Open Knowledge Format (OKF)

July 24, 2026 · View on GitHub

Version 0.2

OKF is an open, human- and agent-friendly format for representing knowledge: the metadata, context, and curated insight that surrounds data and systems. It is designed to be authored by people, generated by agents, exchanged across organizations, and consumed by both.

The format is intentionally minimal: a directory of markdown files with YAML frontmatter. There is no schema registry, no central authority, and no required tooling. If you can cat a file, you can read OKF; if you can git clone a repo, you can ship it.

This document is self-contained: it specifies everything needed to produce and consume OKF v0.2. A summary of what changed from v0.1 is in §13.

1. Motivation

The space of knowledge representation for AI agents is evolving quickly, and many incompatible conventions are emerging. OKF takes the position that knowledge is best represented in commonly accessible, established formats that are:

Readable by humans without tooling.
Parseable by agents without bespoke SDKs.
Diffable in version control.
Portable across tools, organizations, and time.

Increasingly, a knowledge corpus is not authored once and then read: it is continuously written and maintained by agents. When most concepts are machine-generated, a consumer needs answers that a plain markdown-plus-frontmatter convention does not make first-class:

What was this created from, and how was it verified? (provenance)
How much should I trust it? (trust)
Is it still true? (freshness)
Is it the current version? (lifecycle)
Was this number produced the way we said it must be? (attestation)

OKF v0.2 makes provenance, trust, lifecycle, and attestation first-class while keeping the format minimally opinionated. The format is minimally opinionated. It standardizes only the small set of structural conventions needed to make a knowledge corpus self-describing — anything beyond that is left to the producer.

Goals

Define a universal format that producers (people, agents, export pipelines) can write into.
Inform how consumers (agents, UIs, search indexes, deterministic code) should read and traverse it.
Facilitate exchange of knowledge across systems and organizations.
Standardize the small set of frontmatter fields that make an agent-maintained corpus trustable, without prescribing any runtime.

Non-goals

Defining a fixed taxonomy of concept types.
Prescribing storage, serving, or query infrastructure.
Replacing domain-specific schemas (Avro, Protobuf, OpenAPI, and so on). OKF references them; it does not subsume them.
Specifying a packaging or invocation standard for the code an executor or attester points at. OKF fixes the interface, not the packaging.

2. Terminology

Knowledge Bundle (or bundle): A self-contained, hierarchical collection of knowledge documents. The unit of distribution.
Concept: A single unit of knowledge within a bundle, represented as one markdown document. It may describe a tangible asset (a table, an API), an abstract idea (a metric, a business process), or anything in between.
Concept ID: The path of the concept's file within the bundle, with the .md suffix removed.
Frontmatter: A YAML metadata block delimited by --- at the top of a markdown file.
Body: Everything in the file after the frontmatter.
Link: A standard markdown link from one concept to another, used to express relationships beyond the implicit parent/child hierarchy.
Source: A material a concept derives from, external or internal to the bundle, recorded in the sources frontmatter field.
Provenance: The set of sources a concept derives from.
Credibility signal: An objective, per-source fact (author, usage_count, last_modified) used to infer trust; OKF records the signals, not a verdict (see §5.1).
Actor: A string identifying who or what performed an action, using the convention <producer>/<version> for agents, human:<id> for people, and process:<id> for automated processes (see §7).
Trust tier: A level derived from a concept's verified field: unverified, machine-confirmed, or human-reviewed (see §5.3).
Attested Computation: A concept (type: Attested Computation) carrying a sanctioned way to compute a value, so a consumer can confirm the value was produced by running it (see §10).
Executor: Run instructions or code that executes a computation and returns a receipt (see §10.2).
Receipt: The evidence a run returns, shaped by executor.receipt; a runtime artifact, not stored in the bundle (see §10).
Attester: Deterministic (no-LLM) code that inspects a receipt and returns a verdict (see §10.2).

3. Bundle structure

A bundle is a directory tree of markdown files. The directory structure is independent of the domain: producers organize concepts however makes sense for the knowledge being captured.

path/to/bundle/
  index.md                      # Optional. Directory listing for progressive disclosure.
  log.md                        # Optional. Chronological history of updates.
  <concept>.md                  # A concept at the bundle root.
  <subdirectory>/               # Subdirectories organize concepts into groups.
    index.md
    <concept>.md
    <subdirectory>/
      ...

A bundle MAY be distributed as:

A git repository (recommended, since it provides history, attribution, and diffs).
A tarball or zip archive of the directory.
A subdirectory within a larger repository.

3.1 Reserved filenames

The following filenames have defined meaning at any level of the hierarchy and MUST NOT be used for concept documents:

Filename	Purpose
`index.md`	Directory listing. See §8.
`log.md`	Update history. See §9.

All other .md files are concept documents.

Tags remain a first-class concept through the tags frontmatter field (§4.1). OKF does not specify a separate file format for aggregating documents by tag; a consumer that wants a tag-browsing view can synthesize one at consumption time by scanning frontmatter.

4. Concept documents

Every concept is a UTF-8 markdown file with two parts:

A YAML frontmatter block, delimited by --- on its own line at the start of the file and a closing --- on its own line.
A markdown body, containing free-form content.

4.1 Frontmatter

---
type: <Type name>                  # REQUIRED
title: <Optional display name>
description: <Optional one-line summary>
resource: <Optional canonical URI for the underlying asset>
tags: [<tag>, <tag>, ...]          # Optional
# ... trust, lifecycle, provenance, and computation families (see §5, §10)
# ... other producer-defined key/value pairs
---

Required:

type: A short string identifying the kind of concept. Consumers use it for routing, filtering, and presentation. Example values: BigQuery Table, BigQuery Dataset, API Endpoint, Metric, Playbook, Reference, Attested Computation.

Type values are not registered centrally. Producers SHOULD pick values that are descriptive and self-explanatory; consumers MUST tolerate unknown types gracefully, typically by treating them as generic concepts.

type is the only always-required key; a concept carrying just type is fully conformant (§11).

Recommended:

title: Human-readable display name. If omitted, consumers MAY derive a title from the filename.
description: A single sentence summarizing the concept. Used by index.md generators, search snippets, and previews.
resource: A URI that uniquely identifies the underlying asset the concept describes. Absent for concepts that describe abstract ideas rather than physical resources.
tags: A YAML list of short strings for cross-cutting categorization.

The optional provenance, trust, and lifecycle families (§5) and the computation fields for Attested Computation concepts (§10) may also appear.

Extensions: Producers MAY include any additional keys. Consumers SHOULD preserve unknown keys when round-tripping and MUST NOT reject documents with unrecognized fields.

4.2 Body

The body is standard markdown. Producers SHOULD favor structural markdown (headings, lists, tables, fenced code blocks) over freeform prose, since structure aids both human reading and agent retrieval.

There are no required body sections. The following headings have conventional meaning and SHOULD be used when applicable:

Heading	Purpose
`# Schema`	Structured description of an asset's columns/fields.
`# Examples`	Concrete usage examples, often as fenced code blocks.
`# Computation`	The sanctioned computation of an Attested Computation. See §10.

Per-claim attribution to external sources uses markdown footnotes keyed to sources entries rather than a body citations list (§5.1).

4.3 Example: a concept bound to a resource

---
type: BigQuery Table
title: Customer Orders
description: One row per completed customer order across all channels.
resource: https://console.cloud.google.com/bigquery?p=acme&d=sales&t=orders
tags: [sales, orders, revenue]
generated: { by: reference_agent/gemini-2.5-pro, at: 2026-05-28T14:30:00Z }
---

# Schema

| Column        | Type      | Description                              |
|---------------|-----------|------------------------------------------|
| `order_id`    | STRING    | Globally unique order identifier.        |
| `customer_id` | STRING    | Foreign key into [customers](/tables/customers.md). |
| `total_usd`   | NUMERIC   | Order total in US dollars.               |
| `placed_at`   | TIMESTAMP | When the customer submitted the order.   |

# Joins

Joined with [customers](/tables/customers.md) on `customer_id`.

4.4 Example: a concept not bound to a resource

---
type: Playbook
title: "Incident response: data freshness alert"
description: Steps to triage a freshness alert on the orders pipeline.
tags: [oncall, incident]
generated: { by: human:ahormati, at: 2026-04-12T09:00:00Z }
---

# Trigger

A freshness alert fires when `orders` lags more than 30 minutes behind its
expected SLA. See the [orders table](/tables/orders.md).

# Steps

1. Check the [ingestion job dashboard](https://example.com/dash).
2. ...

5. Provenance, trust, and lifecycle

These frontmatter families make "where did this come from," "how much should I trust it," and "is it still current" answerable from frontmatter. All are optional. Their absence carries meaning: an unverified concept is distinguishable from a verified one, but is never rejected (§11).

5.1 Provenance: `sources`

sources records the materials a concept derives from, external or internal to the bundle.

sources:
  - id: ga4-schema
    resource: https://developers.google.com/analytics/bigquery/export-schema
    title: GA4 BigQuery Export schema
    author: team:ga4-docs
    usage_count: 5000
    last_modified: 2026-05-30
usage_window: { from: 2026-06-01, to: 2026-06-30 }

Each sources entry:

resource: REQUIRED within an entry. Names either a concrete artifact a consumer can follow (an absolute URL, a bundle-relative path, or a path into a references/ subdirectory, §6) or a population or scope descriptor it cannot (for example all queries in BigQuery project X).
id: Optional. A stable key used to attribute individual claims (see below). SHOULD be present when the body cites the source.
title: Optional. Human-readable label for the source.
The optional credibility signals author, usage_count, and last_modified, described next.

Source credibility signals. OKF records objective, per-source signals so a consumer can judge how much to trust a concept by judging the sources it was extracted from. It does not store a credibility score: a score is subjective, unportable across consumers, and goes stale. Credibility is inferred from the signals, the same way trust tiers are (§5.3), not stored. Each signal is optional and lives on a sources entry:

author: Who or what produced the source, in the actor convention (§7). An authority signal.
usage_count: How often resource was exercised (dashboard views, query executions, page reads) over usage_window. An adoption and liveness signal. For a single artifact it is that artifact's own exercise count; for a scope descriptor it is the number of exercises within the scope that touch the concept.
last_modified: When the source itself last changed (YYYY-MM-DD). A recency signal, distinct from generated.at (§5.2), which records when the concept was written.
usage_window: Written once as a sibling of sources, it frames every usage_count with a { from, to } date range. A single entry MAY carry its own usage_window to override the shared one.

usage_count is a coarse signal. It is comparable at the alive-versus-dead and order-of-magnitude level, and against a source's own history over time, but not as a precise cross-kind ranking: a scheduled query's executions and a human's deliberate dashboard views do not carry equal weight. Consumers SHOULD read it as liveness and trend, not as a score.

Lineage is expressed through links, not a dedicated field. When a resource points at another OKF concept, the derivation edge already exists in the bundle graph (§6), so a consumer MAY recurse into that source's own sources and let credibility propagate. External leaf sources carry only their intrinsic signals. Deeper lineage (an explicit external derived_from, or data lineage) is out of scope for v0.2.

Per-claim attribution. To attribute a specific claim, use a markdown footnote whose label is a sources[].id:

The `events_` table is sharded daily as `events_YYYYMMDD`.[^ga4-schema]

[^ga4-schema]: GA4 BigQuery Export schema

The footnote label is the join key into sources; consumers resolve attribution through the matching entry, not by parsing the footnote prose. Labels are keyed rather than positional (sources[0]) because agents constantly rewrite these documents: a positional index misattributes silently the moment the list is reordered, whereas a stable id survives reordering.

5.2 Trust: `generated` and `verified`

generated records how the current content was produced. verified records who or what has confirmed the content against its sources or resource. They are kept distinct because who wrote a concept need not be who confirmed it.

generated: { by: reference_agent/gemini-2.5-pro, at: 2026-06-20T22:53:05Z }

generated.by: REQUIRED within generated. An actor (§7).
generated.at: An ISO 8601 datetime marking the content's last meaningful change. Consumers use it to tell a recent edit from a stale fact.

verified:
  - { by: human:ahormati, at: 2026-06-25T09:00:00Z }
  - { by: process:finance-nightly, at: 2026-06-26T02:00:00Z }

verified: A list of verification events, each with by (an actor) and at (an ISO 8601 datetime). Multiple entries capture independent checks, for example a human sign-off plus a nightly process. "How recently" is the latest at.
verified is independent of generated.at: content can change without re-confirmation, and facts can be re-confirmed without regeneration.
A single verifier MAY be written as one { by, at } mapping without the list dash. Consumers MUST treat a bare mapping as a one-element list:

verified: { by: human:ahormati, at: 2026-06-25T09:00:00Z }

5.3 Trust tiers

Consumers derive a trust tier from verified, lowest to highest:

No verified key ⇒ unverified.
verified by non-human: actors only ⇒ machine-confirmed.
verified by a human:<id> actor ⇒ human-reviewed.

A concept with no trust frontmatter is still consumable; consumers MUST NOT reject it (§11). Trust tiers are advisory signals, not access control.

5.4 Lifecycle: `status`

status: stable        # draft | stable | deprecated

draft: not yet reviewed; possibly incomplete.
stable: default; ready for consumption.
deprecated: kept for links and history; no longer current.

Absent status ⇒ stable.

5.5 Lifecycle: `stale_after`

stale_after: 2026-09-23   # absolute date; content is stale on/after this day

Optional. An absolute date (YYYY-MM-DD). A concept is stale when today >= stale_after. An absolute date, not a relative TTL, keeps the staleness decision a plain date comparison with no reference to when the concept was read.

6. Cross-linking and paths

6.1 Links between concepts

Concepts MAY link to other concepts using standard markdown links. Two forms are supported:

Absolute (bundle-relative): begins with /, interpreted relative to the bundle root. This is the recommended form because it is stable when documents are moved within their subdirectory.
```
See the [customers table](/tables/customers.md) for the join key.
```

Relative: a standard markdown relative path.

See the [neighboring concept](./other.md).

A link from concept A to concept B asserts a relationship. The specific kind (parent/child, references, joins-with, depends-on) is conveyed by the surrounding prose, not by the link itself. Consumers that build a graph view typically treat all links as directed edges of an untyped relationship.

Consumers MUST tolerate broken links: a link whose target does not exist in the bundle is not malformed; it may simply represent not-yet-written knowledge.

6.2 Path-valued fields

Several fields name a path or URI: resource, sources[].resource, computation, executor.resource, and attester.resource (§10). A sources[].resource may instead be a scope descriptor (§5.1), in which case it is not a path. Each path-valued field accepts:

an absolute URL (for example https://...),
a bundle-relative path beginning with /, or
a relative path (for example ../computations/revenue.md).

6.3 The `references/` convention

A references/ subdirectory conventionally mirrors external material, run instructions, or code as first-class concepts within the bundle. Sources, executors, and attesters commonly point into it (for example references/attesters/revenue.py). It is a naming convention, not a requirement.

7. Actor convention

Fields that record an identity (generated.by, verified[].by) use a single actor convention:

<producer>/<version> for agents and tools, for example reference_agent/gemini-2.5-pro.
human:<id> for a person, for example human:ahormati.
process:<id> for an automated process, for example process:finance-nightly.

Consumers that classify trust (§5.3) key off the human: prefix, so producers MUST use it for hand-authored or human-confirmed content.

8. Index files

An index.md file MAY appear in any directory, including the bundle root. It enumerates the directory's contents to support progressive disclosure: letting a human or agent see what is available before opening individual documents.

Index files contain no frontmatter, with one exception: a bundle-root index.md MAY carry an okf_version key (§12). The body uses one or more sections, each grouping concepts under a heading:

# Section / Group Heading

* [Title 1](relative-url-1) - short description of item 1
* [Title 2](relative-url-2) - short description of item 2

# Another Section

* [Subdirectory](subdir/) - short description of the subdirectory

Entries SHOULD include the description from the linked concept's frontmatter. Producers MAY generate index.md automatically; consumers MAY synthesize one on the fly when none is present.

9. Log files

A log.md file MAY appear at any level of the hierarchy to record the history of changes to that scope. The format is a flat list of date-grouped entries, newest first:

# Directory Update Log

## 2026-05-22
* **Update**: Added a BigQuery table reference for [Customer Metrics](/tables/customer-metrics.md).
* **Creation**: Established the [Dataplex Playbook](/playbooks/dataplex.md).

## 2026-05-15
* **Initialization**: Created foundational directory structure.

Date headings MUST use ISO 8601 YYYY-MM-DD form. Log entries are prose; the leading bold word (**Update**, **Creation**, **Deprecation**) is a convention, not a requirement.

10. Attested computations concept

An Attested Computation concept carries not just what a value means but a sanctioned way to compute it, so a consumer can confirm the agent ran the blessed computation instead of improvising its own. Provenance (§5.1) answers "where did this claim come from"; attestation answers "was this number produced the way we said it must be." OKF records the computation and the means to check it; it does not execute anything itself.

10.1 A computation is its own concept

A sanctioned computation is a standalone concept of type: Attested Computation. A concept that needs the value (a Metric, a BigQuery Table) links to it with a normal markdown link (§6). Three properties motivate the standalone concept:

runtime defines what parameters mean. A parameter is a SQL bind variable, a dbt var, or a Python argument depending on the runtime. Keeping runtime and parameters in one frontmatter makes the binding semantics self-evident.
One computation, many consumers. The same computation can back a metric, a dashboard concept, and a report; as a concept it is referenced once and reused.
Trust state is per computation. verified, stale_after, and a single attester describe one thing. Revenue, profit, and margin each verify and attest independently, which is three concepts, not three entries in one frontmatter.

10.2 Contract fields

The contract is the concept's top-level frontmatter. In addition to the provenance, trust, and lifecycle families (§5), an Attested Computation concept carries:

runtime: REQUIRED for this type. The single field that says how to run the computation, and so how the executor and attester interpret it and what parameters mean. Example values: bigquery, postgres, dbt, python, Looker.
parameters: A list of the typed, named holes the agent may fill. Each entry: { name, type, required }. Binding semantics follow runtime.
computation: Optional. A path (§6.2) to a file holding the computation, used instead of an inline body fence (see §10.3). Absent ⇒ the body # Computation fence is the computation.
executor: How the computation is run. resource names run instructions or code; a runner (an agent, or deterministic consumer code) follows it. receipt declares the fields a run must return, the evidence the attester inspects (for example a BigQuery job_id and the SQL the job actually executed).
attester: The deterministic check. resource names code (no LLM) that takes a receipt and returns a verdict. It is meant to run consumer-side.

What sits behind a resource (a Skill, a script, a container) is a packaging choice; OKF fixes the interface, not the packaging (§1).

---
type: Attested Computation
title: Revenue for fiscal year
description: Recognized revenue for a fiscal year, per Finance's definition.
status: stable
runtime: bigquery
parameters:
  - { name: year, type: integer, required: true }
executor:
  resource: references/skills/run-on-bq.md
  receipt: [job_id, executed_sql, result]
attester:
  resource: references/attesters/revenue.py
generated: { by: reference_agent/gemini-2.5-pro, at: 2026-06-20T22:53:05Z }
verified: { by: human:ahormati, at: 2026-06-25T09:00:00Z }
stale_after: 2026-09-23
sources:
  - id: rev-policy
    resource: https://wiki.acme/finance/revenue-recognition
    title: Revenue recognition policy
---

# Computation

    SELECT SUM(amount) AS revenue
    FROM finance.recognized_revenue
    WHERE fiscal_year = @year

The computation binds only the declared `parameters`, per the recognition
policy.[^rev-policy]

[^rev-policy]: Revenue recognition policy

10.3 The computation

Provide the computation in one of two ways:

Inline: a single fenced code block in the body under # Computation. Best for a short computation reviewed alongside the contract.
File: set computation to a path (§6.2) and omit the body fence. Best for a long or generated computation, or one already kept as a real file shared with non-OKF tooling.

runtime: bigquery
computation: references/computations/lib/revenue.sql
parameters:
  - { name: year, type: integer, required: true }

The agent MAY only supply values for the declared parameters; it MUST NOT author or edit the computation. Binding computation with the parameter values into the executable artifact is the consumer's job, and the attester independently re-derives that same binding to compare against what actually ran. Because the comparison is on the expanded, compiled artifact the receipt carries (executed_sql, compiled_sql), a rewritten query, a swapped computation file, or a mutated dependency fails the check. A typed, parameter-only surface is what makes "did the sanctioned thing run" a mechanical comparison rather than a judgement call.

10.4 Concepts that use a computation

A document is rarely a single computation. An income-statement overview that discusses revenue, profit, and margin stays one readable concept and links to one Attested Computation per figure:

---
type: Metric
title: Revenue
description: Recognized revenue for a fiscal year.
tags: [finance, revenue]
status: stable
generated: { by: reference_agent/gemini-2.5-pro, at: 2026-06-20T22:53:05Z }
---

# Definition

Recognized revenue sums `amount` over rows booked to the fiscal year,
computed by [the revenue computation](../computations/revenue.md).

Because each computation is its own concept, revenue can be fresh while profit is past its stale_after, and each attests on its own run. Co-locating them is a directory choice (a computations/ folder with an index.md), not a frontmatter one.

10.5 How a consumer uses it (informative)

This subsection is informative, not normative. The runtime artifacts below are not stored in the bundle.

Discover via type: Attested Computation, a frontmatter signal liftable into index.md; a consumer reaches one directly or by following a link from a concept that uses it.
Load the contract from frontmatter and the computation from the body (or the file named by computation).
Parameterize: the agent supplies values for the declared parameters.
Execute: the executor runs the bound computation and returns a receipt shaped by executor.receipt.
Attest: the consumer runs the attester over the receipt. It confirms provenance (the computation that ran equals computation bound with the claimed parameters, not agent-authored SQL) and fidelity (the displayed value matches the receipt's authoritative source, re-read by job id rather than taken from the agent's text).
Gate: refuse to display a failing attestation; warn or refuse when today >= stale_after. On success, surface the verdict (for example a link to the job log) so trust is visible.

10.6 Verification versus attestation

verified (§5.2) and attestation are distinct, and both exist:

verified confirms the definition still matches policy. It is doc-level, slow, and recorded in the bundle.
Attestation confirms a single run produced the value the sanctioned way. It is per-call, runtime, and not stored in the bundle.

A concept with a stale definition can still attest cleanly, and a freshly-verified definition still requires attestation on each run, which is why both are needed.

11. Conformance

A bundle is conformant with OKF v0.2 if:

Every non-reserved .md file in the tree contains a parseable YAML frontmatter block.
Every frontmatter block contains a non-empty type field.
Every reserved filename (index.md, log.md) follows the structure in §8 and §9 respectively when present.

When the trust, lifecycle, provenance, or computation families are present, producers SHOULD follow §5 through §10, and consumers:

MUST treat a bare verified mapping as a one-element list (§5.2).
MUST NOT reject a concept for missing any optional family (§5.3).
SHOULD derive trust tiers and staleness only from the fields specified here, and SHOULD surface, not silently drop, a failing attestation (§10.5).

Consumers SHOULD treat all other constraints as soft guidance. In particular, consumers MUST NOT reject a bundle because of:

Missing optional frontmatter fields.
Unknown type values.
Unknown additional frontmatter keys.
Broken cross-links.
Missing index.md files.

12. Versioning

This document specifies OKF version 0.2. Revisions are versioned as <major>.<minor>:

A minor version bump introduces backward-compatible additions (new optional fields, new conventional section headings).
A major version bump may make breaking changes (renaming required fields, changing reserved filenames).

Bundles MAY declare the version they target with okf_version: "0.2" in a bundle-root index.md frontmatter block (the only place frontmatter is permitted in an index.md). Consumers that do not understand the declared version SHOULD attempt best-effort consumption rather than refusing the bundle.

Considered and deferred

The following are intentionally left to a future revision:

The full runtime protocol: receipt and verdict wire formats, and the attestation lifecycle around a run.
The attester ABI, portability, and sandboxing, likely bundled with future work on serving and Skills.
Attestation caching.
Semantic-layer templates (Looker, dbt) where the attester comparison shifts from SQL equality to model-and-binding equality.

13. Changes from v0.1

v0.2 supersedes OKF v0.1 and is a minor version bump under §12, except for two deliberate breaking changes called out below because they rename or retire v0.1 fields. A v0.1 bundle is consumable by a v0.2 consumer under the fallbacks noted here.

13.1 Breaking changes

timestamp is superseded by generated.at. A concept's last content change is now recorded as generated: { by, at } (§5.2). Consumers MAY fall back to a legacy timestamp when generated is absent.
The body # Citations list is superseded by sources. Provenance moves to frontmatter (§5.1). Consumers SHOULD read sources and MAY still parse a legacy # Citations body list for v0.1 documents.

13.2 Additive changes

All of the following are additive: new optional keys, one new concept type, and one new conventional heading. Their absence yields a plain v0.1 concept.

New frontmatter families: sources with its per-source credibility signals (author, usage_count, last_modified) and the usage_window sibling; generated, verified; status, stale_after (§5).
New concept type Attested Computation and its computation keys runtime, parameters, computation, executor, attester (§10).
New conventional body heading # Computation (§4.2).
The actor convention for generated.by and verified[].by (§7).

Everything else (bundle structure, reserved filenames, the required type, recommended title/description/resource/tags, cross-linking, index files, log files, permissive conformance) is carried forward unchanged.

Appendix A: Worked example, an income statement

One bundle exercising every family, shown as a v0.1 to v0.2 migration of an income statement with two figures, revenue and gross profit.

v0.1 form

A single doc: both figures in one concept, the SQL in prose an agent can read, ignore, or rewrite, citations a flat list, and the only timestamp is timestamp.

---
type: Metric
title: Income statement (fiscal year)
description: Headline income-statement figures for a fiscal year.
tags: [finance, income-statement]
timestamp: '2026-05-28T22:53:05+00:00'
---

# Definition
The income statement reports revenue and gross profit for a fiscal year.

# Revenue
Recognized revenue sums `amount` over rows booked to the fiscal year:

    SELECT SUM(amount) AS revenue
    FROM finance.recognized_revenue
    WHERE fiscal_year = <year>

# Gross profit
Gross profit by segment, per the cost-allocation standard:

    SELECT gross_profit FROM fct_income_statement
    WHERE fiscal_year = <year> AND segment = <segment>

# Citations
- https://wiki.acme/finance/fpa-handbook
- https://wiki.acme/finance/revenue-recognition
- https://wiki.acme/finance/cost-allocation

v0.2 form

The two figures split into attested computations linked from a narrative concept. Every family is populated, and the two computations sit in deliberately different states so one consumer reaches two verdicts.

bundles/finance/
  metrics/income-statement.md      type: Metric  (narrates, links both)
  computations/revenue.md          type: Attested Computation  (runtime: bigquery)
  computations/profit.md           type: Attested Computation  (runtime: dbt)
  references/skills/run-on-bq.md, run-dbt.md
  references/attesters/sql-equality.py, dbt-binding.py

metrics/income-statement.md, the readable doc; trust lives on what it links, not here:

---
type: Metric
title: Income statement (fiscal year)
description: Headline income-statement figures for a fiscal year.
tags: [finance, income-statement]
status: stable
generated: { by: reference_agent/gemini-2.5-pro, at: 2026-06-20T22:53:05Z }
verified: { by: human:ahormati, at: 2026-06-25T09:00:00Z }
stale_after: 2026-12-31
sources:
  - id: fpa-handbook
    resource: https://wiki.acme/finance/fpa-handbook
    title: FP&A reporting handbook
---

# Definition
The income statement reports [revenue](../computations/revenue.md) and
[gross profit](../computations/profit.md) for a fiscal year, per the FP&A
reporting handbook.[^fpa-handbook] Each figure is produced by a sanctioned,
attestable computation; this concept only narrates them.

[^fpa-handbook]: FP&A reporting handbook

computations/revenue.md, BigQuery SQL, human-verified, fresh, and corroborated by a live dashboard source carrying credibility signals:

---
type: Attested Computation
title: Revenue for fiscal year
description: Recognized revenue for a fiscal year, per Finance's definition.
tags: [finance, revenue]
status: stable
runtime: bigquery
parameters:
  - { name: year, type: integer, required: true }
executor:
  resource: references/skills/run-on-bq.md
  receipt: [job_id, executed_sql, result]
attester:
  resource: references/attesters/sql-equality.py
generated: { by: reference_agent/gemini-2.5-pro, at: 2026-06-28T14:00:00Z }
verified: { by: human:ahormati, at: 2026-06-25T09:00:00Z }
stale_after: 2026-12-31
sources:
  - id: rev-policy
    resource: https://wiki.acme/finance/revenue-recognition
    title: Revenue recognition policy
    author: team:finance-fpa
    last_modified: 2026-04-02
  - id: exec-rev-dash
    resource: dashboards/exec-revenue
    title: Executive revenue dashboard
    author: team:finance-fpa
    usage_count: 5000
    last_modified: 2026-06-18
usage_window: { from: 2026-06-01, to: 2026-06-30 }
---

# Computation

    SELECT SUM(amount) AS revenue
    FROM finance.recognized_revenue
    WHERE fiscal_year = @year

Recognized revenue per the recognition policy,[^rev-policy] corroborated by
the executive revenue dashboard.[^exec-rev-dash]

[^rev-policy]: Revenue recognition policy
[^exec-rev-dash]: Executive revenue dashboard

computations/profit.md, a dbt model, process-verified, and past its stale_after:

---
type: Attested Computation
title: Gross profit for fiscal year
description: Gross profit by segment for a fiscal year, per the cost-allocation standard.
tags: [finance, profit]
status: stable
runtime: dbt
parameters:
  - { name: year, type: integer, required: true }
  - { name: segment, type: string, required: true }
executor:
  resource: references/skills/run-dbt.md
  receipt: [run_id, compiled_sql, result]
attester:
  resource: references/attesters/dbt-binding.py
generated: { by: reference_agent/gemini-2.5-pro, at: 2026-06-14T14:00:00Z }
verified: { by: process:finance-nightly, at: 2026-06-12T08:00:00Z }
stale_after: 2026-06-15
sources:
  - id: cost-alloc
    resource: https://wiki.acme/finance/cost-allocation
    title: Cost allocation standard
---

# Computation

    SELECT gross_profit
    FROM {{ ref('fct_income_statement') }}
    WHERE fiscal_year = {{ var('year') }}
      AND segment = {{ var('segment') }}

Gross profit by segment per the cost-allocation standard.[^cost-alloc]

[^cost-alloc]: Cost allocation standard