Unify the ADR-050 Revision 2 (FUSE-crash-forced-drop) and ADR-052 (methodology-load-bearing-with-crash-as-trigger) framings into a single account

Published

May 20, 2026

ADR-075: full-FT OOD drop rationale – unified narrative

Status

Accepted.

Context

ADR-050 was published in two revisions on 2026-05-18:

  • Revision 1: LLM-judge reference scorers (gpt-4o-2024-08-06 + claude-sonnet-4-6) dropped at Phase 4 cost re-estimation when the envelope landed ~16x the original ADR-018 estimate ($14 -> $240) driven by per-row LLM-judge inference being charged at the full input-prompt token count. The vendor_black_box contamination tier therefore carries 0 rungs in this submission; the contamination- stratification gradient compresses from 4 tiers to 3.
  • Revision 2: full-FT OOD inference dropped after the Phase 5 X11 FUSE EIO crash on /workspace MooseFS storage. The 24 LODO direct- source predictions from Phase 2 survived; the post-Phase-5 OOD inference required a re-fire on Low-stock A100 80GB which the operational state didn’t allow.

Hours later the same day, ADR-052 was published, reframing the full-FT drop as methodologically load-bearing (LoRA’s paired-bootstrap evidence already showed fine-tuning was hurting OOD; the FUSE crash was the proximate trigger that exposed an already-implicit decision, not the cause of it).

A skeptical reader reading ADR-050 R2 then ADR-052 sees: “I cut a planned experiment because it crashed -> wait, actually I cut it because the methodology pointed there -> so why did you write ADR-050 R2 with the operational framing?” Same-day retcons in immutable records erode reader trust in any ADR the author wrote, even when (as here) the methodological reasoning is genuinely sound and the operational trigger is genuinely separate.

ADR-050 Revision 1 (LLM-judge cost drop) is unaffected by this; that’s a separate decision on a different axis.

Decision

Replace the two-ADR same-day retcon with one unified account in ADR-075:

Full-FT OOD inference was not run because BOTH (a) the Phase 4 paired-bootstrap CI on LoRA already showed the fine-tuning direction was net-harmful for OOD generalization (LoRA -0.071 AUPRC vs frozen-probe on pooled_ood, CI clears zero; paired_cells.parquet seed=1 + seed=2 stability check 0/40 cells flagged), AND (b) the Phase 5 X11 re-fire attempt crashed on a FUSE EIO before any full-FT checkpoint persisted. The methodological reasoning is load-bearing; the operational event is the trigger that exposed the already-implicit decision. Both halves are real; neither would have independently sufficed for a defensible drop (without the LoRA evidence the FUSE crash would have required a re-fire attempt; without the FUSE crash the methodology argument alone might have deferred to v1.1.x rather than dropping at v1.0.x).

The unified narrative is the prospective citation. ADR-050 Revision 2 and ADR-052 are preserved as historical artifacts in decisions/ — the record of how the decision logic was authored in stages on 2026-05-18. WRITEUP.md §8.1 + WRITEUP/limitations-and-future-work.md §8.1 + WRITEUP/model-rungs.md §4.x cite ADR-075 as the single source of truth on the full-FT OOD drop rationale.

ADR-050 R1 (LLM-judge cost drop) is preserved cleanly — separate decision, separate ADR axis, untouched by this supersession.

Consequences

  • Reviewers reading the ADR sequence + the WRITEUP cross-references see one coherent rationale rather than two competing ones authored hours apart
  • The same-day retcon optic is removed without erasing the historical record (ADR-050 R2 + ADR-052 both remain readable)
  • ADR-050 R1 (LLM-judge cost drop) survives intact on its separate axis
  • WRITEUP cross-references simplify: one ADR citation (ADR-075) instead of two with a “see also” thread between them
  • Future references to the full-FT OOD drop in CHANGELOG, NEXT_STEPS, or other prose can cite ADR-075 directly rather than navigating the ADR-050 R2 -> ADR-052 history
  • ADR-072 (frontmatter backfill for ADR-051 + ADR-052) and ADR-075 both supersede ADR-052 on different axes — ADR-072 on the frontmatter axis (closing_commit + Alternatives Considered), ADR-075 on the narrative axis (unified full-FT drop rationale). Both supersessions are documented; ADR-052’s full historical content remains in decisions/ for audit-trail purposes

Alternatives Considered

  1. Leave ADR-050 R2 + ADR-052 as the two-ADR pair. Rejected — the retcon optic is a real reader-trust risk for any ADR the author wrote. The methodological reasoning in ADR-052 is sound; the narrative structure (two ADRs authored hours apart with different framings) is the issue. Unifying removes the structural problem without changing the methodology content.
  2. Edit ADR-050 in place to incorporate ADR-052’s framing. Rejected — would violate the immutability rule (this is content change, not editorial fix; doesn’t fit ADR-073’s named exception classes). The immutability discipline is the load-bearing constraint; the cost of working within it is the supersession chain.
  3. Delete ADR-052. Rejected — historical record (the decision- trail of the same-day reframe) is preserved by keeping both ADR-050 R2 and ADR-052 in decisions/ while citing ADR-075 prospectively. Deletion would lose the lesson about how same-day retcons happen + why they should be avoided.
  4. Write ADR-052 differently at the time. This isn’t an alternative to ADR-075 (it’s an alternative history); noting it here for completeness. The 2026-05-18 author’s choice to write a second ADR rather than revise the first followed the immutability discipline (correctly); ADR-075 is the deferred-consolidation move the discipline supports.

Linked ADRs

  • ADR-050 Revision 2 (full-FT OOD drop via FUSE EIO): superseded on the prospective-citation axis only; the historical record (operational facts) remains intact
  • ADR-050 Revision 1 (LLM-judge cost drop): NOT superseded; separate decision on a separate axis; remains in force
  • ADR-052 (methodological reframing of ADR-050 R2): superseded entirely in prospective citation; its sole purpose was the ADR-050 R2 reframe which ADR-075 now subsumes. Historical content remains for audit-trail purposes
  • ADR-019 (LoRA training recipe): provides the LoRA evidence that ADR-075 cites as the methodological half
  • ADR-022 (statistical apparatus): defines the paired-bootstrap CI methodology that produced LoRA’s -0.071 AUPRC delta vs frozen-probe
  • ADR-020 (cost discipline): names the GPU spend cap that constrained the re-fire decision
  • ADR-049 (GPU order): names the A100 80GB platform the X11 crash hit
  • ADR-072 (frontmatter backfill for ADR-051 + ADR-052): supersedes ADR-052 on the orthogonal frontmatter axis; this ADR’s narrative supersession + ADR-072’s frontmatter supersession are complementary, not conflicting

Transcript

Transcript file (gitignored per AGENTS.md): the 2026-05-20 audit governance-audit cycle’s identification of the ADR-050 -> ADR-052 same-day retcon as a contradiction class. Captured in ~/notes/prompt-injection-audit-2026-05-20.md (high-impact ADR finding #5) and ~/notes/prompt-injection-audit-2026-05-20-adr-appendix.md (drafted ADR-075).