GOLDEN PATH
HOMECHANGE STATEREAD RESULTWHY HYDRADBTRACE FCOMODELS USEDVERIFY CUSTODYEVIDENCE

Track 03 · Memory + Context Retrieval

HydraMemory

A real LongMemEval-S full500 run materialized temporal sessions, facts, entities, supersession and contradiction in pinned HydraDB, then compared A/B/C/D retrieval. A later pre-registered K5/K10 matrix tested retrieval depth as a controlled perturbation.

Judge walkthroughWhy HydraDB + mathInspect 4D FCGHow to use
FULL500 COMPLETENO POSITIVE B/C/D HIT-RATE SIGNALMODEL BENEFIT NOT ESTABLISHED
Cases500470 retrieval-scored · 30 abstentions
Sessions23,867occurrence-scoped identity
Facts3,506heuristic extraction in historical full500 lane
Entities4,776context-scoped
01 / K=5 RETRIEVAL RESULT

The graph worked. The tested K=5 retrieval advantage did not appear.

B, C and D did not establish a positive Hit@5 advantage over route A at the tested K=5 configuration. Evidence-path coverage can increase while top-k recall declines; HydraDG retains that null/negative result.

RouteHit@5Recall@5Interpretation
A · reference / flat0.96382978720.9065957447reference
B · temporal0.94680851060.8538297872no positive Hit@5 signal
C · temporal + entity/provenance0.94680851060.8525886500no positive Hit@5 signal
D · current/supersession + contradiction0.94468085000.8460283700no positive Hit@5 signal

Evidence class

RECOMPUTED_LIVE_HYDRADB_RETRIEVAL_ABLATION

Claim ceiling: LONGMEMEVAL_FULL500_RETRIEVAL_ABLATION_ONLY_NOT_END_TO_END_QA

Custody state

Hashed, not signed.

Signature: NOT_SIGNED · Merkle: NOT_MERKLE_COMMITTED · independent replication: NOT_ESTABLISHED

02 / K=5 → K=10 DEPTH ABLATION

K=10 improved retrieval depth without establishing a representation or model win.

The pre-registered 2×2 matrix held the frozen dataset and retrieval logic fixed while comparing RAW vs SeedGraph at K=5 and K=10. RAW and SeedGraph were identical at the same K; increasing K improved Hit@K and Recall@K while evidence-path coverage density declined.

DepthMethodHit@KRecall@KEvidence-path coverage
K=5Method A0.963830.906600.63787
K=5Method D0.944680.846030.63787
K=10Method A0.978720.945350.51511
K=10Method D0.970210.922730.51511

Method D depth effect

+2.553 pp Hit · +7.670 pp Recall

K10 minus K5 in the retained local matrix. This supports a ranking-depth/cutoff effect under the tested implementation.

Model involvement

NONE IN PRIMARY MATRIX

The K10 result is not evidence that an LLM improved retrieval. Model-assisted extraction remains a separate future controlled axis.

Open Best Use of HydraDB analysisPre-registrationRetained summary
03 / GRAPH

What was actually written.

RelationCountRole
ABOUT7,012typed graph structure
ASSERTS3,506typed graph structure
CONTAINS23,867typed graph structure
CONTRADICTS4,914incompatible fact state
DERIVED_FROM3,506typed graph structure
HAS_CASE500typed graph structure
MENTIONS6,935typed graph structure
NEXT23,367typed graph structure
PREV23,367typed graph structure
SUPERSEDED_BY2,457temporal replacement
04 / HASHES

Executed result identities.

LongMemEval source

d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442

Result

bdecb4b62cf90040c7f346d283efe78459825b427557cec8d4998f3499ee0324

Statistics

8dcf57f5ac60418d16d3c945ad678b4d17b557b9425fededbd6684add7cff7cc

Receipt

21a29046de961e252372d06fd85d98db767b900982f90421cc720dfb85069365

These digests establish retained byte/object identity under the recorded run; they do not establish correctness by themselves.

05 / FALSIFICATION WALKTHROUGH

Inspect the causal state contract without mutating the public backend.

reference→poison→SUPERSEDED_BY / CONTRADICTS→antidote→restoration check
Open read-only walkthrough tableTrace SeedOfTruthRelationship contract