Decide a fan-out's shared-context layout before dispatch, and verify it after.
When an orchestrator fans out to several subagents, each one re-reads the same upstream context. Fructigen takes the fan-out as it is planned and compiles one immutable prefix plan: which shared segments to send in a cache-safe encoded form, which prefix boundaries to mark as cache breakpoints, and in what order to release the calls so that a prefix is written before it is read. The plan carries the predicted token classes of every call, so the provider's reported cache counters can be checked against it.
Status: alpha. The prefix plan has not been measured against a real provider's cache. The cache
model reproduces three recorded probes and is otherwise built from provider documentation, so every
saving the planner prints is a prediction. One mechanism, encode, has been measured inside a real
fan-out: see Benchmarks.
from fructigen.plan import Call, Part, compile_plan
from fructigen.provider import ANTHROPIC_HAIKU_5_5
trunk = Part("recon-handoff", tokens=1200, encoded_tokens=800)
calls = [Call(f"sibling-{i}", (trunk, Part(f"task-{i}", tokens=100))) for i in range(4)]
plan = compile_plan(calls, ANTHROPIC_HAIKU_5_5)
plan.waves # (('sibling-0',), ('sibling-1', 'sibling-2', 'sibling-3'))
plan.calls[1].usage # Usage(read=800, write=0, uncached=100)
plan.saving # predicted fraction of baseline input costToken counts are the caller's to supply. The library ships no tokenizer.
Fructigen is built for the VRIL LABS Parallel Subtask Agent Workflow and measured in its
benchmark, vril-bench. B0 is naive fan-out: every subagent reads the full upstream context.
Every task is scored by a programmatic verifier against generated ground truth; no model grades the
results. Runs are paired by seed, with 95% t-intervals over seeds. The newest set is shown first;
earlier sets are collapsed below it, dated.
The first Fructigen measurement. B0-E is B0 with one change: before the four downstream tasks
read the recon handoff, encode zone-factors its region map. The re-encoding is lossless and guarded;
it applies only when the answer parses, factors exactly and shrinks by at least 1.213×. Everything
else is identical. The quality criterion was pre-registered and blocking.
Ten clean seeds on Claude Haiku 5.5 via OpenCode Go, 16 zones, 480 calls, $0.39:
| B0-E vs B0 | reading | |
|---|---|---|
| Deliverable quality | +0.0 pp ± 0.0 | pre-registered gate: PASS (but see below) |
| Task accuracy | −0.5 pp ± 1.1 | no regression beyond the 5 pp tolerance |
| Cost | −16.5% ± 3.2 | cheaper in all ten pairs (−10.8% to −26.1%) |
| Context tokens | −32.1% ± 2.2 | the effect the encoding targets |
The quality PASS is weaker than it looks. B0 scored 40 of 40 deliverables and 200 of 200 tasks,
so a regression had no room to show. The one wrong task in either arm was a B0-E sla, the task that
rebuilds zone ids from list position, which is where the encoding's risk sits:
What it does not show: anything about caching, a harder world, another model, or zones ≤ 4, where
the guard never fires by design. Decision record: vril-bench ADR-0012.
2026-11-02 · v0-alpha baseline: V0c closed loop vs B0 (GLM-5.3-Flash, 10 seeds). Not a Fructigen result
These figures measure the workflow's v0 control plane, before Fructigen existed. They are the numbers Fructigen has to improve on, and the reason it exists. V0c is the v0 closed loop: admission control decides whether a subagent starts at all, and agents pass compact manifests instead of transcripts.
Ten clean seeds on GLM-5.3-Flash, 692 calls:
| V0c vs B0 | reading | |
|---|---|---|
| Deliverable quality | +2.5 pp ± 5.7 | no regression; the interval spans zero |
| Task accuracy | +4.1 pp ± 7.4 | no regression; the interval spans zero |
| Cost | −23.3% ± 13.4 | cheaper, interval clear of zero |
| Wall time | −23.3% ± 15.0 | faster, interval clear of zero |
| Total tokens | −24.2% ± 10.9 | fewer, interval clear of zero |
| Context tokens | −43.9% ± 11.3 | the largest and most reproducible effect |
A cost and latency win at parity, not a quality win.
- On a stronger model the saving disappears. On DeepSeek v4.1 Flash, where B0 already scores 100%, V0c's cost difference is +1.1% ± 28.2 over 5 seeds. The context cut survives (−48.6% ± 5.0).
- With a silently corrupted lane, the cheap probe accepts wrong work. V0c accepts 18 wrong attempts; the judge variant V0 holds 7 of them and costs about +10.6% for it.
The one effect that reproduces on every model and world is the context cut, 29 to 51%. The largest unclaimed cost is the verifier: the judge that catches corrupted lanes re-reads each worker's whole prompt at full price. Fructigen targets both: pay for shared context once instead of once per sibling, and let the verifier read the worker's prompt from cache.
Figure sources and checksums: 2026-11-12-encode,
v0-alpha.
| Module | What it does |
|---|---|
segment |
Typed prompt segments, deterministic stable-first rendering, and prefix identity: a scope digest over everything a provider invalidates on, and a chained hash path |
graph |
Validated task graph: nesting-aware topological rank, fan-out |
encode |
Cache-safe encoders. One so far: zone-factoring of a dense region map |
provider |
Capability descriptors per backend, and a deterministic cache model |
plan |
The compiler |
render |
A planned call as request content, with breakpoints where the provider supports them |
attest |
Hash-path check before a call; realisation (reported over planned cache read) after it, with the cause of a shortfall named |
uv venv .venv && uv pip install --python .venv -e ".[dev]"
.venv/Scripts/python -m pytest # Windows; .venv/bin/python elsewhereStandard library only. This package must never import its consumers.
src/fructigen/ the package
tests/ its tests
versions/v0-alpha/ verbatim backup of the modules as they stood in vril-bench before the move
graph adopts the edge and validation design of google/adk-js (Apache-2.0). The scope digest in
segment adopts the continuity-check idea of thu-nics/AKV (Apache-2.0). Both are independent
implementations; no code was copied from either. "Cache-safe" is the term used by
hystax/optscale-kompressor (MIT).
Fructigen is part of the VRIL LABS Whitepaper Series on the Parallel Subtask Agent Workflow. Its name comes from Volume IV, where the fructigen field is the task-context embedding: the seed kernel a task grows from. A shared prompt trunk is that kernel, shared across lanes.
Series page: Parallel Subtask Agent Workflow, vril.li/research.
| Paper | |
|---|---|
| Preface | CSSF |
| Contents | Whitepaper Series |
| Volume I | QSC Pre-Flight Gate |
| Volume II | Dynagen Manifest |
| Volume III | Repulsator Broadcast |
| Volume IV | Fructigen Seed Corpus |
| Volume V | Swarm Thermodynamics |
| Glossary | Key Terms and Definitions |
Start with CONTRIBUTING.md and the quality bar in
CONSTRAINTS.md. Coding agents also read AGENTS.md; a summary for
language models is in llms.txt.
- Bugs, features and prediction mismatches: issues
- Security problems:
SECURITY.md, never a public issue - Getting help:
SUPPORT.md - What changed:
CHANGELOG.md
This project follows the Contributor Covenant.
VRIL LABS Open Source License v1.0. See LICENSE.
Built by VRIL-LABS · Ancient Knowledge · Future Technology