# Babel Researcher — development and hosting status

Checked September 26, 2026. Package version 0.1.1. This is a manual audit snapshot, not a live monitor.

## Assessment

The offline evidence pipeline and read-only knowledge connector work. The AI layer is not currently available because its local services did not respond. The project is a usable research foundation, not a hosted autonomous research system.

## Selected architecture

Research execution, model inference, the library and private evidence packets remain on the local PC. Cloudflare hosts this static website and reviewed status snapshots. Published pages remain available when the PC is off; new AI research requires the PC and model service to be running. No remote job queue, incoming PC connection or automatic report upload is included. Oracle is not used in the selected setup.

Publication is a separate manual step: check the evidence, review the public files, freeze a release with hashes, verify staging, then publish that same release. A deployment does not refresh the research audit date. [Cloudflare documents static asset requests as free and unlimited, with no additional asset storage cost](https://developers.cloudflare.com/workers/static-assets/billing-and-limitations/).

## Fresh evidence

| Check | Result | What this establishes |
| --- | --- | --- |
| Researcher test suite | 26 passed, 0 failed | Deterministic unit and integration behavior, including adapter contract tests |
| Strict local example run | Succeeded in 152 ms | Offline pipeline executed over the three bundled example documents |
| Packet content | 3 sources, 9 exact excerpts, 3 source-observation claims | Traceable local document observations; no model synthesis |
| Review | All 6 gates passed; no blocking failures or warnings | Configured requirements passed; source diversity required only one group |
| Manifest verification | All packet file hashes matched | Saved packet integrity, not truth or independent corroboration |
| Fitness library retrieval | 18 domains, 108 topics, 147 sources, 160 claims, 22 roles | Current read-only SQLite inventory |
| Role contexts | 22 of 22 retrieved with resolvable source references | Each role produced a context packet against the real library |
| Full library integrity | Failed for 4 files | Generated notes and a Canvas differ from their saved manifest |
| AI.OS and Ollama | Neither endpoint responded | AI synthesis and keyless local-model web research unavailable at check time |
| Optional GPT Researcher package | Import succeeded | Package availability only; not proof of a successful web investigation |

The fresh example is `status-local-20260926`. It completed at 20:26:11 UTC. Its elapsed time is pipeline time for tiny local example files, not an inference benchmark.

## Capabilities and limits

- **Local engine:** reads Markdown, text, JSON and CSV; ranks exact lines; records sources, excerpts, source-observation claims, gates, audit events and hashes. It does not write a reasoning-based synthesis.
- **Knowledge connector:** performs read-only full-text search and role-context retrieval. The webpage publishes topic titles and two saved retrieval examples. Browser filtering is not live SQLite search or a new research run.
- **AI.OS adapter:** supplied evidence IDs constrain model claims; deterministic tests pass. A historical acceptance result is not treated as current. Fresh live model acceptance remains pending.
- **GPT Researcher adapter:** optional dependency is installed. The keyless profile configures DuckDuckGo plus local Ollama language and embedding models. No fresh web run was performed because local inference services were unavailable. Search terms leave the PC. External prose stays unverified candidate synthesis: normalized URL records are not independently retrieved source text, and their locator hashes are not page-content hashes.
- **Review:** current gates check coverage, citations, exact-excerpt flags and report presence. The named adversarial-review module is not an independent scientific fact-checking agent.
- **Default configuration:** still uses the deterministic fake engine with networking disabled. Default success is not proof of real research.

The original core has no authenticated hosted API, job queue, scheduler, retention policy or model budget controller. This new website documents the researcher and provides an interactive snapshot explorer; it does not add those missing runtime services.

## Library integrity discrepancy

The verifier reports drift in:

- `10-domains/strength-performance.md`
- `20-topics/pwr-strength-foundations.md`
- `20-topics/pwr-taper.md`
- `Fitness Research Map.canvas`

These files were not overwritten or rebuilt. SQLite retrieval passed, but edited notes may disagree with the database until the changes are reviewed and reconciled. The library holds curated paraphrases rather than verified exact quotations from papers. Its source access inventory includes abstract-only records. No new scientific review was performed in this audit.

## Oracle inspection

Read-only SSH and instance metadata checks found an Ampere `VM.Standard.A1.Flex` server with 2 OCPUs and 12 GB allocated memory. At inspection it had about 10,227 MiB available RAM, 18 GB available root-disk space, very low load and unused swap. The default Python is 3.9.25; Researcher requires Python 3.11 or newer. Node and cloudflared were not found on the default command path; nginx, caddy, httpd, docker and ollama were inactive.

The webpage is a strong capacity fit. The dependency-free Python core and SQLite retrieval are plausible fits after installing a supported Python runtime and validating on ARM Linux. An LLM might fit in memory, but acceptable CPU inference latency, dependency compatibility and sustained workload limits are unproven. No model was installed and no server configuration was changed.

Current [Oracle Always Free documentation](https://docs.oracle.com/en-us/iaas/Content/FreeTier/freetier_topic-Always_Free_Resources.htm) gives A1 an allowance of 1,500 OCPU-hours and 9,000 GB-hours monthly, equivalent to 2 OCPUs and 12 GB continuously. It also lists 200 GB shared block storage and idle-instance reclamation conditions. [Oracle's price list](https://www.oracle.com/cloud/price-list/) separately states a paid-tenancy allowance of 3,000 OCPU-hours and 18,000 GB-hours shared across A1 workloads.

The inspected server matches the smaller allocation. Running this static site on an existing VM would not increase its allocated cores or memory. However, account type, other instances, boot/block volume totals, backups, network usage and billing were not accessible through this SSH inspection. **Zero cost has not been verified.** Do not provision or resize anything on the assumption that capacity equals a free allowance.

The Oracle assessment is retained as background. The chosen implementation uses Cloudflare for the page and the existing PC for execution; Oracle migration is not planned for this release. Any future hosted execution would need authentication, bounded jobs, allowed inputs, budgets, retention and rate limits.

## Next work

1. Reconcile the four library edits with their canonical inputs and rebuild only after preserving intended changes.
2. Start the local AI services and record a fresh strict synthesis run, then a bounded keyless web run with quality and latency evidence.
3. Add independent page retrieval and verified excerpts to the web evidence path.
4. Overhaul the researcher and webpage on top of this local-execution/publication boundary.
5. Add remote job submission only if needed later; it would require authenticated access, explicit limits and a local worker.

## Publication boundary

Public website assets contain sanitized technical status, reviewed bundled-example excerpts, public topic titles, and aggregate capability metadata only. They contain no machine addresses, private filesystem locations, credentials, personal health records, research workspace dump, or callable local-service bridge. The server address and account access remain outside this public report.