One view of the current checkpoint.

Native deployment observation: V151 · source bc1bae37bd4a229664d0da9b72b4adb8e5f9dc42 · succeeded. This is the previous published baseline, not a claim that its archive contains this update.

V151 archive: sha256:a3fadc28c4474f488df24c112f21bb4413f2a1af707a0408e0275f7cd6b7fcc9

e72bcfedf776861f37c3c7a4cb8aca91ef6e96853991910615fce7d08e32ea23

Recompute SHA-256 for each listed file, serialize the ordered path/hash entries as JSON, then hash that JSON. Compare this value with your authorized source checkout and exported readiness report. This is a scoped build fingerprint, not protected attestation or independent deployment proof.

V145: 1,593 Node and 33 Python tests passed. The private signed walkthrough covers FINAL, NOT FINAL, CONFLICT, REOPEN and STOP. V146: 1,624 Node and 33 Python tests passed. These are dated seller-controlled checkpoint results, not test counts for this newer build.

V146 actual APEX-0 diagnostic: 0/20 correct with one candidate and 0/20 with three; generated tokens increased from 640 to 1,920. No quality or efficiency improvement was established. The model remains unqualified.

The current owner-local research runtime checks retained context before model access, between adaptive candidates and before recording results. The new CPU inference path avoids constructing training graphs. A 25-pair warm-process comparison records identical generated outputs; its local timing gain is not improved intelligence, production latency or independent qualification.

A first-party research candidate was trained for 4,000 fixed CPU steps from the retained V148 model. It answered 117 of 129 held-out compact addition expressions correctly; V148 answered 0 of 129. Both used the same 19,800-parameter architecture and four-token inference limit. Reversed operand pairs remained in the same split. All frozen predictions reproduced in a second local replay.

The older natural-language arithmetic diagnostic remains 0 of 20. This is one seller-visible synthetic task and one training seed, not general reasoning, independent qualification or a frontier comparison. The candidate is not admitted to the serving registry. Energy and total cost were not measured.

Five permanent programs pursue stronger intelligence, computational efficiency, governed self-improvement, autonomous invention and independent competitive qualification. The aim is to understand an authorized organization, invent useful improvements and prove what changed across successive generations. Production authority stays with Finality Core and Control.

The protected evaluation workspace now builds a tenant-scoped map from explicitly approved offline source exports. Reviewers choose categories, fields and relationships. Current signed full snapshots produce source-linked facts and connections; hypotheses, contradictory values and missing categories remain visible. Revoked, expired, partial, failed and stale sources are excluded from the current map.

Duration targets and existing observation diagnostics produce reviewable improvement hypotheses. They do not establish causes, realized savings or permission to change a system. Two synthetic organization scenarios were exercised locally; actual customer and unassisted browser acceptance remain unverified. This is an offline signed-snapshot connector, not automatic integration with every enterprise system.

An approved company duration finding can become a versioned read-only experiment. The owner reviews the exact source baseline, design, assumptions, risks, recovery and acceptance criteria, then approves the specific draft with an expiry. The workspace calculates a projected comparison and can separately compare a newer signed source observation. Financial arithmetic uses exact integer minor units and counts one modeled time-capacity benefit.

Projections, synthetic observations, reported before/after data, experimental attribution and realized benefits remain distinct. Quality stays unknown when it has not been measured. No customer system changes are executed and no simulated gain becomes realized savings. The portable outcome report replays with a separate configuration pin; external buyer and authenticated browser acceptance remain unverified.

The frozen candidate scored 0/120 on larger operands and 0/60 when spacing changed. On 30 diagnostic time-delta proposals it scored 0/30; exact deterministic arithmetic scored 30/30. No new training, general reasoning gain or model promotion is claimed. The prior compact-addition result remains a narrow historical result.

The protected company workspace groups duplicate relationship claims while retaining their source records, shows missing endpoints and type conflicts, and identifies declared dependency cycles. Three optional illustrative scenarios use exact units and source-linked inputs. Their combined value is withheld because time, labor and rework may overlap. No scenario grants approval or establishes realized savings.

Eight industry descriptions are metadata rather than trained specialists. Original signed imports, owner approvals, expiry, revocation and portable company and outcome formats are preserved. No unsigned candidate store, browser tenant token or external model connection replaces them.

On 60 changed-format expressions using previously disclosed arithmetic pairs, the frozen V149 candidate answered 0 correctly from raw prompts and 59 after syntax-only normalization; exact arithmetic answered 60. This is a research-only grammar adapter, not new model learning, general reasoning or an independent hidden benchmark. It changes no customer inference or production admission. The larger-operand and other V151 transfer failures remain preserved.

Sign in, create a synthetic walkthrough, inspect the approved source map, check readiness and export the evidence package and scoped report. Keep the configuration pin separate. The report includes this build fingerprint, package digest and explicit unresolved acceptance checks.

Browser acceptance is not verified. Pilot remains blocked. High-assurance production remains on HOLD with all 18 gates open. Independent validation and customer-live operation remain unestablished. Historical Site 80, V126 hosting observations and V132 component measurements retain their original meaning.