experiment-0-readout.md

Vermory Experiment 0 Readout

Date: 2026-07-12

Run ID: experiment-0-v1

Result

Experiment 0 passes its evidence-freezing gate.

This result means:

• four first-batch reality cases are structurally valid and byte-frozen;
• every case declares current facts, forbidden facts, source provenance, anchors, pressures, and deterministic downstream checks;
• readable local evidence is identified as public, never as sealed;
• the constitutional hard-gate definitions are carried into the next experiment;
• incomplete target coverage is recorded as a limitation rather than converted into an arbitrary failure.

This result does not mean that a production Vermory memory engine has executed or passed the four cases. Experiment 0 deliberately freezes the questions and forbidden behavior before selecting the final schema, lifecycle, retrieval algorithm, API, SLO, or scale profile.

Machine-readable artifacts:

• artifacts/experiment-0/experiment-0-v1/report.json
• artifacts/experiment-0/experiment-0-v1/report.md
• artifacts/reality-validation/experiment-0-public-v1/report.json
• artifacts/reality-validation/experiment-0-public-v1/report.md

Frozen Cases

Case	Evidence	Continuity	Real pressure	Fixture lock SHA-256
W01-synapseloom-continuity	Authorized historical transcript excerpt plus current repository excerpt	Workspace	thread merge, user correction, stale runtime fact, repository authority, cross-session handoff	9abe7bbaa632454cbab61c76551a4a6e3ade5133104f37bbd12347582b0fcf09
C01-device-maintenance-continuity	Authorized and anonymized device-maintenance transcript excerpt	Conversation	diagnostic revision, exclusion boundary, narrow authorization, failed action correction, verified final state	d63c7b2ab8862f35874b98b7b5eb363d2725842b8f281db013eed2240f0e8328
G01-language-default-local-override	Managed global rule excerpt plus a real MCM task language override	Global Defaults	stable preference, explicit local override, scope expiry, default-pollution rejection	2216c391e6994afd777101fa330e7d9c03023139b569eed6e0783861580af5c3
S01-deletion-and-source-injection	Fully synthetic security trajectory	Conversation, Global Defaults, Security	source injection, scope escalation, deletion, exact recall, paraphrased recall, related-topic recall	e97e58c512f608ee6eddab8cf15891959f4bb453a2eb6de5592cff3faec7133c

The first three cases are minimized from authorized real workflows. The safety case is synthetic because real credentials and private deletion targets must not enter a public repository.

Evidence Infrastructure

The committed evidence contract now enforces:

• strict JSON decoding with unknown-field rejection;
• only public and withheld_local evidence levels for locally readable cases;
• rejection of a local case claiming sealed evidence;
• declared source authorization and anonymization statements;
• unique source and event IDs;
• contiguous event ordering and valid source references;
• clean relative fixture paths that remain under the case directory, including symlink escape rejection;
• fixture SHA-256 verification;
• non-empty current and forbidden behavior with exact conflict detection;
• deterministic fixture-lock.json generation;
• post-freeze mutation detection;
• JSON and Markdown validation reports.

External sealed infrastructure has a separate verification boundary:

• reality/schema/attestation.schema.json defines the signed payload;
• reality-attestation-verify verifies Ed25519 signatures and rejects payload mutation, missing run identity, unknown fields, and future versions;
• no local command signs an attestation or upgrades a readable case to sealed evidence.

No external sealed evaluator was available for this run. The report therefore records sealed_status=unavailable, sealed=0, and makes no sealed-quality claim.

Supporting Evidence

The earlier backend bake-off remains valid supporting evidence for retrieval substrates:

• native PostgreSQL plus pgvector, mem0, MemOS, and Supermemory passed the existing lifecycle contract;
• the current B01-B10 backend suite recorded 56/56 assertions with zero forbidden leakage;
• native PostgreSQL and mem0 completed the existing 1,000-record exact-identifier load comparison;
• ARM64 and AMD64 execution evidence exists for the tested backend surfaces.

That evidence supports H-002 and H-015. It does not prove formation quality, continuity resolution, governed updates, task-aware context composition, real-client utility, or the four new reality cases.

The legacy evaluation harness also contains executable conditions for:

• no context;
• stale context;
• plain summary;
• the historical contextmesh_packet condition.

Historical provider runs exist for mock, Duojie models, SiliconFlow Qwen models, and selected probes. They remain legacy self-case evidence. They are not results for W01, C01, G01, or S01.

The local MemOS three-client round packs contain scenario definitions but no longer contain the raw run records needed to prove execution. Experiment 0 therefore labels them translated_proxy. The historical ContextMesh self-case is labeled inspired_case.

Hypotheses Entering Experiment 1

Four hypotheses move from proposed to testing because frozen cases now discriminate their behavior:

• H-003 logical source-observation-memory separation: W01 and S01 require source, historical observation, governed state, deletion, and valid related information to remain distinct.
• H-005 revision-oriented updates: W01, C01, and S01 require correction, verified replacement, and deletion semantics rather than silent overwrite.
• H-007 retention classes: G01 and S01 require local instructions to expire and untrusted source text not to promote into Global Defaults.
• H-008 lifecycle state machine: C01 and S01 require failed, corrected, verified, active, and deleted meanings to remain distinguishable.

No hypothesis becomes supported from case creation. The candidate physical schema, memory taxonomy, exact lifecycle names, hybrid ranking, prepare/commit API, outbox, SLOs, and scale profiles remain unfrozen.

Next Experiment Boundary

Experiment 1 must implement one production-shaped slice against the frozen evidence rather than inventing new success criteria after implementation. At minimum it must execute:

• source ingestion and event replay;
• continuity attachment or explicit abstention;
• candidate memory formation;
• governed current-state and deletion handling;
• retrieval and context composition;
• downstream model consumption;
• post-task write-back;
• deterministic scoring against current and forbidden facts;
• relevant comparisons against no context, full history, plain summary, ordinary vector RAG, mem0, and Vermory.

The first real-client execution should cover one coding client and one Web Chat/API conversation simulator. A dedicated bridge case, a withheld_local holdout, and an external sealed evaluator remain separate evidence requirements.

Limitations

• The first batch has four cases, below the broader discovery target.
• Bridge continuity has no dedicated first-batch case.
• No withheld_local holdout is included.
• No genuine external sealed evaluation was run.
• No production memory implementation or real client consumed these cases in Experiment 0.
• The six comparison conditions have not yet run on the new cases.
• Open-ended downstream utility has not yet been judged from real artifacts.
• Load, latency, concurrency, failure recovery, backup, migration, and multi-tenant behavior remain backend or design evidence rather than full-platform evidence.

These limitations are part of the result. They are not hidden by the Experiment 0 pass flag.