Evidence: harness probe 2026-07-27 on the glm-4.7-flash arm (RS-LoCoMo-Full-v4, conv-26, 168 claims), report archived at /opt/remember-stack/.benchmark-runs/v4-glm2/harness-probe-grok-45.md. Grok-4.5 acting as the retrieval agent over the public recipe API only.
- qa/0000: the store contains "went to the LGBTQ support group yesterday" — reachable via
claims_verbatim/explain — but no absolute date anywhere; claim_valid_from null. Gold "7 May 2023" unreachable in principle.
- qa/0001: "Melanie painted a lake sunrise last year" reachable; the year 2022 appears in no claim, span, or field.
- Session D1's own timestamp was never extracted as a claim, while D5/D6 timestamps were — coverage of session-header time is accidental, not systematic.
D41 gave claims valid_from/valid_until, and the Claimify prompt instructs models to use conversation timestamps, but on this arm only 13/168 claims got any valid-time and none of the gold ones. The miss is at extraction: relative expressions ("yesterday", "last year") should be resolved against the session timestamp that is present in the source header at extraction time — that context exists in the E2 input window.
Distinct from #153 (propagating already-stamped valid-time into fact layer / timeline): this issue is that the stamp is almost never produced for relative expressions in the first place. Both are needed before any cat-2 (WHEN) LoCoMo question can be answered from claims.
Acceptance: on the smoke corpus, the two gold WHEN facts (support-group visit, sunrise painting) carry grounded claim_valid_from (day and year precision respectively) after ingest, on at least the pinned extraction binding.
Evidence: harness probe 2026-07-27 on the glm-4.7-flash arm (RS-LoCoMo-Full-v4, conv-26, 168 claims), report archived at
/opt/remember-stack/.benchmark-runs/v4-glm2/harness-probe-grok-45.md. Grok-4.5 acting as the retrieval agent over the public recipe API only.claims_verbatim/explain— but no absolute date anywhere;claim_valid_fromnull. Gold "7 May 2023" unreachable in principle.D41 gave claims
valid_from/valid_until, and the Claimify prompt instructs models to use conversation timestamps, but on this arm only 13/168 claims got any valid-time and none of the gold ones. The miss is at extraction: relative expressions ("yesterday", "last year") should be resolved against the session timestamp that is present in the source header at extraction time — that context exists in the E2 input window.Distinct from #153 (propagating already-stamped valid-time into fact layer / timeline): this issue is that the stamp is almost never produced for relative expressions in the first place. Both are needed before any cat-2 (WHEN) LoCoMo question can be answered from claims.
Acceptance: on the smoke corpus, the two gold WHEN facts (support-group visit, sunrise painting) carry grounded
claim_valid_from(day and year precision respectively) after ingest, on at least the pinned extraction binding.