Hack Hydra 2026 · Track 03 · HydraDG
See the context move.
HydraDG keeps changing memory as a traversable custody graph: current state, prior state, contradiction, supersession, source, and claim ceiling remain linked instead of collapsing into one answer.
LongMemEval full500 executedHosted canonical FCG readback receipt recordedOne SHA-256 identity per canonical FCO
Context Iceberg · T0–T2 scientific fixture
Antidote / restoration
G*—
ΔG*—
Cloud Drift—
Retrieval ΔN/AN/A for synthetic fixture
FCG root: synthetic control / no hosted root in this fixture · HydraDB root: synthetic control / no hosted root in this fixture
NOT_SIGNED · NOT_MERKLE_COMMITTED. T3–T5 release states are shown separately; no scalar context score is fabricated without a declared distribution.
Public data plane
Hosted HydraDB not connected on this deployment
BackendCHECKING
DatabaseNOT_CONFIGURED
CollectionNOT_CONFIGURED
TraceabilityNOT_CONFIGURED
This card is only a live server-side connectivity/readback diagnostic for the current deployment. HydraDG's retained local HydraDB executions, graph use case, deterministic math and bounded hosted receipts are separate evidence objects. Credential values are never returned to the browser.
01 / GOLDEN PATH
Open table walkthroughReference → poison → antidote.
One fact changes. HydraDG preserves the old state, records the relationship that diverged, and restores the declared current state without deleting the perturbation history.
Reference / normal
Violet · frozen comparison state.
Poison / mutation
Orange · controlled divergent state.
Antidote / restoration
Blue · recovery while poison history remains.
The graph worked. The tested retrieval advantage did not appear.
The completed full500 ablation constructed and queried the typed graph but did not establish a positive B/C/D hit-rate signal over the flat route at the tested configuration. HydraDG retains that null/negative evidence instead of optimizing it away.
Scientific ceiling
Retrieval ablation only.
Not an end-to-end LongMemEval QA improvement claim. Hash identity is not correctness.
Custody boundary
History stays visible.
Negative, null, superseded and restored states remain first-class graph objects rather than being deleted.
Scientific state and release state are not collapsed into one score.
| Timepoint | Class | Scalar context score | Evidence |
|---|---|---|---|
| T0_REFERENCE · Reference / normal | SYNTHETIC_FIXTURE | G* -0.061230 · ΔG* 0.000000 · Drift 0.0000 | Declared deterministic fixture baseline. |
| T1_MUTATION · Poison / mutation | SYNTHETIC_FIXTURE | G* 0.572956 · ΔG* 0.634186 · Drift 40.3629 | Declared controlled perturbation fixture. |
| T2_RESTORATION · Antidote / restoration | SYNTHETIC_FIXTURE | G* -0.027496 · ΔG* -0.600452 · Drift 1.8729 | Declared restoration fixture; counterevidence remains in the graph. |
| T3_HOSTED_MIGRATION · Hosted HydraDB migration | PRODUCTION_RELEASE_STATE | N/A — no declared distribution | Hosted parity receipt: 36 canonical FCOs, 24 edges, zero canonical set/edge/content-hash delta. |
| T4_CONTEXT_VS_ENTROPY · Context vs Entropy | PRODUCTION_EXPERIMENT_STATE | N/A — no declared distribution | 18,567 raw findings; 18,555 context-classified; 12 abstentions; 99.9354% coverage. |
| T5_FINAL_JUDGE_RELEASE · Final judge release | PRODUCTION_JUDGE_RELEASE_STATE | N/A — no declared distribution | Exact deployed Git SHA and WebsiteRelease FCO are resolved at build/runtime; no scalar G*/Cloud Drift is fabricated without a declared distribution. |
T3–T5 are real project/release transitions, but HydraDG does not fabricate a G*/Cloud Drift scalar when the required distribution is not declared.
One custody spine. Three graph problems.
The final judge path emphasizes Track 03 while preserving the shared custody spine used by the broader HydraDG track experiments.