Helianth Project - A fractal decision architecture from a single decaying primitive

Helianth is a decision architecture built from a single primitive, run at every scale. An agent holding many concurrent commitments — projects, threats, maintenance, its own attention — must repeatedly choose which to act on. Each loses value while unattended, and each has a point where waiting stops being cheap. Rather than score tasks on separate axes of importance, urgency, risk and cost, Helianth models every one of them as the same object: a thread whose value drifts and can suddenly kink, formally the state of a piecewise-deterministic Markov process whose jumps fire at state-dependent hazard rates, with five effects exhausting the jump kernel. Selection is then one comparison — act when the expected loss averted exceeds what the effort costs at the shared reservoir's current level. Because shared resources are threads too, and because the containment rule composes into a lattice rather than a strict hierarchy, the rule that picks a next action is the rule that allocates a quarter, unchanged. That comparison is an index policy of the Whittle/Gittins family, and no novelty is claimed for its form. What this work contributes is a measurement of what its one-step version costs. On rested arms, where the Gittins index is exactly optimal and therefore ground truth, the rule is exactly optimal on declining reward profiles and up to 35.94% short on back-loaded ones. On restless arms — the regime the architecture targets, where Whittle is only asymptotically optimal, so an exact joint dynamic program is the yardstick and both indices are heuristics measured against it — the rule lands within 1.3% of optimal on six of eight instances, never more than 0.3 percentage points behind Whittle on those six, exactly optimal where Whittle is not on one, and a flat 1.5–2% behind it out to 250 threads. One structural blindness of the one-step form is identified rather than hidden: it prices the loss averted now and nothing of the future safety a service buys, tying two arms Whittle separates. This deposit contains: Act Before the Kink (short) — 10 pages. Start here. The selection rule as an index policy and the full measurement above. Self-contained; nothing in it depends on the companion. Act Before the Kink — 46 pages. The architecture the rule belongs to: the decaying-thread primitive, the five-effect grammar, the fractal lattice composition, risk as trajectory dynamics rather than a tuning parameter, and a battery of worked falsifications, each stating what it was built to break and whether the design survived. demo/ — the runnable engine, a browser demonstration, a world builder, and a dependency-free test suite of 444 tests across 72 files that reproduces every table in both papers. Requires Node.js 24+ and nothing else; no install step. © 2026 Michael K. Turner. The papers are licensed CC BY 4.0; the code is licensed Apache-2.0.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-07-27
DOI
https://doi.org/10.5281/zenodo.21413002
Primary Topic
Leadership, Behavior, and Decision-Making Studies
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

Helianth Project - A fractal decision architecture from a single decaying primitive

M Turner
Zenodo (CERN European Organization for Nuclear Research)
Leadership, Behavior, and Decision-Making Studies
preprint

Helianth Project - A fractal decision architecture from a single decaying primitive

M Turner
preprint en

Abstract

Helianth is a decision architecture built from a single primitive, run at every scale. An agent holding many concurrent commitments — projects, threats, maintenance, its own attention — must repeatedly choose which to act on. Each loses value while unattended, and each has a point where waiting stops being cheap. Rather than score tasks on separate axes of importance, urgency, risk and cost, Helianth models every one of them as the same object: a thread whose value drifts and can suddenly kink, formally the state of a piecewise-deterministic Markov process whose jumps fire at state-dependent hazard rates, with five effects exhausting the jump kernel. Selection is then one comparison — act when the expected loss averted exceeds what the effort costs at the shared reservoir's current level. Because shared resources are threads too, and because the containment rule composes into a lattice rather than a strict hierarchy, the rule that picks a next action is the rule that allocates a quarter, unchanged. That comparison is an index policy of the Whittle/Gittins family, and no novelty is claimed for its form. What this work contributes is a measurement of what its one-step version costs. On rested arms, where the Gittins index is exactly optimal and therefore ground truth, the rule is exactly optimal on declining reward profiles and up to 35.94% short on back-loaded ones. On restless arms — the regime the architecture targets, where Whittle is only asymptotically optimal, so an exact joint dynamic program is the yardstick and both indices are heuristics measured against it — the rule lands within 1.3% of optimal on six of eight instances, never more than 0.3 percentage points behind Whittle on those six, exactly optimal where Whittle is not on one, and a flat 1.5–2% behind it out to 250 threads. One structural blindness of the one-step form is identified rather than hidden: it prices the loss averted now and nothing of the future safety a service buys, tying two arms Whittle separates. This deposit contains: Act Before the Kink (short) — 10 pages. Start here. The selection rule as an index policy and the full measurement above. Self-contained; nothing in it depends on the companion. Act Before the Kink — 46 pages. The architecture the rule belongs to: the decaying-thread primitive, the five-effect grammar, the fractal lattice composition, risk as trajectory dynamics rather than a tuning parameter, and a battery of worked falsifications, each stating what it was built to break and whether the design survived. demo/ — the runnable engine, a browser demonstration, a world builder, and a dependency-free test suite of 444 tests across 72 files that reproduces every table in both papers. Requires Node.js 24+ and nothing else; no install step. © 2026 Michael K. Turner. The papers are licensed CC BY 4.0; the code is licensed Apache-2.0.

Zenodo (CERN European Organization for Nuclear Research)
Leadership, Behavior, and Decision-Making Studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.