Inspiration
GroundLoop started from a problem I kept running into as a researcher: the data can look meaningful before the claim is actually safe.
A curve changes. A peak appears. A device response shifts. The first reaction is to explain it with the mechanism we expected. But in real experimental work, the same signal can still be caused by contacts, geometry, drift, heating, calibration, sample history, or a missing control.
I did not want to build another tool that just finds more papers or gives a smoother AI explanation. I wanted a workspace that asks the harder question:
What does this experiment actually identify, and what should I measure next?
What it does
GroundLoop: ControlFirst is a local research decision workspace for materials, electronics, and functional-device experiments.
A researcher starts with a mechanism claim, method notes, bounded CSV measurement data, and literature candidates. Codex with GPT-5.6 does the scientific reasoning: it reads the claim, inspects the measurement context, reviews source excerpts, decomposes the claim into measurable signatures, compares those signatures against local data facts, and proposes the next control experiment.
GroundLoop is the evidence boundary around that reasoning. It keeps track of source provenance, data hashes, researcher-confirmed bindings, review status, frozen evidence packets, deterministic data facts, four alignment states, and the final control contract.
The result is not "AI says the claim is true." The result is a Convergence Map:
- Observed — the required signature is directly present in the data.
- Confounded — the data is compatible, but another explanation still survives.
- Missing — the required observable was not measured.
- Contradicted — the data goes against the required signature.
Then GroundLoop ends with one ControlFirst proposal: the smallest next experiment that would separate the proposed mechanism from the main confound.
How a judge can try it
The workflow is intentionally local and reproducible.
- Clone the repository and run the setup commands in the README.
- Register the GroundLoop MCP server with Codex.
- Start the local UI with ./scripts/demo.sh.
- Open the app, create or load a Run, and copy the Codex brief.
- Paste that brief into a Codex session using GPT-5.6.
- Codex reviews sources, stops at the human freeze gate, then continues through data facts, signatures, alignments, one control contract, and export.
The demo includes a generic CSV path and a deeper electrical transport fixture. The transport fixture is deliberately conservative: a two-wire resistance-temperature trace shows a real change, but GroundLoop refuses to call it a bulk mechanism until a matched four-terminal control is run. That is the point of the product: useful skepticism that ends in a concrete next step.
Why it is different
Most research AI tools are strongest at search, summary, or chat. GroundLoop is built for the moment after that: evidence-bound judgment.
Codex is allowed to reason broadly. It can bring in literature, think top-down from theory, think bottom-up from the dataset profile, and design controls. GroundLoop only constrains what can be committed as evidence.
A search snippet is not evidence. A title is not evidence. A real signal is not automatically mechanism proof. An Observed or Contradicted alignment must cite GroundLoop-materialized data evidence. A Confounded alignment must name the alternative explanation that still remains. A direct source must be semantically reviewed before it can support a role such as theory basis, method limit, or discriminating control.
That split is the product: keep the LLM useful, but make the final research decision auditable.
How we built it
GroundLoop is a Python and React application with a typed local evidence core, a loopback-only FastAPI service, and a stdio MCP server for Codex.
The core accepts bounded UTF-8 CSV measurements with arbitrary headers, profiles them conservatively, stores artifact hashes, requires explicit column bindings, materializes deterministic data facts, validates alignment records, and exports Markdown/JSON reports. The UI owns the Run, source review state, evidence freeze, Convergence Map, and report view. The model does not call the UI directly; Codex works through the MCP tools.
The implementation now supports a generic CSV workflow, imported literature candidates with provenance, source-role review, stale-state handling when evidence changes, and method-aware capability surfaces for deeper demos without turning recipes into rigid scientific rules.
How we used Codex and GPT-5.6
Codex was both the build partner and the runtime operator.
During development, Codex helped design and implement the Run state machine, MCP contracts, local API, React UI, deterministic evidence operations, source provenance rules, tests, and export path. It also drove repeated QA loops from the point of view of a first-time user and a hackathon judge: install, run locally, copy a brief into Codex, review sources, freeze evidence, generate a Convergence Map, and inspect the final report.
At runtime, GPT-5.6 is used through Codex as the reasoning host. GroundLoop does not require an OpenAI API key, and the local web UI does not make cloud model calls. GPT-5.6 performs the semantic work; GroundLoop validates what can be saved.
Challenges
The hardest part was avoiding two extremes.
If GroundLoop is too loose, it becomes another confident research chatbot. If it is too rigid, it suppresses the LLM's ability to reason across unfamiliar hypotheses and measurement types.
The current design keeps capability packs and recipes as advisory evidence surfaces, not scientific handcuffs. Codex can still reason. GroundLoop catches the places where reasoning becomes an unsupported claim.
The second challenge was making a conservative answer feel useful. "Not established yet" only helps if the system also shows why and tells the researcher what to do next. That is why the report is organized around claim, evidence, gap, and control.
What we are proud of
- A real local workflow that judges can run from the README.
- Codex/GPT-5.6 used as the reasoning surface through MCP, not as a hidden chatbot.
- A clear split between literature support, local data facts, and mechanism interpretation.
- Source candidates that stay untrusted until reviewed.
- Reports that preserve provenance, hashes, roles, rationales, and control outcomes.
- A product stance that helps a researcher make a better next experiment, not a louder claim.
What's next
The next step is more real-use stress testing: spectra, I-V curves, time series, grouped comparisons, hysteresis, and multi-artifact controls. I would expand deterministic evidence operations only where repeated use shows they help Codex avoid overclaiming.
Longer term, GroundLoop should become a practical workspace for research groups to ask: given our current theory, methods, literature, and data, what can we responsibly claim, and what control should we run next?
Log in or sign up for Devpost to join the conversation.