Your first study

Run a complete, inspectable research workflow on your own computer. In this example you will compare two estimators, keep a failed and interrupted attempt, inspect the resulting claims, and recompute the report from its evidence. There is no model session, API bill, GPU, or Python package installation.

Get the small source checkout

You need Python 3.11 or newer, Git, and macOS or Linux. Run python3 --version first. This sparse checkout leaves the large historical research archives out of your initial download:

git clone --depth 1 --filter=blob:none --sparse https://github.com/alohays/allagma.git
cd allagma
git sparse-checkout set allagma adapters contracts methods recipes policies profiles \
  templates tools conformance examples/toy-study examples/first-research evals/context-retention
python3 -m allagma check

The catalog command returns JSON with "status": "pass". The public checkout requires no GitHub authentication. If you already have a checkout, start with that last command. The Python runtime is independent of the documentation site’s Node toolchain. A wheel or pip install allagma is not the supported distribution in this release candidate.

Run the experiment

python3 -m allagma toy --destination work/my-first-study

Use a new destination for each run. An existing study is never erased. The command prints a JSON handoff when complete; it may be quiet while running. The example has a 60-second recorded execution budget and 28-attempt ceiling; those are bounds, not a runtime promise.

The question is deliberately small: does adding 0.25 to a sample mean increase its squared error? Each observation is equally likely to be −1 or +1. Two known-answer pilot seeds check the runner. Then 24 new seeds each generate 64 observations for the paired comparison.

One attempt really fails and another is interrupted after a checkpoint. Both are retained, and their retries receive new IDs. These are deliberate recovery demonstrations. They do not count as extra independent observations. See the full toy specification.

Read the result

Open this file in your editor or Markdown viewer:

work/my-first-study/campaigns/toy-v1/analyses/a001/paper/manuscript.md

The representative result is:

Quantity	Value
Sample-mean MSE	0.01529948
Mean-plus-0.25 MSE	0.08040365
Paired increase in MSE	0.06510417
Approximate 95% interval	[0.03985104, 0.09035729]
Independent confirmation seeds	24

These are generated results, not values typed into a report template. The interval is the seed-level mean ±1.96 standard errors, a normal approximation. The mathematical expected increase is 0.25² = 0.0625; the realized estimate need not equal that expectation.

The adjacent claims.json marks the average comparison supported. It marks “greater error on every seed” contradicted: seeds 104, 108, 118 and 121 are counterexamples. A third claim records leave-one-seed-out sensitivity. An average effect does not justify a claim about every observation.

Follow a claim back to evidence

Within campaigns/toy-v1/, inspect these files in order:

File or directory	What it lets you check
analyses/a001/paper/claims.json	Claim status, scope, limitations and references
analyses/a001/outputs/summary.json	Numerical values used in the report
analyses/a001/raw-manifest.json	Exactly which attempts were included or excluded
runs/	Individual attempts, raw measurements and failure records
materials/ and lock.yaml	Frozen study programs and the exact Allagma bundle

The figure is analyses/a001/outputs/paired-differences.svg. It shows individual seed differences rather than only the favorable average. Files are ordinary Markdown, JSON and SVG; no hosted dashboard is required to read them.

Verify the package

python3 -m allagma campaign audit --study work/my-first-study --campaign toy-v1

Look for "verdict": "pass", an empty findings list, and 521 checked references for the default full example. The audit checks content digests, reruns the analysis and regenerates the manuscript. It appends a new review; the original remains. This is deterministic verification, not independent scientific peer review or evidence of a language model’s research quality.

The campaign ID is toy-v1; its protocol revision is toy-v2. These identify different things and are intentionally distinct. The state separates phase (audit), execution (completed) and assurance (deterministic).

Continue from here

Turn these retained results into an additional paper with the runnable toy-to-paper walkthrough. It covers explicit attribution, provided-only references, scoped reviews and stale-review recovery. The base walkthrough stays offline; TeX compilation is a separate optional step.

The Linux/Python 3.13 qualification receipt records these exact commands on Debian 12/aarch64 with CPython 3.13.16, including a checkout path with spaces. It covers the offline toy and conformance suite in that container environment; native agents and model quality need separate evidence.

To change the offline question without a model session, follow prepare an adapted toy study. It stops before execution so you can inspect the changed bias, source lineage and finite plan, then start, run, analyze and audit the new campaign explicitly.

Use your own question with a native agent, learn to resume and reproduce, or browse the real study showcase. If a command fails, use troubleshooting before changing a lock or deleting evidence.

---

Canonical source · Generated from the maintained repository document at this build’s commit.