03 · Tunables (the customization surface)¶
Everything in Axioms and Guardrails is
[INVARIANT]. Everything here is[TUNABLE]— the dials a project sets when it adopts CompassAlpha.
In plain terms: this is the section of settings you get to choose. Some rules in CompassAlpha never change; the ones here are the knobs — how fast, how careful, how much to spend — that each team turns to fit its own work. It's for anyone setting up CompassAlpha for a project and deciding how it should run.
New here? Start with the mental model.
[TUNABLE]
What "tunable" means¶
CompassAlpha separates its surface into two halves:
[INVARIANT]— the load-bearing rules (firewall, hard labour, bus protocol, persistence law, dual validation). These cannot be changed without forking the framework. They are structural correctness.[TUNABLE]— the performance/cost surface. Concurrency, context depth, model choice, work granularity, enrichment investment, memory rate. Each project sets these to fit its risk profile, budget, and maturity.
Tuning operates only on the performance/cost surface — never on structural correctness. A "speed-up" that violates the firewall isn't a faster federation; it's a different (more polluted) one. Whenever you read a tunable page, the invariants are still in force behind it.
The 5-axis trade-off¶
Every tunable moves the federation along five axes that are in tension:
| Axis | What it measures |
|---|---|
| Speed | Wall-clock time from cycle activation to deliverable. |
| Intelligence | Depth, correctness, coverage — the federation's "IQ floor." |
| Cost | Compute (tokens, API calls) + founder cognitive load + maintenance burden. |
| Risk | Chance of pollution, hallucination, drift, replay, or stale-snapshot decisions. |
| Predictability | How accurately cycle duration and deliverable scope can be estimated up front. |
No combination optimizes all five. Push speed up and risk tends up + intelligence down. Push intelligence up and cost tends up + speed down. Push cost down and founder load or risk tends up. See Overview for the full geometry.
The pages in this section¶
| Page | The dial it covers |
|---|---|
| Overview — the 5-axis trade-off | The tension space; why no combo wins everywhere. |
| Axis declarations | Defining your own work axes (the 4-field + 3-role declaration). |
| Concurrency modes | Sequential / pipelined / parallel-independent / parallel-doer. |
| Context patterns | Mentor lifecycle, doer scope, cross-tier visibility, memory rate, digest verbosity. |
| The retrieval surface | How a tier pulls the slice on demand — lexical vs vector grade, corpus, chunking, cadence. |
| AI model choices | Per-tier model selection; single vs mixed providers. |
| Work granularity lanes | The four lanes — matching ceremony to change size. |
| Stage taxonomies | Per-axis lifecycle vocabularies used by status grids. |
| Invariants, toolings & agents | The three enrichment surfaces; intelligence ↔ cost compounding. |
| Memory policy | What carries forward across cycles; inheritance + garbage collection. |
| Full parameter matrix | Every tunable in one table, with defaults and ranges. |
Five categories of tunable¶
The full matrix groups every dial into five categories. Knowing the categories helps you reason about which decisions are one-time setup versus ongoing per-cycle choices:
- Project-level — name, repos, host, agent harness, founder identity. Set once at adoption.
- Axis-level — per-axis tier names, deliverable type, cycle stages, Charter posture. Set per axis.
- Operational — concurrency mode, lifecycle, effort, provenance strictness, rotation cadence. Tuned per cycle.
- Cultural — founder role scope, founder-call boundary, pacing. Set by the project's working culture.
- Tech-stack — VCS, remote, commit discipline, worktree mechanism. Mostly invariant; a few choices.
How to start¶
Most adopters do not tune individually. They pick a named operating preset (Conservative, Balanced, Throughput, Cost-optimized, Risk-averse, Research, Bootstrap) and adjust from there. Presets live in Toggles → Operating presets; the parameters each preset sets are documented across this section.
If you are brand new, start with Bootstrap (LAYGO concurrency, top-tier everywhere, minimal enrichment) and graduate as your first cycle proves the rhythm.