I want to explore a local assistant that gradually learns about the world, recognizes its own limits and grows through tools and checked observations. The AMA experiments begin with clean data and random initialization, keep resource use modest and preserve failures alongside successful runs.

AMA-1 Genesis started from a small v0.1 prototype. Codex launched AGY to implement changes and independently reviewed the work. After two test rounds exposed repairs that covered only known paths, I authorized Codex to contribute separately identified patches. The result is a runnable foundation. The full capability target is still unfinished.

The goal and the current system

The six directions are training from scratch, a truthful self-model and stable personality, world and object knowledge, reviewed continual learning, lightweight tool use, and trustworthy networking. The intended order is self-model, objects, tools, correction and learning, followed by more natural conversation and multimodal perception.

The component actually trained here is a 516-parameter linear routing classifier using character bigrams. Its 24 synthetic examples distinguish self, world, search and other requests. It is not a language model. Object attributes in SQLite do not constitute visual recognition.

The personality contract takes inspiration from Kal’tsit’s calm, rational independence while AMA-1 keeps its own identity. Subjective consciousness remains an open research question. Functional state reporting does not establish subjective experience.

Implementation and review had separate jobs

The original files were inventoried and frozen before a clean Git baseline was saved. No old project code, datasets or model weights were transplanted. Review used source code, actual command output and checkpoint provenance; historical writeups were not accepted as evidence of a new run.

The initial 16 tests passed. Additional checks found malformed checkpoints being accepted, contradictory knowledge surviving correction, unsafe URL handling and CAPTCHA pages being counted as successful searches. AGY added typed object records, episodic events, a tool registry, candidate manifests and rollback operations, together with tests.

Those additions had their own gaps. Redefining an object could discard its identity. New property APIs bypassed review. A matching hash could still describe an invalid model. A corrupt backup could replace a valid model. Even a fixed search endpoint could redirect toward a private address. The network reproduction used a mocked transport and made no actual private-network request.

How the two corpora were separated

Both corpora were frozen and hashed before evaluation, and the implementer was explicitly prohibited from reading them. The original 24 training rows stayed unchanged. Review checked literal and Unicode/case-normalized overlap and inspected visible tool traces.

This was an implementer-hidden independent evaluation, not a formal double-blind experiment. There was no OS access-control separation, and the reviewer knew the questions. The classification task also shares intent labels and domain semantics across its splits.

Once first-round feedback guided repairs, that corpus became regression data. The second round used a different, previously undisclosed corpus. Repairs after those rounds were evaluated as regression work, not fresh blindness.

Routing evaluation Correct Accuracy Majority baseline
Round 1 9/24 37.5% 25%
Round 2 8/24 33.3% 25%

The training-majority baseline is 25% in both rounds. These are routing results, not measures of object recognition or language understanding. An emoji input could still be classified as search, and reported confidence had no independent calibration. A prediction therefore cannot grant tool permissions.

Passing repairs must not erase failures

The second round found invalid training schedules, partial weight mutation before a later bad sample was rejected, control characters in URLs, truthy string authorization, alias shadowing and missing backup lineage checks. AGY handled routing and URL validation. Codex separately patched knowledge review, permissions, typed corrections, transactions and rollback checks.

The final full regression run passed 85 tests, with zero failures, errors or skips and exit code 0. Another 17 actual CLI invocations exercised candidate properties, approval, aliases, correction, arithmetic and default network denial.

Codex-authored patch checks are author regression evidence, not an independent audit of those patches. Both rounds’ failures, predictions and false positives remain in the repository. Training accuracy reached 100%, but unseen results still show inadequate competence. The actual checkpoint remains a candidate and was not promoted into production.

Resource and network claims stay within the evidence

The observed peak RSS was about 17.72 MiB in the standalone training process and 25.34 MiB in a separate CLI process. These are single samples from different processes. RSS did not fall after object unloading, so the experiment does not establish that memory returned to the operating system.

The 6GB VRAM budget is still a target. CUDA allocations, reserved memory, global GPU footprint and future perception-model requirements were not measured.

Networking is denied by default. Explicit consent is required before sending a query. Search results remain untrusted links: no automatic knowledge promotion, downstream fetching or CAPTCHA bypass. Live provider success was not verified. Independent-source checks, ownership clustering and complete GEO defenses remain unfinished.

What is available for analysis

The source and audit snapshot are published in jsw-teams/train. A successor commit replaces the current tree while retaining the previous remote history. Public evidence redacts machine paths; full raw execution records remain local. Exact dataset, checkpoint and frozen-corpus bytes are preserved so newline conversion cannot invalidate their hashes.

The next stage needs representative material with traceable sources and licenses, plus a new isolated evaluation corpus. The complete learning loop, a scratch language model, natural conversation, photo recognition and speech are not implemented. Published or feedback-exposed questions must not become the next model’s supposedly unseen test.

See the AMA model family page, audit report, executed test results and training-data manifest. Release of the final capability target remains blocked.