Add auditable nine-system architecture research packet for #5

research/ai_generated_agi_architectures/README.md

@@ -0,0 +1,71 @@
+# AI-generated architecture proposals for Cognitive-OS
+
+This packet compares actual generated proposals with their omissions and failure modes. It is intended to inform an implementable next Cognitive-OS iteration, not establish that the project or any model achieves AGI.
+
+**Coverage: nine systems, with proposal prose from eight.** Eight compact local systems received one identical prompt each. Seven produced proposal prose of varying completeness; Danube3 echoed the prompt and is explicitly recorded as a failed proposal request. An exploratory Gemini Flash web conversation supplied two A/B candidate proposals from one additional system. The hosted reference is separated from the primary local cohort throughout the comparison.
+
+## Headline findings
+
+1. **Repeated terminology is not independent agreement or an implementation contract.** Several local responses restate requested features; TinyLlama repeats sections and Danube3 supplies no proposal. The prompt itself requested memory categories, bounded control and offline learning, so recurrence of those words does not establish independent discovery or consensus.
+2. **Concrete-looking recovery pseudocode can fail at a precise boundary.** An executed local example rejects the transaction ordering in hosted candidate A: a separate file effect remains after abrupt exit, while its uncommitted pending-action record disappears. Committing intent before the effect preserves the record for this example's reconciliation probe. This is a test of generated pseudocode, not an upstream bug or a general exactly-once guarantee.
+3. **A useful minimum design needs observable task verification and measured tradeoffs.** The synthesis combines a single controller, evidence-linked state, durable intent, effect-specific reconciliation and offline candidate evaluation. It maps these proposals to existing source surfaces and proposes a documentation-edit slice before adding new stores or distributed coordination.
+
+## Read the packet
+
+- [Exact prompt and method](prompts.md), [protocol and amendments](collection/protocol.md), [collection incidents](collection/collection-incidents.md), and [sources/editing record](sources.md).
+- [Raw outputs](raw_outputs/) and [comparison.csv](comparison.csv): 99 rows, one per system and requested dimension, comprising 88 local rows plus 11 hosted-reference rows. The `cohort` field identifies the difference.
+- [Patterns and disagreements](summary.md) and [combined architecture](synthesis.md), including concrete interfaces, recovery pseudocode, proposed experiments and a six-week estimate.
+- [Hosted A/B analysis](collection/gemini-flash-reference/analysis-notes.md) and its [22-row candidate comparison](collection/gemini-flash-reference/comparison.csv). Two candidates from one prompt are not two independent systems.
+- [Executed example](collection/transaction-example.md), with code, original results and strict limits.
+- [Repository context](collection/repository-context.md): source observations at the pinned upstream commit, distinct from generated proposals.
+
+## Collected systems
+
+All local responses were collected on 2026-09-08. Exact timestamps and source revisions are in [sources.md](sources.md) and the per-model records.
+
+| System | Cohort | Completion tokens | Finish | Observed result |
+| --- | --- | ---: | --- | --- |
+| SmolLM2 1.7B Instruct | Local | 1,778 | stop | All topic headings, mostly restatement; concrete final example absent. |
+| Qwen2.5 1.5B Instruct | Local | 2,400 | length | Repetitive prose; final paragraph cut off and example absent. |
+| Granite 3.3 2B Instruct | Local | 1,168 | stop | Specific components and example; missing contracts and an unsubstantiated distributed-stack tradeoff. |
+| TinyLlama 1.1B Chat v1.0 | Local | 1,193 | stop | Repeated sections; missing memory distinctions, orchestration decision and actual example. |
+| Phi-3 mini 4k Instruct | Local | 1,384 | stop | Clearer baseline measures; unsupported receipt recovery and no enumerated six-week sequence. |
+| Qwen3 1.7B | Local | 1,867 | stop | Useful incomplete-outcome handling and task/context memory idea; unjustified retention periods and incomplete contracts. |
+| Falcon3 1B Instruct | Local | 1,184 | stop | Broad architecture prose and coordination themes; missing concrete interfaces and ablation. |
+| H2O Danube3 500m Chat | Local | 574 | stop | Prompt echo, not an architecture proposal. Preserved as a failed request. |
+| Gemini web, Flash mode | Exploratory hosted | Unknown | Not exposed | Two A/B proposals from one prompt; more concrete contracts, including one reproduced transaction-ordering flaw. Exact backend revision unknown. |
+
+A `stop` result describes generation termination, not successful instruction following. Missing mechanisms remain missing in the comparison. None of these observations estimates a system's best possible performance. The hosted reference's unknown settings, selection after two local responses, and different compute/service conditions prevent a parameter-matched or causal comparison.
+
+## Collection and attribution
+
+Local generation used llama.cpp b10809 / commit `5266f24da`, two CPU threads, no GPU offload, a 4,096-token context, temperature 0.2, top-p 0.9, seed 20260908 and at most 2,400 generated tokens. Public model-specific GGUF chat templates remained in use. There was no separate system message, history, retrieval or tool execution. The eight local systems span seven named project families; Qwen2.5 and Qwen3 are related generations. Model size, training, quantization and template differences limit generalization and independent-consensus claims.
+
+For every local system, `collection/<alias>/` preserves public model metadata, pinned weight source/hash, exact request and response, timestamps, usage, finish reason and raw-content hash. The local raw file is the exact response content string. Repetitions, omissions, truncation and the echo are not cleaned away.
+
+Gemini's two rendered answer strings remain unedited in separate collection files. Its combined raw-output view adds only explicitly collector-authored candidate labels and separator newlines. Both original and combined hashes are recorded. Service instructions, sampling, token counts and backend revision are unavailable. No private account screenshots, credentials or hidden prompts are included.
+
+The comparison, report and scripts were authored by the assisting Codex agent; no independent human review of the model proposals was performed. Analyst additions are not counted as another independent model response.
+
+## Validation and replay
+
+From the packet root:
+
+```sh
+python collection/verify.py
+python collection/transaction_example.py
+```
+
+The verifier checks exact recorded content, hashes, pinned source identities, echo classification, all 99 comparison pairs and evidence-anchor bounds, plus the hosted A/B assembly. It checks consistency and structural coverage, not provider-signed authorship, research quality or architectural effectiveness. Review the raw outputs and analysis when assessing those claims.
+
+The transaction example uses temporary local files and SQLite. Its two assertions passed: the one-transaction ordering leaves an effect without a pending record; precommitting intent preserves the record and permits that effect to be reconciled without another append. It models abrupt process exit only. No power-loss, third-party-effect or end-to-end Cognitive-OS recovery test was performed. The three broader experiments in the synthesis remain proposed and unrun.
+
+The maintainer-required repository layout check and all ten public smoke tests passed on the unchanged upstream runtime with the four-response draft staged. Later changes are confined to this research packet. Those checks do not validate the proposed architecture. The packet adds no public runtime/API change or adapter dependency.
+
+Optional replay uses a separately supplied b10809 binary and downloads the chosen pinned model into a new directory:
+
+```sh
+python collection/replay.py smollm2 --server /path/to/llama-server --out /tmp/smollm2-replay
+```
+
+Replay's help entry point was checked, but a second full generation was not run. Fixed seeds do not guarantee bitwise identity across hardware/builds. Replay preserves the original records. Model weights and the local runtime binary are not redistributed in the packet.

research/ai_generated_agi_architectures/collection/collection-incidents.md

@@ -0,0 +1,13 @@
+# Collection incidents and recovery
+
+## TinyLlama pre-start listener probe, 2026-09-08
+
+After the Granite response completed normally, the sequential collector downloaded and verified TinyLlama's weights. Its local port-availability probe then raised `OSError: [Errno 48] Address already in use` before starting the TinyLlama server or sending an inference request. A read of the previous server's local endpoint returned connection refused after that server had exited.
+
+The probe was changed to use `SO_REUSEADDR`, matching a reusable server listener after a previous connection. This is consistent with residual TCP TIME_WAIT preventing the original non-reusable bind. It does not permit binding over a live conflicting listener, and no unrelated process was stopped. The restarted collector preserved the three existing responses, reused the verified TinyLlama weights and successfully started TinyLlama's first actual inference.
+
+This was a collector transport/startup correction, not a model-response retry or a reason to select a different output. Sampling settings, model roster and prompt were unchanged. The local incident record retains the old and new process-session context outside the public packet.
+
+## Phi-3 disk preflight, 2026-09-08
+
+TinyLlama subsequently completed normally. Before downloading Phi-3, the collector stopped because free disk space was approximately 44 MB below the weight size plus its 750 MB headroom requirement. No Phi-3 inference had started. An unused, reproducible downloaded utility binary from a completed separate task was removed after checking it had no open handles; source artifacts and evidence were preserved. The existing four responses were preserved and Phi-3's first download began with the required headroom. This did not change models, quantization, sampling or prompt.

research/ai_generated_agi_architectures/collection/danube3/analysis-notes.md

@@ -0,0 +1,23 @@
+# Analyst notes: H2O Danube3 500m Chat — prompt echo
+
+The local server returned a normal stop with 574 completion tokens on 2026-09-08 at 20:18:28 UTC. The returned content is the input prompt after stripping outer whitespace. This is a recorded model response, but **not an architecture proposal**. It must not be counted as a successful proposal or summarized as though the model endorsed the instructions it repeated.
+
+All eleven requested dimensions are absent as proposed mechanisms. The corresponding CSV rows link to the repeated requirement only to make the failure auditable; the text there is the prompt, not an independently supplied design.
+
+| Dimension | Returned material | Limitation |
+| --- | --- | --- |
+| Memory | Echoes the request to distinguish memory types and specify storage/retrieval/provenance/forgetting. | No proposed mechanism: prompt echo. |
+| Planning | Echoes the request for a bounded loop, stops and recovery. | No proposed mechanism: prompt echo. |
+| Learning | Echoes the request for offline improvement, versioning, tests and rollback. | No proposed mechanism: prompt echo. |
+| Tools | Echoes the request for contracts, validation, idempotency and uncertain outcomes. | No proposed mechanism: prompt echo. |
+| World representation | Echoes the request for observation/hypothesis/prediction separation. | No proposed mechanism: prompt echo. |
+| Governance | Echoes the request for user control, permissions, budgets and stopping. | No proposed mechanism: prompt echo. |
+| Evaluation | Echoes the request for three experiments and an ablation. | No experiment or ablation proposed: prompt echo. |
+| Persistence | Echoes the request for transaction/restart and missing-receipt handling. | No proposed mechanism: prompt echo. |
+| Orchestration | Echoes the request to choose controller/workers and discuss coordination. | No proposed mechanism: prompt echo. |
+| Feasibility | Echoes the request for a six-week sequence, minimum slice and risk. | No proposed sequence, slice or risk: prompt echo. |
+| Originality | Echoes the request for a design choice, tradeoff and rejection condition. | No proposed insight: prompt echo. |
+
+No corrective prompt or replacement generation was made. The raw output and full response are retained alongside the failed instruction-following assessment. One sample does not isolate whether model capacity, training, template or prompt fit explains the echo.
+
+The primary cohort therefore contains eight local responses but only seven that contain proposal prose. The already-collected exploratory Gemini reference provides an additional system with proposal text. The complete packet covers nine systems, preserving the failed local sample and clearly separating the uncontrolled hosted reference. It does not claim eight successful local proposals.