RLHF and alignment
RLHF covers reward-model training and human-feedback alignment used to shape model behaviour. Because this work is typically iterative and its outcome uncertain until late, it usually reads as research under IAS 38 or is held below the probable-to-complete threshold by significant development uncertainty under ASU 2025-06, and is expensed.
Routing by framework
Reward-model training and human-feedback alignment are usually iterative and exploratory, so they read as research and are expensed unless a defined development-phase asset can be demonstrated.
Novel alignment work commonly carries significant development uncertainty, which defers capitalisation under the probable-to-complete threshold.
Why it usually expenses
Alignment work refines behaviour that cannot be fully specified in advance, so the entity often cannot demonstrate technical feasibility or a probable defined benefit while the work is ongoing IAS 38§54. Under US GAAP, novel alignment carries significant development uncertainty, which defers capitalisation until resolved ASU 2025-06.
Where a defined asset can emerge
A reward model built to a fixed specification for a scoped, feasible internal system can, once the recognition tests are met, produce capitalisable development-phase cost like any other component. The distinction is between open-ended behaviour shaping and building a specified, feasible artefact.
An entity runs several rounds of human feedback to improve a support assistant. The exploratory rounds establishing whether the approach works are expensed. Once the behaviour target is fixed and feasible and the build is funded, further directly attributable cost may be capitalised. All figures are illustrative.
- S1IAS 38 Intangible Assets, IFRS Foundation (IFRS). https://www.ifrs.org/issued-standards/list-of-standards/ias-38-intangible-assets/
- S2Handbook: Software and website costs (ASC 350-40 internal-use software), KPMG (US GAAP). https://kpmg.com/us/en/frv/reference-library/2026/handbook-software-website-costs.html
- S5ASU 2025-06 on internal-use software costs, BDO (US GAAP). https://arch.bdo.com/new-asu-on-internal-use-software-costs-guidance