Dendra — brand mark concepts (round 2)

Round 1 shipped four concepts (Dendrite / Ascend / Curve / Gate). Ben's feedback: "A looks like a Y and the asymmetry is something I'm still trying to process; B appeals; C conceptually best but least favorite image; D has subtle power." Round 2 applies that feedback to each concept with orientation + execution variants, then hands the full set to three AI design critics for assessment.

A · Dendrite

Name etymology — the tree of decisions.

A1 · original Asymmetric Y. Accent on right-branch tip. Reads as "one graduation earned."
A2 · symmetric Same Y, but the accent moves into the split node. Perfectly mirrored. "The graduation point IS the ember."
A3 · canopy Two-level branching (4 terminal tips) — less Y-literal, more tree-literal. Accent at center node.
A4 · rooted Trunk rises from a three-root splay. No canopy — the graduation is the ascent itself, accent at the top.
A2-horiz A2 rotated 90°. Input enters from the left, forks at the decision node, two outcome paths. Reads "data-flow / decision fork," not "tree."
A3-horiz A3 rotated 90°. Compound-branching decision tree — one input, four terminal outcome paths. Process time reads left→right.

A1 at sizes · wordmark

16
32
96
200

B · Ascend

Lifecycle as shape — six phases, literally.

B1 · original Six staircase steps rising from the rule floor, top step accent.
B2 · compact Tighter rhythm, fills the central region more fully. Reads as a single dense object rather than scattered steps.
B3 · stack Six horizontal bars at the same x-position — pure vertical stack. Reads as "foundational strata," not as motion. More architectural.

B1 at sizes · wordmark

16
32
96
200

C · Curve

Measurement as mark — Figure 1 collapsed.

C1 · original Curve + 5 checkpoint dots + rule floor. Last dot accent.
C2 · minimal Stripped to one curve + one terminal accent dot + rule floor. Single confident gesture.
C3 · tick marks Checkpoints as vertical tick marks crossing the curve. Reads as a measurement ruler — more "instrumentation."
C4 · step-function Curve rendered as a stepped staircase. Discrete phase advances along a rising path — bridges Curve and Ascend.

C1 at sizes · wordmark

16
32
96
200

D · Gate

Theorem as architecture — the threshold crossed.

D1 · original Floor + two posts + lintel + accent tick rising through the center.
D2 · rising Accent starts below the floor, parts the floor, rises through the gate. Dynamic and directional.
D3 · open threshold Drop the lintel. Two posts on the floor, accent rising between. More gesture, less architecture.

D1 at sizes · wordmark

16
32
96
200

Critic assessments

Three AI design critics with distinct perspectives. Each one saw all 14 marks (4 originals + 10 variants), the B-Tree Labs brand system, and the innovation context, and was asked to give their sharpest opinion.

Critic 1 · Brand strategy + system coherence

Top pick: D1 (Gate · original)

"The only mark in the set that reads as architectural infrastructure — a threshold, a ground, a passage — at 16 px, at 200 px, and five years from now when 'AI-branching-tree' marks look like a generation-tagged cliché."

Requested fixes: widen the posts (taller proportions, more torii, less TV-set); break the lintel-post joint perception (butt-join with a tiny pin at each corner, echoing B-Tree Labs' node vocabulary); thin the accent by ~20% so it reads as a value being measured rather than a third post. Vetoed A1, A3, C1, C3.

"Marks that illustrate the thing they're for age badly. Stripe's mark isn't a payment flow; Cloudflare's isn't a proxy cloud; Temporal's isn't a state machine. They're abstract containers that earn meaning through the product. The Dendra set is trying to pre-load meaning into the glyph, which is exactly the move mature infra brands don't make."

Honest close: "There's no designer in the loop. Pick a direction here, then spend $6-12k with an identity designer (Hoodzpah, Character, Mackey Saturday) for two weeks of refinement." Full transcript →

Critic 2 · Execution + craft

Top pick: D1 (Gate · original)

"The only mark in the set that is architecturally defensible as a standalone glyph. The Π form is closed, symmetrical, and legible as a single silhouette — it owns a shape the way the B-Tree Labs cross owns a shape."

Three specific production-readiness fixes (shipped as D1' — gate-refined in round 3 below): shorten the accent from ~104 px to ~72 px so it reads as a short extension past threshold, not a third post; merge posts + lintel into a single polyline with miter joins; add a 28 r graphite center node at the lintel midpoint so the accent emerges from a dock-point rather than floating on the lintel stroke (ties to B-Tree Labs' node vocabulary).

Small-size survivors: D1, D3, A4, A2. Collapse-to-noise at 16 px: A3, C1, C3, all three B variants, D2. Scrapped entirely: A3, B3, C1, C3, D2.

On D2 specifically: "The 48-px parted-floor gap is wider than the 34-px accent passing through it — the accent doesn't pierce the floor, it walks through a doorway." Fix shipped as D2' — gate-rising-refined: narrow the gap to exactly the accent stroke-width (34 px).

Critic 3 · Conceptual depth + semiotics

Top pick: D2 (Gate · Rising)

"The mark's argument: 'A floor exists, a threshold exists, and the thing that advances is evidence — a single stroke that begins below the floor, punctures it, and crosses the gate.' That is, almost literally, the theorem. The rule-floor is parted BY the accent — the geometry asserts that the floor is not broken but crossed at a measured point, which is exactly what a McNemar rejection does to the null."

First-read signification: Dendrite reads "neural net," Ascend reads "number go up KPI slide," Curve reads "VC pitch hockey stick," Gate reads "threshold / passage." Only Gate's primary semantic field is passage rather than progress — and passage is what Dendra is actually about.

Of the 14 marks, only two couldn't be lifted wholesale into another AI company's identity: D2 · Gate · Rising and (distant second) A4 · Dendrite · Rooted. Everything else is generic.

Fifth direction proposed (shipped as E · Contingency below): "The contingency table. Dendra's mechanism is a paired-proportion test — a 2×2 of {rule_correct, rule_wrong} × {ml_correct, ml_wrong}. A mark built from a square partitioned into four cells — three graphite, one (the b-cell: 'ml correct where rule was wrong') accent — would be the only logo in the category whose geometry is the product's actual math. It reads as a window, a pixel, a confusion matrix, a crosshair. Distinctly not a tree, a chart, or a neural net. Pairs orthogonally with the Axiom cross. A first-time viewer doesn't need to know McNemar to read it as 'four possibilities, one of them is what we're betting on' — which is fundamentally the pitch."

Round 3 · Critic-driven refinements

Refinements applied per the craft critic's specific construction fixes, plus the semiotic critic's proposed fifth direction built fresh.

D1' · gate refined D1 with three fixes: shorter accent (72 px), polyline joins on posts+lintel, 28 r center node at lintel midpoint (docks to B-Tree Labs node vocabulary). Top pick for two of three critics.
D2' · gate rising refined D2 with floor gap narrowed to exactly 34 px (the accent stroke-width). Accent now pierces the floor cleanly instead of walking through a doorway. Top pick for the semiotic critic; "carries the theorem."
E · contingency (light) Critic 3's proposal. Four cells of the McNemar table as glyph; the b-cell (ml correct where rule was wrong) is accent. Only concept in the set whose geometry is the product's actual math.
E · contingency (dark) Dark-ground inversion. Structure cells become warm off-white; b-cell stays accent orange. The accent's semantic role — "the one cell different from the other three" — holds regardless of ground.

Round-3 candidates at sizes · E wordmark

D1' 32
D1' 96
D2' 32
D2' 96
E light 32
E light 96
E dark 32
E dark 96

Next step

Pick among three round-3 candidates:

Once picked, follow-up PR exports the full asset set (favicon 16/32/180/512, dark-ground, monochrome light/dark, stacked wordmark, 1200×630 Open Graph card) matching the B-Tree Labs brand-kit structure, replaces the placeholder favicon, adds the chosen mark to the landing header, and commits a proper brand/ directory at repo root.

Critic 1's closing advice: "Pick a direction here, then spend $6-12k with an identity designer (Hoodzpah, Character, Mackey Saturday) for two weeks of refinement." That's the right order of magnitude for any of the three.