How much data is in this batch?

33 independent yes/no questions, 429 runs including rewired-connectome, no-connectome and random-graph controls. Full per-run tables are published in the repository.

What controls does BioReservoir score against?

A no-connectome baseline (same embedding and readout wiring, no brain), a rewired connectome that preserves degrees, weights, signs and cell types, and an Erdős–Rényi random graph of the same size and density.

Did BioReservoir's validation pass?

Partly, and that is reported directly. The male connectome (MaleCNS) reproduces Shiu et al.'s published sugar-feeding response and shows a real, consistent left/right lateral bias. The female connectome (BANC) fails the same lateral-bias validation, collapsing to a constant response regardless of which side is stimulated -- reported as a failure, not hidden.

Are outcomes scored yet?

No. This batch's questions resolve by 2026-12-15 and no outcomes have been recorded yet, so no accuracy, Brier score or calibration numbers exist yet -- only the process numbers on this page.

Data

The data behind BioReservoir

This page is the landing page for anyone who wants the numbers, not the interactive experience — link here instead of the home page. Everything below is computed from BioReservoir's own ledger and committed calibration files, not asserted. Where the ledger does not have enough to support a claim, that is said explicitly instead of filled in.

Every run is a live event, not a cached reply: the simulation runs fresh on IG Digital Lab's own CPU servers in Sacramento, California, and takes about 66 seconds end to end. /api/now reports the current queue and cycle time in real time.

33independent yes/no questions

429runs including controls

99real-brain runs (33 × 3 phrasings)

330control runs

Volume

33 independent yes/no questions · 429 runs including controls. Verified directly against experiments/001-fly-oracle/ledger.sqlite: 99 runs on the real male connectome (33 questions, each asked as its original phrasing, its negation, and a paraphrase — 10 simulation trials per run) plus 330 control runs (see below). All questions are about public events; per the owner's decision, this page does not name individual questions or report which way any one of them leaned — only aggregate, batch-level numbers.

Controls

Every question is also scored against these controls, not just against nothing:

No-connectome baseline — the same text embedding and readout wiring, with no brain at all. Isolates how much of any answer comes from the embedding itself rather than the connectome.
Rewired connectome — the same neurons, degrees, synapse weights, signs and cell types, with connections shuffled. Isolates the real wiring pattern from everything else that is held constant.
Erdős–Rényi random graph — a random graph of the same size and density as the real connectome (auxiliary control).

No-connectome runs are deterministic given their seed (one trial each); real, rewired and random-graph runs are stochastic simulations (10 trials each). Full per-run counts for every condition, including controls not named above, are published in the repository (see below).

Invariance and confidence

Computed from the real-brain runs only (33 questions × 3 phrasings, malecns/real condition), comparing each question's original phrasing against its negation and its paraphrase:

Full per-run measurement tables are published in the repository: experiments/001-fly-oracle/RESULTS.md. Read plainly: rephrasing a question — negating it or paraphrasing it — moves the answer only slightly, for every condition tested. That is consistent with the rest of this site's framing: the signal is real and measurable (see Validation below), but it is small, and this batch does not show it producing a confidently one-sided answer to individual questions.

Not available yet: accuracy, Brier score and calibration against actual outcomes. This batch's questions resolve by 2026-12-15 and no outcomes have been recorded in the ledger's resolutions table yet — that scoring will only exist after resolution, and will be published here when it does, including if the brain performs no better than its controls.

Methodology

Encode. A small text-embedding model (all-MiniLM-L6-v2) turns the question's text into a vector, which becomes Poisson input rates on real sensory neuron populations: gustatory (sugar/bitter), olfactory, photoreceptor and mechanosensory, defined by each connectome's own cell-type vocabulary.
Simulate. A whole-brain leaky integrate-and-fire (LIF) network — parameters ported unchanged from Shiu et al. (2024) — runs for 250 simulated milliseconds per trial, 10 trials per (question, phrasing, condition).
Read out. The primary readout is the lateral bias of the descending-neuron population, (L−R)/(L+R) (~650 neurons per side) — not a single named cell pair, which turned out to be too small a sample to trust on its own.
Correct for handedness. Every brain has its own intrinsic left/right asymmetry, independent of any question. A set of 24 short, neutral reference sentences establishes that baseline bias per (brain, condition), and it is subtracted from every question's raw bias before scoring.
Map to yes/no. Each question gets its own pre-committed, randomly seeded assignment deciding whether a leftward corrected bias reads as "yes" or "no" for that question — set before any run, so a brain that happened to have a constant left/right lean could not read as a uniform yes-or-no skew across the batch.

Full parameter table, citations and calibration detail: live at /api/validation and on the validation section of the home page.

What validated, and what didn't

Reproduced Driving the real sugar-sensing neurons (the male connectome's own labellar taste population) fires the real proboscis-extension motor neuron, MN9, at 128.5 ± 21.3 Hz (mean ± std, 10 trials, 150 Hz input) against 0 spikes with no input — matching Shiu et al.'s published direction and shape of the response. Adding bitter input on top of sugar suppresses that firing by 96.8%, reproducing the paper's feed-forward inhibition result.

Failed — published anyway The same lateral-bias readout, run on the second, female connectome (BANC), does not track which side is stimulated at all: driving either the left or right side of a balanced sensory population produces the identical bias, collapsing to a constant response every trial. This is a real limit of BANC's current public reconstruction (a proofreading-driven left/right imbalance in the released data, not a code bug — verified by checking reachability and by trying a rebalanced subgraph, which did not fix it), not a result we are hiding. The male connectome (MaleCNS) does not have this problem: its lateral bias tracks stimulus side correctly and consistently. Live numbers for both: /api/validation.

Try it yourself

Ask the fly a question and watch the real spikes, or read the press kit for sources, images and contact. The pipeline described above — encoding, simulation, readout, handedness correction, scoring — is open source: github.com/igdigitallab/bioreservoir, Apache License 2.0. This page and the live experience are built and operated by IG Digital Lab.