> ## Documentation Index
> Fetch the complete documentation index at: https://docs.conformly.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Knowledge Graph

> Curated cross-references between automotive standards

<style>
  {`
      p, li {
        text-align: justify;
        text-justify: inter-word;
      }
    `}
</style>

## What it is

The Standards Knowledge Graph is Conformly's curated layer encoding
the alignments between ASPICE 3.1, ISO 26262, and ISO 21434.

The same engineering practice — say, "verify software units with MC/DC
coverage for ASIL D code" — is required by both ASPICE SWE.4 and ISO
26262-6 §9. They use slightly different language and structure, but
the underlying work product is identical. Without a knowledge graph,
analyzing a document against ASPICE and against ISO 26262 means
running the analysis twice. With one, the same evidence satisfies
both clauses in a single analysis pass.

This is what lets Conformly answer the procurement question:

> "How does Conformly handle cross-standard analysis? We need to be
> assessed against ASPICE 3.1, ISO 26262, and ISO 21434 simultaneously
> — we don't want to upload our documents three times."

The answer is: you don't have to. The knowledge graph automatically
maps your evidence across all three standards.

## Where to find it

Sidebar → Library → **Knowledge Graph**.

You see four KPIs at the top (clauses, cross-references, standards
covered, most common edge type), a filter row for searching and
narrowing by standard, and a list of clauses below. Clicking any
clause expands it to show its full set of relationships.

## What you can do here

The page is **read-only**. You can:

* Browse the full clause set across all three standards
* Search for specific clause IDs, titles, or descriptions
* Filter by standard (ASPICE 3.1 / ISO 26262 / ISO 21434)
* Click into a clause to see its outgoing and incoming relationships
* Read the **justification** for each relationship — why this
  cross-reference exists, with a citation to the source standard or
  alignment guide
* See the **confidence score** of each cross-reference (0–1.0)

You cannot edit cross-references from the UI. The knowledge graph is
intentionally **version-controlled**, not user-editable. Every entry
goes through the same code review discipline as a code change.

## Why the KG is read-only

The knowledge graph is the product's **defensibility moat**. The whole
point of having one is that it's curated to a high standard — every
entry is sourced from published industry alignment guides
(VDA QMC, ASPICE 4.0 Annex, etc.), every relationship has an explicit
justification, every entry has a provenance tag.

Letting users add cross-references through a UI would mean:

* LLM-generated suggestions getting saved without review
* Customer-specific mappings polluting the global graph
* No way to roll back a bad addition
* No way to track who added what and why

Instead, the seed lives in a single YAML file
(`conformly-core/app/services/kg/seed_clauses.yaml`) that's loaded
into the database via a one-shot loader script. To add a new
cross-reference, an engineer (or domain expert) opens a pull request
against that file. The PR is reviewed, the cross-reference is run
against the eval benchmark to verify it improves the F1 score, and
only then is it merged and loaded into production.

## Provenance tags

Every cross-reference has one of three verification sources:

| Tag                    | Meaning                                                                                                                      |
| ---------------------- | ---------------------------------------------------------------------------------------------------------------------------- |
| **industry\_standard** | Explicitly aligned in published standards or alignment guides. Citation in the justification field. Used by the AI pipeline. |
| **domain\_expert**     | Reviewed by a human automotive compliance professional. Used by the AI pipeline.                                             |
| **ai\_proposed**       | LLM-suggested, awaiting human review. **Not used** by the AI pipeline — visible to admins for review only.                   |

The pipeline injects only `industry_standard` and `domain_expert`
entries into evaluation prompts. `ai_proposed` entries are stored
for review but never leak into customer-facing analysis. This is
enforced in code, not by convention.

## How the AI pipeline uses it

When you analyze a document against ASPICE SWE.4, the pipeline:

1. Looks up SWE.4 in the knowledge graph
2. Finds the cross-reference to ISO 26262-6 §9
3. Injects a "Cross-Standard Notes" section into the LLM evaluation prompt
4. Asks the LLM to (optionally) declare which related clauses the same
   evidence satisfies
5. Captures the LLM's claim in a `cross_standard_satisfied` field on the result

So when you analyze your unit test plan against ASPICE SWE.4 and the
test plan happens to satisfy ISO 26262-6 §9 too, the result tells you
both — without you having to upload the document twice or pick both
standards manually.

## Browsing tip

Click into a clause to see **everything connected to it**. Both
outgoing edges (this clause → other clauses it satisfies / cross-references)
and incoming edges (other clauses → this clause). This is the most
useful view when you're trying to understand "what does this clause
relate to?"

The justification text on each edge is short on purpose — it has
to fit a procurement review. If you want the full context, follow
the citation in the justification to the original standard or
alignment guide.

## What's in the v1 seed

| Standard   | Clauses                                          |
| ---------- | ------------------------------------------------ |
| ASPICE 3.1 | SYS.2, SYS.3, SWE.1, SWE.2, SWE.3, SWE.4, SUP.10 |
| ISO 26262  | Part 4 §6, Part 6 §6, §7, §9; Part 8 §8          |
| ISO 21434  | Clauses 9, 10, 11                                |

That's 15 clauses and 10 cross-references in v1. The seed grows
incrementally as we add benchmark cases — every new cross-reference
must improve the eval benchmark F1 score by ≥0.5 points before it
ships. Quality over quantity.