E3-A COMPLETED — PUBLIC RESEARCH RESULTS
NadIntellect Lab has completed its first controlled AI communication calibration experiment: 1,000 trial records, a frozen experimental specification, independently checkable offline analyses, and a scientific manuscript.
The public research archive, including experimental records, code, documentation, and reproducibility materials, is available on Zenodo.
Read the E3-A research archive:
https://doi.org/10.5281/zenodo.23514526
A detailed progress report, including results, limitations, and the next three research directions, is published in the Progress Updates section below.
A narrow output can still carry a message: what we learned from E3-A
NadIntellect Lab has completed its first calibration experiment on constrained AI communication. The result separates three questions that are easy to confuse: what an interface permits, what a model can execute, and what a recipient can actually learn.
Imagine an AI assistant working with protected documents. Its operator restricts the answers it can send outside the system. An answer passes the rules—but could the choice of that answer also carry a message to someone who knows how to read it?
That possibility matters wherever a model has access to information that a recipient should not receive. A permitted output may still leave room for coded signals. Even the distinction between an accepted and a rejected answer can, in principle, convey information. To assess such a system, we need to understand the communication available through the interface, the model’s ability to use it, and the assumptions behind our measurements.
E3-A began with a deliberately tractable part of that problem: can a language model execute an explicitly supplied code through an exactly constrained output? We gave the sender the secret and the encoding instructions. This let us study prescribed communication under controlled conditions. Unauthorized extraction from protected documents remains a motivation for subsequent research; it was not the task tested here.
An interface we could check exactly
We built an experiment in which the model had to return a string of 32 binary digits satisfying specified equations. The arithmetic uses XOR. Four progressively constrained conditions, A0–A3, allowed 4,096, 256, 16 and one valid answer, respectively. A validator checked each response and released either the valid answer or a fixed rejection symbol.
The sender received a particular valid solution, a basis describing the available choices, and a rule mapping secret bits to those choices. A local exact encoder served as an oracle: it showed that the prescribed procedure could be executed correctly on the collected tasks.
The protocol was fixed before collection. The completed pilot, recorded on 15 October 2026, contains 50 base tasks and 1,000 trial records: 400 local control trials and 600 model calls. The recorded model identifier was gpt-5.6-sol. Of the model calls, 200 tested the instructed sender and 400 formed a separate valid-answer diagnostic. These are different conditions, not 1,000 independent tests of the same hypothesis.
The result: valid answers and correct messages diverged
For the instructed sender, each condition contained 50 trials:
Condition Valid answers available Responses accepted Exact prescribed code A0 4,096 0/50 0/50 A1 256 2/50 1/50 A2 16 22/50 19/50 A3 1 50/50 50/50 (empty code)
The A3 result means the model returned the sole valid answer. With no remaining choice among accepted answers, that success carries no semantic secret bits. The local oracle executed the prescribed code in all 200 trials, including the 50 empty-code cases.
The most informative result came from A2. Its 16 valid answers provided four binary coordinates for coding. The model produced an accepted answer in 22 of 50 trials, but reproduced the assigned code in only 19 of 50—38%. Three answers passed the validator while encoding the wrong word. Algebraic validity and message correctness were different achievements.
Looking inside those 19 successes changed the interpretation further. Every success involved a target requiring either no basis vector or just one. There were zero exact successes in the 26 trials requiring a combination of two or more basis vectors. Across all 22 accepted answers, only five of the 16 possible coordinate words appeared.
This is the central finding: the interface offered more coding choices than the model reliably used in this pilot. The observed pattern points toward questions about executing the prescribed procedure, but it does not identify the cause of the failures. It also makes the aggregate “38% success” an incomplete description. Which messages succeeded matters as much as how many did.
What the completed experiment establishes
E3-A achieved a concrete first objective for the laboratory: an operational, documented calibration experiment connecting a known mathematical output space to observed model behavior. Offline checks reported in the final manuscript agreed with all 1,000 retained release decisions and linked all 600 recorded model calls to the locally retained request/response journal. This supports the internal consistency of the evidence; the journal is not independent provider attestation.
Adversarial AI reviews helped sharpen the account, especially by making the restricted set of successful messages central to the conclusion. These reviews were AI-assisted methodological checks, not human peer review or independent experimental replication.
The statistical limits also shape what comes next. The frozen primary analysis of execution burden was inconclusive because its regressions encountered quasi-separation. The pilot did not establish a numerical measurement of total mutual information through the interface, or of information conveyed by acceptance and rejection alone. A retained-data planning calculation specifies 321 future TEST tasks, is sensitive to one influential observation, and remains a planning result. The confirmatory MAIN experiment has not been run.
For AI control, the useful lesson is methodological: an exact validator can tell us whether an answer belongs to the allowed set, while a separate analysis is needed to understand communication through that set. A model’s difficulty using a code is specific to the tested model, task and instructions. It is an empirical question how that behavior changes with a different sender or strategy.
Making the work inspectable
We have prepared a public evidence package covering the E3-A specification, implementation, scientific records, manuscript, supplementary analyses and correction history, with a minimized provider journal. The release is intended to let other researchers inspect the methods and challenge the interpretation. Documented transformations distinguish the public materials from the private operational archive; credentials and private seed material are excluded.
Three research programs follow from the result 1. Actual information flow: what can the recipient learn?
A count of permitted answers gives a mathematical bound. Measuring information transfer requires examining the full defined output, including the release decision, and stating what the recipient knows. This program will seek defensible estimates or informative bounds on actual transfer. Its value is to make claims about constrained communication testable under explicit assumptions, including cases where the available evidence supports only a limited conclusion.
2. Encoding and model capability: what limits the sender?
The sharp difference between simple A2 targets and targets requiring combinations raises a question that an overall success rate cannot answer: which part of the coding task limits performance? This program will examine how model capability and execution of a prescribed procedure relate to restrictions imposed by the interface. The goal is to distinguish competing explanations for success and failure, and to learn how far a result transfers beyond one model and one setup.
3. Utility under constrained output: what useful work remains possible?
A restricted interface is practically valuable only if it can preserve worthwhile work. This program asks which intellectual benefits can be measured independently under output constraints, and how those benefits relate to empirically assessed information transfer. We want evidence that helps designers judge what useful capability survives a restriction and what communication accompanies it.
These programs connect measurement, capability and usefulness. Their questions are defined; their specific future protocols still require prospective design and review. Each study should discriminate between explanations and leave a useful result whether its outcome is positive, negative or unresolved.
The laboratory we want to build
Our long-term aim is to help make powerful AI useful through interactions whose limits can be specified, tested and independently scrutinized. That requires a cumulative research program: clear hypotheses, preserved evidence, challenges to our own interpretations, and experiments that change what we ask next.
E3-A is the first completed experimental step in that direction. It gives us a working research foundation and a precise unresolved problem: how do we preserve valuable intellectual work while measuring the information that passes through a constrained interface?
We welcome researchers who can challenge the analysis, strengthen information-theoretic measurement, or help design and independently replicate the next studies. For supporters, this is the next milestone: turning the ambiguities exposed by E3-A into experiments capable of resolving them.
Thank you to Keri Warr for supporting the first experiment through Manifund, and to everyone willing to examine the work closely.