From a working demo to governed memory: one week building the next layer

11 September / Local and cloud

Repair the demo without rewriting its history.

The new working freeze preserves executable 16d3c8ce. Both supplied Spain-Spanish sessions reached ordinary conversation, explicit live and recent-history vision, public research, interruption/resume and normal shutdown. Local mode used Qwen cache16 with James/DDG; cloud mode used Groq with Exa answers. Core, Cohere transcription, Scylla synthesis and camera/perception ownership stayed on Thor.

The repair work covered Exa streaming compatibility, cloud cancellation, video selection relative to speech end, requested history windows, Spanish visual phrasing, the camera snapshot interface and nonblocking ASR/audio ownership. Across the two completed runs, 54 ASR requests reported no failures. Camera processing stayed around 29.6 FPS, with joined drains and no retained video at shutdown. An earlier operator-interrupted local startup remains a separate unsuccessful attempt.

These were Spanish execution observations, not a fresh bilingual acceptance corpus or proof of factual accuracy. Short cloud visual follow-ups after pause/resume still sometimes received no images; one answer was incomplete, and an isolated continuation command was missed. Latency gates remained failed. The later conversation-reliability repair is implemented on the development lineage, but this review found no new physical run qualifying that repair.

12 September / Core and persistence

A memory is more than a sentence in a database.

The governed-memory work now has executable contracts and an isolated PostgreSQL backend. Retention, use, disclosure and execution are separate decisions. A record carries source evidence, scope, purpose, retention and revision information; a model-generated sentence does not supply its own permission to become permanent memory.

The GM-01–GM-06 sequence adds deterministic policy, reviewed durable writes, correction and revision history, restriction, revocation, expiry, suppression and separate payload cleanup. Recall and output are bound to current rights rather than an old permission decision. Inspection, export and relationship controls are explicit. Action approvals are bound to exact requests, but the executors in this programme are database-only mocks—not robot skills or general external actions.

Synthetic authentication, SQL roles, row-level isolation and encrypted cold restore followed in GM-07. These tests establish behaviour in the fictional test deployment. They do not establish production process isolation, protection of every disk/log/WAL copy, independent key custody or an independently stored recovery copy. The open platform recovery requirement has not disappeared.

This path uses fictional profiles and explicitly enabled fixtures. It does not import family conversations, silently retain the full demo transcript or attach protected memory to the cloud/internet lanes. The existing conversation ledger remains session history; durable memory is a separately governed decision. Core remains the owner, and the identity created once at MS-06 is not recreated.

14–15 September / Meaning and reviewed input

Remember what was said, including its conditions.

The MI work builds on that foundation rather than replacing it with another memory framework. It preserves exact source statements and their digests, task and familiarity conditions, explicit dates and links between revisions. A correction of an error, a later change of preference and an unresolved contradiction are different operations. Missing times remain unknown; the system must not invent a date merely to fill a field.

A reviewed terminal now joins acquisition, questions, correction, change/conflict links, deletion and restart into one synthetic workflow. Supported English or Spain-Spanish statements produce a proposal for inspection. A separate confirmation approves only the displayed operation; cancellation does not commit it. Suppressing a record and purging its payload are deliberately separate steps with separate status.

The newest task-statement interpreter accepts complete requests within a reviewed task catalog. It preserves negative preferences rather than converting “I do not want…” into an affirmative instruction. The interpreter proposes annotations; it has no database, model or approval authority. Unsupported wording, mixed topics and ambiguous conditions require clarification. This is bounded bilingual input, not unrestricted understanding of everything a person says.

Scoped quotation reports provide a useful narrow result: return the eligible statement with its conditions and provenance instead of asking the language model to silently decide which preference is universally true. This also makes correction and disagreement inspectable. Broader semantic acquisition, event learning, mixed-topic decomposition and the complete cognitive-memory research programme remain open.

12–15 September / Governed conversation

A fast answer must still belong to the current session.

GM-08 first connected synthetic governed recall to the existing local Qwen text provider, then to Cohere and Scylla. The later incremental-speech candidate allows bounded speech units while generation continues, rather than waiting for the whole answer. Each output remains subject to current authorization and cancellation. The ordinary frozen demo is not silently converted into this protected-memory experiment.

A physical attempt exposed an important failure: repeated cancellation during cleanup left a stale active-generation slot, so later questions were refused. The repair separated and joined the cleanup task without reopening the cancelled output. In the supplied 14 September retest at b8d7983c, 402 backend cases passed; four replies completed and two were interrupted. Fresh English and Spanish replies completed after the interruptions under the same synthetic login, with final cleanup and report statuses at zero.

First software playback writes took 2.139–2.666 seconds across those six generations. This shows bounded recovery and overlap, not acoustic onset, scored answer correctness or an accepted latency improvement. A later prefetch/revocation problem and its repair remain in the record. The newest bilingual revoke/refusal/reauthentication/fresh-reply check is implemented but has not yet supplied physical verification. G8 remains open.

15 September / Latest supplied target result

903 required methods passed. The scope matters more than the count.

The operator-reported run at 1122c4c4 passed all 903 required backend methods with zero required skips: 341 host methods and 562 PostgreSQL methods, including the new task-workflow cases. It exercised reviewed synthetic proposals, preserved negative quotations, confirmation, persistence, process restart and current-rights behaviour. The full source-check completion marker was also present.

The source runners separately reported 3,276 unittest cases with 33 skips, and 4,093 pytest passes with 542 skips and five collection warnings. Those runners overlap and must not be added together. Nor are earlier 795-, 826-, 845-, 868- and 881-method checkpoints extra independent trials of the latest candidate: they are cumulative development snapshots with their own scopes.

The evidence is a sanitized repository record derived from the supplied Thor log and its pinned command/source association, not a new independent hardware run for this publication. A backend pass does not score an arbitrary model answer. Separate actual-model probes ended at 74/80, 79/80, 62/64, 63/64 and 127/128 against their respective strict gates; all remained failed. Schema, citation and task-selection mistakes cannot be erased by later deterministic report tests.

Current boundary / Next controlled work

Close the synthetic proof before opening the participant programme.

The immediate sequence is G8 governed-conversation qualification, then G9 consolidated synthetic acceptance. G9 needs a requirement/case/source/evidence map, not a conversion from test count into a completion percentage. The original 60 bilingual scenario specifications, 12 fault schedules and the newer 96 cognitive-memory specifications are different inventories; the latest backend pass does not qualify all of them.

Deployment security and independent recovery remain G7 work. G10 and MS-10 require separately approved participant, authentication, retention, withdrawal and incident protocols, followed by the defined adult evaluation. MS-07, MS-08 and MS-09 retain their internet, voice and perception gates. Head presence remains architecture-planned under MS-11; powered motion, arms and mobility have separate safety programmes. No assembled head or motion result is claimed by this week of software work.

Foundation v1.2 still controls. Foundation v2.0.1 and Alpha v0.3 remain unapproved, including the unresolved staging conflict with section 74. The achievement is a much stronger synthetic foundation for remembering responsibly—not a claim of perfect memory, unrestricted cognition, validated household use or Alpha readiness.