Instance #2 — Kosplora, and why it (not Global Steward) is the rule-of-three test

Scoping analysis, 2026-06-20. Companion to THE-VALUES-LAYER.html §7 (the rule of three) and STANDARD-v0.html. Grounded in research on open-education rubrics and civic-consensus systems.

One instance is an app; two is a substrate. A second instance exists to falsify the claim "this substrate is general." A falsification test is only informative if it reuses the substrate and varies the domain. Judged by that rule, Kosplora is the right instance #2; Global Steward is instance #3 (the commons layer), not #2.

1. The verdict

**Build Kosplora as instance #2. Park Global Steward as instance #3 — built on top of a proven engine, not instead of one.**

2. Why Kosplora is a clean configuration, not a rewrite

Open education is the rare domain where the rating dimensions are already professionally formalized, sourced, and public-domain — you ingest the criteria rather than invent them. That makes it an unusually honest transferability test: if the engine ranks learning resources as cleanly as it ranks banks, generality is demonstrated; if it leaks, you find out cheaply.

Entity = a learning resource (a course, book, tool, channel, OER object). (The Achieve rubric's "object = smallest stand-alone meaningful unit" answers the granularity question.)

Criteria — drawn straight from the existing instruments, each sourced exactly like CC's:

Values — the learner's portable passport, re-weighting those criteria. This is the whole point, and the place every incumbent fails. Class Central gives one global Bayesian ranking; MERLOT gives one set of expert scores. None of them re-weight by the individual. Yet a self-taught career-switcher weighs openness + time_to_value; a university adopter weighs accessibility + standards-alignment; a depth-seeker weighs rigor + deeper-learning. Re-weighting by portable values is precisely what CC's engine already does — which is the strongest possible signal that the substrate transfers.

Note how this maps onto the existing 8 themes with almost no new vocabulary: openness→Honesty/Privacy, rigor/pedagogy/effectiveness→a (new) Learning theme or People, accessibility/time_to_value→Cost&ease, currency→Health(freshness) or its own. The Values Passport is almost domain-portable as-is — and the small gaps it reveals are exactly what a rule-of-three test is supposed to surface.

3. The one genuine stress (a feature, not a blocker)

Currency decays. A 2019 web-dev course can be obsolete; a brand's labor record ages far more slowly. Open education makes time-decay of a criterion load-bearing in a way consumer ethics doesn't. This is a parameter on an existing criterion, not a new primitive — and it's the ideal kind of mild stress: it teaches the Standard whether Criterion needs an optional half_life / decay field, discovered on a real second domain rather than guessed. Likely outcome: v0 → v0.1 adds an optional decay hint. That is the rule of three working as designed.

4. Why it's a real product (its reason to exist)

Open educational resources have one chronically documented failure: discoverability. Learners and teachers "can't find high-quality OER suitable for their needs" because repositories are fragmented by region/institution/project and metadata is absent or non-uniform. A values-weighted ranking over a unified Entity×Criteria index is a direct answer to a problem the literature names explicitly. And there's timing: Common Sense Education is pausing its edtech reviews in February 2026 — a curation vacuum opening exactly as a values-ranked alternative could fill it. Kosplora is a genuine product, not a demo.

5. Honest caveats (carry these in)

• effectiveness is the hardest to source — true learning-transfer data rarely exists. Proxy it transparently (completion + review sentiment + engagement) and label it a proxy, per the Standard's provenance-first invariant. Don't fake it.
• Crowd-review corpora (Class Central / Goodreads-style) are often proprietary; the expert rubrics (Achieve CC-BY) are reusable, the review data may not be. Confirm data rights before depending on them — prefer open/own-collected signal.
• Don't re-skin CC. The transfer test only means something if Kosplora gets its own presentation (L3) — proving the engine/data (L0–L2) separate cleanly from the shell. A learning atlas should look nothing like a consumption map.

6. What "instance #2" concretely is (the minimal honest build)

The point is not a full product — it's the falsification test, run as cheaply as possible:

1. One Kosplora lens in the Standard's format: ~15–25 real learning resources, the criteria above, sourced provenance (ingesting Achieve/MERLOT/Common Sense scores where licensed).
2. The same engine.js, imported unchanged. If it ranks learners' resources by their values without a single edit to the engine, L0–L2 transferability is proven. If it needs an edit, that edit is the abstraction the rule of three was built to find (likely: the currency decay field).
3. A distinct minimal shell (even a rough one) over the same engine — proving L3 skinnability.
4. Write down what leaked. Every edit the engine needed → a Standard v0.1 revision. That is the deliverable: not Kosplora-the-product, but Standard v1, forced into existence by a real second domain.

This can be a weekend's proof, not a quarter's project — which is itself the test of whether the substrate is real.

7. Global Steward — instance #3, the commons layer (why later)

Global Steward is the more important and more novel project. It is the wrong transferability test now, because it doesn't exercise the values-ranking engine at all — it forces a different machine:

• A write/commons layer. Users create and deliberate on statements/proposals (the pattern of Pol.is, Decidim, Loomio), with state, moderation, and binding thresholds. CC's substrate is deliberately read-only; this is the very layer we deferred. Building it to satisfy a rule-of-three would front-load the most speculative, least reusable infrastructure first — backwards.
• Pseudonymous, Sybil-resistant identity. Cohort consensus needs one-human-one-voice without surveillance. The research is blunt: there is no clean deployed answer — BrightID (social-graph, surveillance-light, and a natural fit for Dunbar cohorts), Gitcoin Passport (stacked weak credentials), ZK proof-of-personhood are all trade-offs on the {Sybil-resistance, pseudonymity, decentralization} triangle. World ID's biometric orb maximizes Sybil-resistance but destroys pseudonymity — the anti-pattern. Betting transferability on an open research frontier is unwise.
• Bridging-rank, not values-rank — the deep one. CC's engine ranks Entities by Criteria weighted by an individual's Values (deliberately personal: my ranking ≠ yours). Civic consensus does the opposite: Pol.is clusters opinion (k-means) and elevates statements agreed across clusters; Community Notes uses matrix factorization over a person×statement matrix, publishing a note only when people who usually disagree both rate it helpful (the note-intercept = broad helpfulness; the note-factor = polarization). That is a different objective function — find the common ground, not serve the individual. You would not reuse the ranking core; you'd write a new one (and rightly draw on Ovadya & Thorburn's Bridging Systems and the Prosocial Ranking Challenge's evidence that bridging measurably lowers polarization).

So Global Steward is a coherent, exciting instance #3: bridging-rank as a sibling to values-rank, a write/commons layer, and surveillance-light cohort identity — built after Kosplora has shown the read engine generalizes, so the commons extends a proven core instead of substituting for it.

8. Recommendation

1. Build the Kosplora falsification test (§6) when ready — a thin lens + the unchanged engine + a distinct shell. Its real output is Standard v1, revised by contact with a genuinely different domain.
2. Park Global Steward as instance #3 / the commons layer — and, when its time comes, co-design L5 (the decentralized commons) with it, since its governance/identity/bridging needs are exactly what L5 must answer.
3. Sequence honored: read-engine generality (Kosplora) before write-commons novelty (Global Steward). Prove the brick before pouring the foundation for the cathedral.

And the standing caveat from the north star: none of this should pull effort from getting Conscious Consuming itself in front of real people. Instance #2 is the forcing function for the architecture; real users are the forcing function for the project. The second matters more.

9. Result — the test was built and run, and it PASSED (2026-06-20)

The falsification test exists, twice over:

• Headless & reproducible — research/kosplora_test.js imports app/engine.js unchanged (via module.exports) and ranks 12 learning resources for four learners.
• Visual — kosplora/index.html, a standalone page with a deliberately different skin (dark "learning atlas," indigo, serif display), loading the same ../app/engine.js.

What transferred (unedited): score(), band()/bandFill(), scoreTier(), the coverage/confidence model, the "strongest on" reasons — and even the FLIP "living ranking" animation — all worked on a new domain with zero edits to the engine. The ranking core is genuinely domain-agnostic.

What leaked (exactly one thing): the value vocabulary (THEMES + KEY2THEME) was baked into the engine, so a learner's portable values couldn't reach education-specific criteria — themeDefaults projected unknown keys to neutral. This is precisely the abstraction the rule of three exists to find.

The fix → Standard v0.1: themeDefaults(criteria, passport, k2t) now takes the key→theme map as an optional parameter (defaulting to CC's, so CC is untouched). Kosplora injects its own vocabulary (rigour / open / ease / fresh); the engine bumped to v0.1. One backward-compatible line — the entire deliverable of the test.

The result:

Rigour-first      #1 Stanford Encyclopedia of Philosophy   top3: SEP, 3Blue1Brown, freeCodeCamp
Well-taught/easy  #1 Khan Academy                          top3: Khan, 3Blue1Brown, freeCodeCamp
…the full list reorders for every value profile…
PASS — different values → different top picks, via a vocabulary the engine never knew. The substrate transfers.

A rigour-seeker and a well-taught-&-easy learner get different #1s from the same engine and the same facts. That is transferability, demonstrated — not argued.

One honest finding (kept, not hidden): the 3↔5 weighting is intentionally mild, so a broadly-excellent option (here SEP — deep, open, current) holds the top against single-axis prioritisation, exactly as in CC. Values reorder the field and move the top pick (Khan wins "well-taught & easy"), but don't flip it for every value. A real property of the engine, surfaced only by running a second instance.

Verdict: the values-ranking substrate is transferable — proven, at the cost of one backward-compatible revision (Standard v0.1). One instance was an app; two is a substrate. Global Steward (instance #3) remains the home for the eventual write/commons layer.