docs(blog): OKF vs StrayMark — "What the open format left to the producer" (EN/es/zh-CN) (#291)

Publishes the dev-blog post on Google's Open Knowledge Format (OKF v0.1) and how
it relates to StrayMark, in all three locales (website/blog + i18n/es + i18n/zh-CN).

The post's spine is the paradigm difference, not a feature gap: OKF builds
cognition *for the agent* (lifting the burden of knowing the system off people,
in line with the long-running-autonomous-agent trend); StrayMark builds cognition
*for the agent and the human engineer together* (AI-augmented engineering — the
human repositioned into judgment/direction, not displaced; Loom as a didactic aid
for spatio-temporal location). Non-defensive throughout: convergence as
validation, interop over competition.

Also refines the experiment-okf analysis docs with the same lens:
- Correct the genealogy: StrayMark reached the primitives independently; the
  Karpathy gist is OKF's lineage, not StrayMark's (removes "DNA"/"siblings"/
  "headwater both drink from"/"same intuition under StrayMark").
- Elevate the "who is the knowledge for" axis above the data-vs-governance domain
  split in 01/02/README.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: montfort

experiment-okf/01-okf-vs-straymark-analysis.md

@@ -7,9 +7,15 @@
 
 OKF and StrayMark share the **same mechanics** — the "LLM-wiki pattern":
 markdown + YAML frontmatter + a graph of links + an agent that maintains the
-corpus — but operate in **orthogonal domains**. OKF describes *the data and
-systems agents consume*; StrayMark records *the governance of what agents do*.
-Not a frontal competitor; a strong validation of the design bet and a real
+corpus — which StrayMark reached **independently** (OKF credits Karpathy; we
+didn't get here from him). But they obey **opposite paradigms of AI use**: OKF
+builds cognition *for the agent* — lifting the burden of knowing the system off
+people, in line with the dominant long-running-autonomous-agent trend — while
+StrayMark builds cognition *for the agent and the human engineer together*,
+keeping a person oriented and in command. The domains differ too (OKF describes
+*the data and systems agents consume*; StrayMark records *the governance of what
+agents do*), but that's downstream of the deeper split: **who the knowledge is
+for.** Not a frontal competitor; a strong validation of the design bet and a real
 interoperability opportunity. The risk is **mindshare capture**, not
 replacement.
 
@@ -46,13 +52,25 @@ Google arrived — independently — at almost every design decision StrayMark m
 | Tolerant consumption | broken links are "not-yet-written knowledge", unknown keys preserved | dangling refs kept as first-class `resolved:false` edges |
 
 That Google converged on markdown+frontmatter+graph+agent is the strongest
-possible evidence the bet was right. They even cite Karpathy and the "LLM-wiki
-pattern" as the genesis — the same intuition under StrayMark.
+possible evidence the bet was right — *because the two projects started from
+different places.* OKF cites Karpathy and the "LLM-wiki pattern" as its genesis;
+StrayMark reached the same intuition on its own, without that lineage. Independent
+arrival at the same primitives is what makes the convergence evidence rather than
+imitation.
 
-## 4. The decisive difference: domain and intent
+## 4. The decisive difference: who the knowledge is for
 
 This is where they stop touching.
 
+The split isn't only domain; it's **who each format's knowledge is ultimately
+for.** OKF encodes a system so an *agent* can understand and operate it with as
+little human involvement as possible — the autonomous-agent direction the industry
+is racing toward. StrayMark encodes how a system was built so a *human engineer*
+stays oriented and in command alongside the agent — the "AI-augmented engineering"
+position, where the human is repositioned into judgment and direction, not
+displaced. One faces the agent; the other faces the agent **and** the human at
+once. The domain difference below is a consequence of that, not the root of it:
+
 - **OKF = a semantic / context layer over data.** It encodes *what this table
   means, what this metric is, the join paths, the runbook* so an AI agent can
   reason over an organization's data estate (BigQuery, Knowledge Catalog). It is

experiment-okf/02-karpathy-llm-wiki-genesis.md

@@ -1,12 +1,15 @@
-# The Karpathy gist — genesis of the LLM-wiki pattern, and what it means for StrayMark
+# The Karpathy gist — OKF's headwater, and an independent check on StrayMark's bets
 
 *Prepared 2026-06-17. Source: Andrej Karpathy, gist `442a6bf555914893e9891c11519de94f`
 (the pattern OKF's announcement explicitly credits as its origin), plus the
 community replies on that gist.*
 
 This is the "extra analysis step" requested: the gist is the intellectual
-headwater both OKF and StrayMark drink from, and its comment thread already
-mapped out the failure modes — several of which StrayMark answers by design.
+headwater **OKF's lineage** drinks from — StrayMark reached the same shape
+independently, without knowledge of it. What makes the gist worth reading anyway
+is its comment thread, which already mapped out the pattern's failure modes —
+several of which StrayMark answers by design. That's **corroboration, not
+inheritance.**
 
 ## 1. The proposal
 
@@ -32,10 +35,10 @@ rather than auto-resolving). The rationale: humans find wiki upkeep tedious;
 it in Vannevar Bush's 1945 Memex — associative trails that finally have a tireless
 maintainer.
 
-## 2. Why this is StrayMark's DNA too
+## 2. Where StrayMark lands on the same shape — independently
 
-StrayMark is, structurally, an instance of this pattern aimed at one domain —
-software governance:
+StrayMark independently lands on the same structural shape, aimed at one domain —
+software governance. It didn't derive from the pattern; it converges with it:
 
 | Karpathy's pattern | StrayMark |
 |---|---|
@@ -47,9 +50,12 @@ software governance:
 | Lint (contradictions/stale/orphans/gaps) | `validate`, `charter drift`, follow-ups drift, dangling-ref classification |
 | "LLMs don't forget cross-references" | typed frontmatter links resolved by `core::graph` |
 
-So OKF and StrayMark are **siblings, not rivals** — two specializations of the
-same parent idea. OKF specializes it for *data context*; StrayMark for
-*development governance*.
+So OKF and StrayMark are **convergent, not derivative** — two projects that
+arrived at the same shape from different starting points (OKF down Karpathy's
+lineage, StrayMark on its own) and specialized it differently. And not only for
+different domains — *data context* vs *development governance* — but for different
+**readers**: OKF builds cognition for the agent; StrayMark builds it for the agent
+and the human engineer together.
 
 ## 3. The community replies already mapped the failure modes — and StrayMark answers several
 
@@ -97,10 +103,11 @@ where StrayMark's opinionation earns its keep.
 
 ## 4. Takeaways
 
-1. **Convergence at three levels.** The same idea produced (a) Karpathy's
+1. **Convergence, not descent.** The same shape recurs across (a) Karpathy's
    pattern, (b) Google's OKF spec, (c) community implementations (memwiki), and
-   (d) StrayMark — independently. The pattern is no longer a differentiator; the
-   **domain and the opinionation** are.
+   (d) StrayMark — which reached it independently, with no knowledge of the gist.
+   The pattern is no longer a differentiator; the **reader it serves** (agent
+   alone vs agent + human) and the **opinionation** are.
 2. **StrayMark answers the gist's hardest critique.** Typed edges that preserve
    contradiction (pursultani) are designed-in via `supersedes`/`alternatives`/TDE
    — exactly where OKF chose untyped links. This is the sharpest line to draw.