hills

Local verification environments for AI research agents

Quickstart · What a hill is · How it stays honest · Reference · Spec · Agent skill

When an agent runs research experiments autonomously, the same model writes the solution, evaluates it, and reports the result. It grades its own homework, so its numbers cannot be trusted. Hills separates the two roles.

A hill is a packaged evaluation task: a task description, an evaluator, and data. The agent (the climber) develops however it likes, but an official score comes only from hills eval, which runs the hill's evaluator in a separate process against a frozen, committed version of the hill and returns a signed report. The agent can query the verifier; it cannot modify it or grade itself.

The mental model is git. A hill is a versioned directory, hills commit freezes it, and every score is tied to the exact version that produced it.

Quickstart

Two minutes, no GPU, no network after install. We will create the hello-world hill, freeze it, score a submission against it, and then show what happens when someone edits the score.

1. Install.

uv tool install hills

If you want your coding agent to drive this instead, install the skill and it bootstraps the CLI itself:

npx skills add autolab-hq/hills

2. Create a hill. circle-packing ships with the tool: place 26 circles in the unit square without overlaps, maximize the sum of the radii.

$ mkdir demo && cd demo && git init -q .
$ hills new circle-packing -t circle-packing
hills: created machine state at ~/.autolab/hills
created demo/.hills/circle-packing
  template   circle-packing
  version control  demo/.hills/circle-packing/.vc (empty; nothing committed yet)

Nothing was asked of you and no file of yours was edited. .hills/ carries a .gitignore containing *, so it excludes itself from your project's git the way uv excludes .venv.

3. Check it, then freeze it.

$ hills check circle-packing
  ok    manifest             circle-packing 0.1.0
  ok    layout               eval.py, README.md, pyproject.toml
  ok    dependencies         uv.lock is up to date
  ok    evaluator contract   eval(submission, final, n, tolerance) imports and binds
  ok    tests                7 passed in 1.27s

circle-packing: all checks passed

$ hills commit circle-packing -m "initial"
  private.lock  0 file(s), 0 bytes
  blobs.lock    0 file(s), 0 bytes

committed circle-packing 0.1.0
  tree hash  0a93cd12b4360ba5434b7524897c8d4ce4ba5c68
  commit     20c61d03ca7245791b756238d8a7fa7626b08199

Scores from here on are tied to this tree hash. A new commit starts a fresh history.

commit runs check as a gate, then regenerates the lock files from disk. The tree hash is the hill's identity, and every score from here on is tied to it.

4. Score a submission. A submission is just a directory. The hill ships one:

$ cp -r .hills/circle-packing/examples/grid ./my-packing
$ hills eval ./my-packing -H circle-packing -o report.json

circle-packing @ 0a93cd12b436
  submission   ./my-packing
  hash         sha256:b415d43098bf…
  params       n=26  tolerance=1e-09
  PASSED       sum_radii=2.5414 (max)
  config       n=26*  mode=validation*  tolerance=1e-09   (* = primary)

The full report goes to stdout as JSON; the summary above is stderr, so hills eval ... > report.json does what you expect.

{
  "hill": "circle-packing",
  "tree_hash": "0a93cd12b4360ba5434b7524897c8d4ce4ba5c68",
  "commit": "20c61d03ca7245791b756238d8a7fa7626b08199",
  "submission_hash": "sha256:b415d43098bf9d11f8a0b910760d28ba13f68ca0ee54ba2288a085758ecebf08",
  "submission_git": null,
  "passed": true,
  "config": [
    {"name": "n",         "value": 26,           "primary": true},
    {"name": "mode",      "value": "validation", "primary": true},
    {"name": "tolerance", "value": 1e-09,        "primary": false}
  ],
  "metrics": [
    {"name": "sum_radii", "value": 2.5414, "direction": "max"}
  ],
  "details": {"min_radius": 0.0414, "max_radius": 0.1},
  "params": {"n": 26, "tolerance": 1e-09},
  "final": false,
  "official": true,
  "official_reason": null,
  "tool": {"version": "0.1.0", "sha256": "463ea637f2f79fb9…"},
  "timestamp": "2026-08-10T02:11:05Z",
  "report_version": 1,
  "signature": "hmac-sha256:97553de336afdd8a3088b27392cfec0c5fee9d28c134f315294fe48599f5f3d2"
}

5. Now try to improve the score by editing it.

$ hills verify report.json
signature valid: circle-packing @ 0a93cd12b436
  PASSED  sum_radii=2.5414 (max)
  signed 2026-08-10T02:11:05Z by hills 0.1.0

$ sed -i '' 's/2.5414/2.9/' report.json
$ hills verify report.json
signature INVALID: this report was edited, or it was signed on another machine.

That is the whole point of the tool, in one command.

Where to go next. hills describe circle-packing prints the contract your agent would read. hills new <name> scaffolds a blank hill for your own task, and hills new <name> -t nanogpt-10min starts from the flagship example: a timed training run scored on a held-out split the climber never sees.

What a hill is

A directory, versioned by its own embedded git repository:

.hills/circle-packing/
  hill.yaml        minimal manifest: watchdog, typed params, blob rules
  README.md        the contract, written for the climbing agent
  eval.py          THE entrypoint: def eval(submission: Path, **params) -> dict
  private/         evaluator-only content; never enters git
  examples/        minimal demonstration submissions
  tests/           checks on the hill itself, run by `hills check`
  .vc/             the hill's own git dir, named so it cannot collide with yours

private/ is the only special directory. Everything else is climber-readable by design, including eval.py: transparency about how you are judged is a feature. The consequence is that anything answer-revealing has to live in private/, not inline in the evaluator.

The evaluator contract

One fixed function, at the hill root:

from pathlib import Path

def eval(submission: Path, *, final: bool = False, **params) -> dict:
    return {
        "passed": True,
        "metrics": [{"name": "val_bpb", "value": 1.043, "direction": "min"}],
        "config":  [{"name": "gpu", "value": "rtx4090-24gb", "primary": True},
                    {"name": "torch", "value": "2.9.1", "primary": False}],
        "details": {},
    }

A submission is a directory. That is the whole input contract: a codebase, model weights, or a single JSON file are all just files in a directory.

The tool never imports eval.py in-process. It runs a shim as a subprocess in the hill's own uv environment, so per-hill dependencies stay isolated, a watchdog can kill a hung evaluation, and an evaluator crash cannot take down the tool. For timed tasks the evaluator launches the submitted code itself and enforces the deadline, so the climber's code never runs the official clock.

config entries describe the conditions of the measurement. Primary entries define comparability: two reports are comparable only if their primary config tuples match. Metrics are an ordered list with per-metric direction, and ranking is lexicographic in that order. Any set of reports therefore sorts into ranked groups with no configuration beyond the reports themselves.

How it stays honest

Identity is content, not history

A hill's identity is its git tree hash, not its commit hash. Tree hashes are a pure function of content, so the same files give the same identity on any machine. All state is keyed by tree hash: a new hill version starts a fresh attempts history, because a changed evaluator is a new game.

Locks bind what git must not hold

Two lock files, regenerated from disk at every commit:

The tree hash commits to private content through the lock without containing it. There is no blob store, no symlink farm, no content-addressed cache: the lock file is the tracking, and integrity is enforced at the two moments it matters, at commit (locks regenerated from disk) and at eval (disk verified against the locks at HEAD; a mismatch is a hard error naming the file).

Reports are signed, history is chained

The dict your evaluator returns is the report core. The tool wraps it in an envelope: hill name, tree hash, commit, submission hash, submission git lineage, params, tool version and package hash, timestamp, and an HMAC signature over the canonicalized report. The key lives at ~/.autolab/hills/key, mode 0600, deliberately outside any project directory an agent works in.

Every eval appends to an attempts log whose entries are HMAC-chained to the one before, so a deleted or edited line is detectable. hills attempts prints the break rather than hiding it.

Trust posture

This tool defends against self-deception: an agent loop accidentally or opportunistically grading its own work, editing its own scores, or drifting its own evaluation criteria.

It does not defend against a malicious human operator. Private files are ordinary files, protected by convention and by the tool only handing their location to the evaluator process. Signed reports are tamper-evident, not tamper-proof.

The honest claim: your agent cannot fake a hills report; you could, but then you're only lying to yourself. Disputed results are re-runnable, because hills are inspectable and submissions are hashed, so verification is ultimately by replay, not by trust in any single machine.

Working with a coding agent

The agent skill ships in this repo, version-locked to the CLI, and installs with npx skills add autolab-hq/hills or hills setup. You do not have to know what a hill is to use it: it triggers on any request to improve a number by iterating, and builds the hill as part of the job.

It runs in four phases.

1. Confirm the project. A minute, no more. What is this, and is it what you want to optimize?
2. Agree on a plan. Goal and direction, files in scope, files that are read-only, what is held out, the constraint that makes runs comparable, what would count as cheating, the run command, and when to stop. You confirm or edit it before anything is built.
3. Build and freeze the hill. The plan becomes a hill: read-only files are frozen into it, held-out data moves into private/, and the scoring code is copied rather than imported so it cannot drift with your project. The agent red-teams its own draft, then presents a brief listing the gaming vectors it closed and the ones that remain open as your decision. You run hills commit. The agent that wrote the evaluator does not get to freeze it.
4. The experiment loop. A fresh subagent starts from hills describe and nothing else, then loops: edit, commit, dev-run, hills eval, decide. It does not stop to ask permission, and it runs until your stopping criteria are met.

The third phase is the one a self-graded loop skips, and it is the reason the numbers at the end mean something. Total human surface: one install, one commit, one "proceed" with stopping criteria.

Reference

Commands

hills eval flags: -p key=value (repeatable), --final for test mode, --force, --current, --queue, -v to stream evaluator output, -o to also write the report to a file.

It evaluates HEAD, never the working tree. A dirty hill is an error: commit it, --force to score the last committed version anyway, or --current to test a draft evaluator against a real submission (unofficial, tree_hash: null, logged separately).

Machine state

~/.autolab/hills/
  registry.json                       name -> path, current tree hash
  key                                 the per-machine signing key (0600)
  state/<name>@<tree_hash>/attempts.jsonl   append-only, HMAC-chained
  runs/<name>/<timestamp>-<id>/       materialized hill, submission snapshot, logs, report
  envs/<name>/<tree_hash>/            the uv environment for that hill version

Hills are stateless: a hill emits signed reports and remembers nothing. Eval history is tool state; the climbing agent keeps its own working notes.

Repository layout

src/hills/          the library and CLI
skills/hills/       the agent skill, version-locked to the CLI
examples/           the example hills, also usable as `hills new -t <name>`
docs/SPEC.md        the design specification this implements
tests/              tests for the tool

The two shipped hills: circle-packing, the hello-world used above, exact arithmetic and no private data; and nanogpt-10min, a timed training run whose evaluator owns the clock, keeps its splits in private/, and reports a normalized GPU profile as primary config.

License

MIT. See LICENSE.