[Amendment] Daml Code Assistant - Deliver Milestone 2 and 3 Tool Tranches after Milestone 4
Development Fund Proposal Amendment
Proposal file: proposals/2026-06-IEU-Daml Code Assistant.md
SIG label: daml-tooling
Amends: Daml Code Assistant, PR #10, as updated in PR #687
Champion: IntellectEU Proposal Author: Nathan Cornille, IntellectEU (@nathancornille-ieu)
Motivation for Postponing M2 and M3 Tool Tranches
We propose completing M4 – Federated Learning before M2 and M3 Tool Tranches, while claiming the M2 Benchmark Tranche in line with the original proposal date. We are not proposing a change in the payout amounts for any of the milestones.
The main constraint on the M2 and M3 models today is not technology or infrastructure, but the amount of high-quality DAML training data available.
- Our experiments show that model performance continues to improve as more training data is added, with no clear plateau yet.
- M4 directly addresses this bottleneck by enabling us to collect more diverse training data in a privacy-safe way.
- This additional data should improve both the quality of generated DAML code and the model’s ability to generalize to a wider range of real-world use cases.
- Much of the required training infrastructure is already in place, meaning M2 and M3 can be developed more efficiently once sufficient data is available.
- We are already able to improve the performance of less capable local agents with our current training data.
See the build-rate performance of our model as training data increases. The horizontal line shows GPT-5.6 Sol performance with High effort, the dashed line is a simple linear extrapolation.
<img width="1000" height="640" alt="build-rate-vs-training-data" src="https://github.com/user-attachments/assets/c6372aff-025e-4169-a196-97b8f42b0c36" />Reordering the milestones therefore gives us a clearer path to delivering stronger M2 and M3 results while making better use of the work completed under M4. For a more in-depth explanation, including our experimental results and supporting literature, please check the Appendix.
Appendix - Daml Code Assistant amendment.pdf
Checklist
- Proposal file amended under
/proposals/ - Rationale documented
- Funding amounts and acceptance criteria unchanged