Bias Weights — Edge vs Cloud

Sep 29

4 min read

Parts 1–2 got you to a connected Blip and a first collapse. Now we steer the field without flipping the light switch early.

Same math. Different resolution.

Bias is how you lean the spinning coin. Edge is one matrix — one honest angle, fast. Cloud is an ensemble — when one angle is not honest enough. If part 1 was “what collapse is,” this is “how to aim the measurement.”

What bias actually does

Bias is the primary control for directional lean inside Binary Probability Matrices. It decides how probability mass concentrates before the system collapses to a single high-leverage prompt.

Paired with exponents n and m, bias creates a controllable classical superposition over structured alternatives — roles, reasoning patterns, structure, quality, constraints, ethos. Higher bias → stronger lean toward decisive, high-agency outputs.

Mental model from our Bias Weights Guide: bias is the tilt of a spinning coin whose surface has been expanded into a probability grid. n and m control how sharply that tilt is expressed on each axis. Steps let the lean propagate through the matrix. Measurement returns the single most probable structured prompt under that lean.

You will see the shape in docs and code: biasⁿ on one side, (1 − bias)ᵐ on the complement — directed fractional contributions, outer-product blending into the matrix, normalize, evolve, collapse. Classical hardware. Conserved probability mass. No mysticism required.

Recommended lean bands

From the published guide (starting ranges for operators — not quality guarantees):

0.50 – 0.65 — Exploratory / multi-perspective. Best for: Early ideation, trade-off mapping, broad coverage
0.65 – 0.78 — Balanced coverage. Best for: Analysis that still surfaces alternatives
0.78 – 0.85 — Strong but controlled (default band). Best for: Most technical, strategic, and execution work
0.85 – 0.93 — Decisive / high-agency. Best for: Clear commitment, first-principles, shipping focus
0.93 – 1.00 — Maximum lean. Best for: When you need the system to commit hard

Defaults in production tools (guide table):

Edge generate_optimal_prompt — often bias 0.82, n=3, m=1, steps=7, grid_size=5
Ensemble generate_ensemble_prompt — often bias 0.80, steps=6, with num_cubits (example default 10 in the guide)
generate_controlled_prompt — flexible surface near the same default lean

Near 0.50 with high steps / larger grids explores more of the space but can feel less decisive. Very high bias + high n can over-concentrate and reduce diversity of reasoning patterns. That is tuning feedback, not a product defect. Progressive Completeness: start in the default band; move only with intent.

Other knobs that ride with bias

steps — evolution iterations before measurement; more steps refine under the chosen lean.
grid_size — matrix resolution; common default 5; larger grids allow finer distinctions.
num_cubits — ensemble size in Cloud mode; cost scales with cubits.
seed — makes Edge deterministic for a given configuration (A/B your lean).
goal + extra_context — the problem statement and hard constraints that seed the field.

Supporting tools (list_prompt_atoms, set_blip_mode, set_tools_power) let you inspect and steer without turning the session into a shopping cart of templates.

Edge vs Cloud — same substrate

Edge (single matrix)

One Binary Probability Matrix. Fast. Deterministic when seeded. Interactive. Metered in the cheap class (product narrative: Edge optimal ≈ 1 Qx). Best when you already know a good lean band and want rapid iteration on a known goal shape.

Cloud / ensemble

Multiple independent cubits — parallel matrices with slight parameter diversity. Each evolves separately. The system returns the global champion — highest-scoring collapsed prompt across the ensemble. Higher ceiling. Cost scales with cubits (GTM materials illustrate scaling examples; quote live metering in-product, not folklore). Use when one angle is not honest: contested goals, high-stakes briefs, or when you want the field stress-tested before you let the decision leave.

There is no “show me the full probability landscape forever” product mode. The design philosophy is forced high-leverage measurement. The coin spins; then it blips. That is the through-line of this series in one sentence.

Practical workflow

Write a clear goal (+ optional extra_context for voice, length limits, domain background).
Choose Edge first for most work.
Set bias in the 0.78–0.85 band unless you have a reason to explore or commit harder.
Optionally tune n/m/steps; typical settings keep n > m to emphasize the bias direction.
Call the tool via your Infinity Blip MCP connector. Collapse. Feed the prompt to the host.
Iterate. Seed + fixed parameters give reproducible Edge results — useful for A/B on lean.

Compare coherence, specificity, and reduced need for human repair against a raw prompt. That comparison is your real benchmark. We will not publish fake win-rate charts.

What we will not claim

No guaranteed quality. No invented benchmarks. No “always use 50 cubits.” Bias is a control surface. Your goal quality still dominates. Street-smart rule: if the goal is mush, the matrix evolves mush elegantly.

Soft door

From the portal connector you already set up, run Edge on a real goal this week. Change only bias between two runs. Feel the lean. Then decide whether ensemble earns its Qx.

Try Edge first from Members: https://mcp.qxbin.com/members

Next chapter: zoom out from product controls to the chessboard underneath — Magic Grid / Binary Probability Matrices, cubit mind, bias winds, and measurement harvest.