing. For Enterprise (custom integrations, dedicated infrastructure, on-prem), contact us.

Patent posture​

The Anchored reasoning mechanism is patent-pending. Specifically, the constrained-attention layer with evidence-density selection and explicit-gap signaling is the subject of an in-progress patent application. The encryption substrate the reasoning runs on is patent-pending separately.

Self-host availability​

Anchored reasoning is available for hosted, single-tenant, and self-host deployments under the Anchored tier or Enterprise license. The same code that runs the hosted product runs in your environment — no feature gating between deployment postures.

Frequently asked questions​

Is Anchored reasoning a separate product? No. It's a tier of Agience. You buy Agience; you upgrade to Anchored when you need the answer-level guarantees. One brand, one product, one upgrade path.

Will it work with Claude / VS Code / Cursor? Yes. Anything that speaks MCP. The reasoning runs server-side on Agience; your client just makes regular MCP tool calls.

Does Anchored reasoning train on my data? No. The encryption guarantees of the platform tier still apply — storage gets ciphertext; we never see your data. The reasoning mechanism operates against your authorized light-cone of encrypted artifacts.

What happens to claims with insufficient evidence? They surface as explicit gaps in the answer. You'll see a statement like "the committed evidence does not warrant a conclusion about X" instead of a fabricated answer.

How is this different from RAG? RAG retrieves from a (typically plaintext) index and injects retrieved chunks into a prompt. Anchored reasoning retrieves from an encrypted index and constrains generation to retrieved evidence at every step. Different mechanisms, different guarantees, different patent posture.

Can I get Anchored reasoning without the encryption layer? No. Anchored reasoning depends on the platform's encrypted substrate and provenance model. The two are designed together; the upgrade ladder is Free → Pro → Team → Anchored.

See also​