A lens for grading any capability, principle, or technique on two independent axes — Human and AI — not one balance bar. They are not a tradeoff: a thing can be Human 5 and AI 4. Each axis decomposes into five facets (a penta), and every item lands somewhere on the Human × AI plane, which sorts into four zones. This is the model behind the design scorecard and the score badge on the design docs.

The core claim, inherited from Higher-Order Generativity vs Higher-Order Judgment: what stays human is the part that is not interpolation — out-of-distribution origination and accountable discernment, not the median synthesis between them. AI is strongest at producing and recalling the average of everything, fast and cheaply, in verifiable domains.

The plane and its four zones

Plot Human (vertical) against AI (horizontal). Because the axes are independent, “both high” is its own meaningful cell:

• Own — Human high, AI low. Your call to make; agents can’t. Master these.
• Augment — both high. The agent does the work, you direct and judge. The collaboration sweet spot.
• Delegate — AI high, Human low. Hand it off.
• Low-leverage — both low. Context-bound or parked.

The derived field on every scored item is its Zone, not a Human-minus-AI skew (skew would erase the Augment cell, which is the whole point).

The Human axis — five facets

Two pillars: Discernment (choosing the good) and Origination (bringing the good into being, and owning it).

Taste · Judgment · Originality · Specific Knowledge · Accountability

Discernment

• Taste — recognizing what’s good or great, and what to cut, before you can say why. Durable because AI makes creation free and fills the world with average; when generation is cheap, the scarce act is selection, curation, the eye.
• Judgment — the accountable call in a messy, multi-constraint situation with incomplete information, where being wrong is costly and there’s no retry. Durable because reliability under genuine novelty is where models are weakest.

Origination

• Originality — the out-of-distribution move made with intent: something genuinely surprising that shifts the trajectory, not interpolating the median. Durable because models interpolate within their training; stepping outside the system (Gödel breaking the formal system; “art” as meaningful OOD behavior) is the ceiling.
• Specific Knowledge — un-trainable, curiosity- and obsession-grown knowledge; the idiosyncratic blend of experience that resists schooling and automation. Durable because it isn’t in the training set — it’s yours.
• Accountability — taking the risk under your own name: signing off, carrying the consequences, the trust that forms only when you’re on the hook. Durable because a model cannot be accountable — it can’t be punished, rewarded, or trusted to stand behind a call. The one facet categorically impossible for a model.

On Agency. The disposition to act and direct (“what should I build?”) is not a sixth human facet — it sits above the five as the will that deploys them. Its gradable form lives on the AI axis as Autonomy. See the Naval “intelligence vs agency” debate in The AI Industrial Revolution.

The AI axis — five facets

Two pillars: Production (raw output capability) and Reliability (whether you can trust it unsupervised).

Fluency · Knowledge · Scale · Verifiability · Autonomy

Production

• Fluency — generating coherent, integrated output (prose, code, design); the generativity now at or above the median professional.
• Knowledge — breadth of recall and pattern-matching across a vast corpus; the average of everything, on demand.
• Scale — speed, parallelism, near-zero marginal cost; tireless repetition. “Waste tokens, save time.”

Reliability

• Verifiability — how checkable or specifiable the task is; whether the agent can self-check against a clear success criterion. The boundary condition: AI wins decisively where verification is cheap, and degrades at the creative frontier where it isn’t.
• Autonomy — agentic multi-step execution: decompose a goal, plan, use tools, and run it end-to-end without hand-holding. (Agency’s gradable home.)

Model snapshots (dated)

Relative profiles on the five AI facets — judgment calls, not benchmarks, snapshot 2026-07 (Sonnet 5 replaces Sonnet 4.6 this pass). The point: every “higher-order” model is a larger polygon in a different direction, which is why order is the polygon’s size, not one of its axes. Opus leads the Reliability side (Verifiability, Autonomy) and Knowledge; Sonnet 5 matches its Verifiability at 0.6× the price; Fable leads Fluency; Composer and Haiku lead Scale.

Model	Role	Fluency	Knowledge	Scale	Verifiability	Autonomy
Opus 4.8	frontier reasoning + agentic	4	5	2	5	5
Fable	generative / creative	5	4	3	2	3
Sonnet 5	near-Opus, cheaper	4	4	4	5	4
Composer 2.5	mechanical / agentic coding	3	3	5	4	4
Grok 4.3	generalist, mid tier	3	4	4	3	3
Haiku 4.5	fast / cheap tier	3	3	5	3	2

Opus 4.8 frontier reasoning	Fable generative	Sonnet 5 near-Opus, cheaper
Composer 2.5 mechanical / code	Grok 4.3 generalist, mid	Haiku 4.5 fast tier

Composer 2.5 is the mechanical workhorse — Scale and a verifiable, agentic coding loop carry it; it isn’t reached for breadth or prose. Grok 4.3 sits a notch below the frontier today; its July release, trained on Composer data, is projected near Opus parity — projected ≈ Fluency 4 · Knowledge 5 · Scale 3 · Verifiability 5 · Autonomy 5, to be re-graded on release, not asserted now.

The snapshot rots: the Reliability pillar (Verifiability, Autonomy) is closing fastest, and the cheap tier (Composer, Haiku) keeps rising on Scale. Re-grade when a model jumps; the floor moves, the spread persists.

Usable intelligence — Sonnet 5 vs Opus 4.8

For controlled-vocabulary authoring behind a fail-closed gate — the tsumugu-ed and tsumugu-core content lanes — Sonnet 5 is the default author and Opus 4.8 is the orchestrator and the fallback for calls the gate can’t check. Same job, 0.6× the price (0.4× on Sonnet 5’s intro rate through 2026-08-31), near-identical measured gate outcomes.

The comparison is cost-adjusted deployable capability, not peak reasoning. Both publish a 1M context window, 128K max output, and the same low–max effort ladder (both carry xhigh); both run adaptive thinking. They share one tokenizer (the Opus-4.7-era tokenizer), so identical text bills the same token count on each — the sticker gap is the real gap. Opus costs 1.67× Sonnet 5 per token ($5/$25 vs $3/$15 per MTok), 2.5× against the intro rate ($2/$10 through 2026-08-31).

Facet grades (this snapshot): Opus 4.8 = F4 K5 Sc2 V5 Au5; Sonnet 5 = F4 K4 Sc4 V5 Au4. Sonnet 5 trails by one on Knowledge and Autonomy, matches on Fluency and Verifiability, leads by two on Scale (cheaper, faster, no fast-mode premium to buy back throughput).

Polygon area (order = size, this lens’s own scalar; A = ½·sin72°·Σ rᵢrᵢ₊₁ on equal 72° spokes): Σ = 85 for Opus, 88 for Sonnet 5 → area 40.4 vs 41.9. Sonnet 5’s usable-capability polygon is ~3% larger: Opus’s Scale-2 (slow, expensive) cancels its Knowledge/Autonomy edge once cost and speed are on the axes. Raw ceiling favors Opus; deployable capability is a wash.

Per dollar (area ÷ output $/MTok): Opus 40.4/25 = 1.6; Sonnet 5 41.9/15 = 2.8 sticker (1.7× Opus), 41.9/10 = 4.2 intro (2.6× Opus).

Empirical, observational — different lessons through the same fail-closed gates, not a blind bake-off:

• Sonnet 5, tsumugu-core companion readings (author → critic → repair, 2026-07-01): 31 drafts, 3 critic-caught repairs → 28/31 (90%) clean on the first author pass, 100% final gate pass, 100% grammar coverage, ~100% vocab, ~85% of articles rated “good.”
• Opus 4.8, tsumugu-ed simplified A1→C2 encoding (70 parallel agents, 2026-06-21): 982/982 authored; 11/982 (1.1%) needed a repair on the register/leak sweep (3 glyph leaks, 8 register); final 0 leaks, 4 advisory flags all legitimate.
• Both clear the gates at high first-pass rates on different task shapes; neither collapses. The measured quality gap sits inside the confound — task shape, floor, date — so it can’t be the deciding variable. Cost and speed can.

The split, already in use: the 2026-07-01 run ran Opus as orchestrator and Sonnet 5 as author/critic/repair. Opus’s Knowledge and Autonomy edge pays on planning and the long-horizon loop; Sonnet 5’s Verifiability at 0.6× cost pays on the high-volume authoring the gate verifies.

Flip condition: when a lane’s quality bar moves to what the gate can’t check — subtle naturalness, cross-lesson coherence, taste — Opus’s Knowledge/Autonomy edge starts earning its price; promote authoring back to Opus there. The other soft spot is the equal-weighted area itself: reasoning-bound, verification-scarce work should up-weight Knowledge and Autonomy, where Opus leads outright, and the per-dollar verdict narrows or inverts.

How it’s used

• Score badge at the top of every standalone design doc: two independent bars (Human, AI), value-colored (green 4–5, amber 3, red 1–2), with Build and Learning as relevance. On the published site each axis expands to its penta.
• Master scorecard: every Universal Principle and Refactoring UI rule graded Human / AI / Build / Learning, with a Zone, in one sortable table.
• Graduation rule: high-signal items — especially Human-5s like Wabi-Sabi — get their own page; the rest live as rows.

Sources

• Higher-Order Generativity vs Higher-Order Judgment — the generativity/judgment split and the OOD ceiling.
• The AI Industrial Revolution — Naval roundtable; verifier role, intelligence vs agency, “waste tokens save time.”
• The Almanack of Naval Ravikant — specific knowledge, accountability, leverage.
• Naval Ravikant, on judgment & taste as the moat under infinite leverage: Wealest summary, Office Chai — design as an AI moat.

Related

• Design Two-Track Extraction — the Agent/Human split this lens grades.
• What the Model Names Signal — what Opus/Sonnet/Haiku/Fable, and Composer vs Grok, signal.
• Essential AI Skills 2026
• Don’t Outsource the Learning
• Global Workspace and J-space — the interpretability substrate for these axes: automatic processing underwrites the AI Production and Scale facets, deliberate access-conscious work underwrites the human Judgment and Accountability facets.

Open Questions

• Are the two axes truly independent, or does very high AI verifiability quietly lower the Human score over time (today’s Augment sliding to Delegate)?
• Does Accountability belong on the Human penta or one level up with Agency, as a precondition rather than a facet?
• How fast is the AI Reliability pillar (Verifiability, Autonomy) closing? The lens is load-bearing only while that gap holds.