ADR-0063: Relax ADR alternative thresholds to stop fabricated options

Status: accepted | Date: 2026-09-16

Tags: validation

References: ADR-0042, ADR-0027

Context

External user feedback (September 2026) reports that current models produce bloated ADRs of 100-200 lines: three or four alternatives with two or three pros and cons each, even when only one viable option exists. The structural root cause is the gate threshold established in ADR-0042: writing decision requires at least two alternatives with at least one rejected, and adr accept requires at least one accepted and one rejected alternative.

For genuinely uncontested decisions — one viable direction, no real second option — that threshold cannot be met honestly. Agents satisfy it by fabricating straw options and boilerplate trade-offs, which is worse than no alternatives: fabricated deliberation is noise for human reviewers and misrepresents the decision record. The --force escape hatch becomes routine rather than exceptional, which erodes the gate habit ADR-0042 was designed to build.

ADR-0042’s core protection remains worth keeping: the write-time gate is what prevents conclusion-first writing, and it costs nothing when a real choice exists. The defect is only the rejected-alternative quota.

Decision

Relax the alternative thresholds established by ADR-0042 while keeping its two-gate structure. Writing decision requires at least one alternative. Accepting an ADR requires at least one accepted alternative, and — only when two or more alternatives exist — at least one rejected alternative, so the losing option of a real comparison is always marked. The --force escape hatch for historical backfills is unchanged. ADR-0042 is superseded by this decision.

Consequences

Positive: uncontested decisions produce short, honest ADRs again; the fabrication pressure that caused the reported bloat is removed at its structural root, complementing the softened adr-writer and adr-reviewer wording; gate bypass stays exceptional.

Negative: uncontested ADRs no longer demonstrate deliberation structurally. Mitigation: the adr-writer skill requires stating the uncontested nature plainly, and the reviewer flags fabricated-looking padding instead.

Neutral: existing accepted ADRs are unaffected; the gates apply only to future writes and acceptances, as before.

Alternatives Considered

Keep both gates but require only one accepted alternative; require a rejected alternative only when two or more alternatives exist (accepted)

• Pros: Removes the fabrication driver: uncontested decisions pass honestly with only the chosen option recorded, Keeps the write-time gate, the only structural barrier against conclusion-first writing
• Cons: Uncontested ADRs state rather than demonstrate deliberation

Remove both gates entirely and rely on skill guidance (rejected)

• Pros: No enforcement surface to maintain
• Cons: Conclusion-first writing returns; nothing prevents assertion-as-decision
• Rejected because: The write-time gate costs nothing when a real choice exists and is the only structural protection ADR-0042 provided that remains worth keeping.

Keep the gates unchanged and rely on softened skill wording (rejected)

• Pros: No code change
• Cons: The accept gate is a hard refusal, so agents still fabricate a rejected option to pass it
• Rejected because: Soft guidance cannot counteract a hard gate that requires content which may not exist.