Datafew Talk to us

AI write-path control

Control the change. Prove the result. Reuse the work.

Control the AI write path.

Watch a changeSee the evidence

The concrete risk

Keep the scope.

An edit still needs permission. Datafew makes scope visible.

✓

Authorized edit

retry.ts#backoff

Independent checks follow.

×

Locked region

retry.ts#retryPolicy

No write is admitted.

How it works

One controlled loop.

The model supplies inference. Only admitted writes can leave.

Model / LLM provider inference

Datafew control plane

Coding agent / IDE / internal bot
  1. 01Authorizebind edit scope
  2. 02Gatereject unsafe writes
  3. 03Verifyexecute independently
  4. 04Reuseadmit passed work only

Append-only audit throughout

Repository / CI / production admitted write only
Read the control semantics

The full agent loop stays inside. First, bind the allowed scope. Then, gate each proposed write. Verify by independent execution. Keep an append-only audit trail. Reuse only admitted work.

First-party evidence

Inspect the proof.

Real execution judges each run. The same model serves both arms.

Admitted passesn = 128 runnable tasks
Raw
41 / 128
Control plane
114 / 128
128 tasks. Same model.2.89× total tokens, with repair.Not production traffic.
Conditions and replication

73 passes only with Datafew. Zero passes only in the raw arm. Per-call cap: 16,384 tokens. Temperature: 0.1 in both arms. Total token budgets differ. Audited locally on a clean commit. Not an official submission.

Later: five fresh tasks. A different model served both arms. 4/5 admitted. 1/5 raw. 1.6× model calls with Datafew. A direction, not a general rate.

Verified reuse

Use the work again.

Reuse starts from verified work. New tasks may still need repair.

Same task · BigCodeBench/19

2First-run model calls
0Reuse model calls

Independent verification passes. Execution checks pass again.

One same-task observation.Not a cross-task result.No production savings measured.

New tasks · BigCodeBench-Hard

0 / 20passed as-is
20 / 20admitted after bounded repair
21model calls across 20 tasks

Every task found prior work. None passed without repair. 19 needed one repair round. One needed two rounds.

20 previously unseen tasks.21 model calls. No cold arm.No cost comparison.
Controlled warm-path study

Separate 20-task warm suite

Both arms use the same tasks.Only the warm path changes.

20 → 5model calls
36%of baseline cost per safe edit
19/20 → 20/20safety rate
20-task controlled suite.5 measured model calls, not zero.Not production traffic.

Fit and verification

Show us your setup.

We are preparing deep integrations. No customers are in production today.

Under NDAYour engineers can clone it.Run the pinned verification suite.Dataset hashes are pinned.

Deterministic checksUse your own environment.Model keys are not needed.Negative results are included.

Contact

Let’s talk.

Describe your current setup. We will say if Datafew is useful.

datafew / direct contact

  • bowen@datafew.com
  • yicai@datafew.com

Copy an address to get in touch.