# OERSI Word Lab — implementation and deployment report

**Date:** 2026-10-29
**Host:** `srv7700305575` (`89.44.240.18`)
**Status:** live experimental client-side candidate generator; every proposal requires human review.
**Feature commit:** `f03962d` on `release/oersi-word-lab-v0.1.1-20261029`

## 1. What was started and deployed

- Added OERSI Word Lab 0.1.1 to the published Language Forge page.
- The browser-side generator combines ordered compounds, locale-specific endings and string blends from user-supplied roots.
- Each candidate records its concept, definition, locale, seed roots, formation rule, deterministic ID, generator version, and review status.
- It checks collisions only against the loaded `core-operations.json` package.
- JSON export is explicitly user-initiated. User-entered concept text is not posted to the Word Lab server or saved in local persistence.
- All output remains `needs-review`. The algorithm is deterministic and does **not** call a language model.

Example with Persian roots `شواهد، حافظه، پیوسته`: 24 reproducible candidate proposals were produced, including compound candidates such as `شواهد‌حافظه` and productive-ending candidates such as `شواهد‌محور`. These are candidates, not certified new words.

## 2. Source control and rollback

- Forgejo branch: `release/oersi-word-lab-v0.1.1-20261029`
- Feature commit: `f03962d` — `Add experimental OERSI word proposal lab`
- Pre-deployment backup: `/opt/darkoza/backups/language-forge/word-lab-predeploy-20261029T102006Z`
- Deployment consisted of static-file copies; no application service was restarted and no model weights/configuration were changed.

## 3. Test results

- Word Lab Node.js tests: **9/9 passed**.
- Existing Language Forge Python regression tests: **42/42 passed**.
- JavaScript syntax checks: passed.
- Git whitespace/diff check before feature commit: passed.
- Browser E2E could not run: the installed Chrome binary fails to launch because the host is missing `libatk-1.0.so.0`. No OS package was installed just to overcome this blocker. The UI, actual download action and mobile layout therefore remain unverified in a real browser session.

## 4. Live production verification

The following seven public assets returned HTTP 200, and each downloaded body's SHA-256 matched the corresponding file in `/opt/darkoza/asi-site/language-forge/` at verification time:

- `https://darkoza.com/asi/language-forge/`
- `https://darkoza.com/asi/language-forge/README.md`
- `https://darkoza.com/asi/language-forge/WORD-LAB-REPORT.md`
- `https://darkoza.com/asi/language-forge/word-lab.html`
- `https://darkoza.com/asi/language-forge/word-lab.js`
- `https://darkoza.com/asi/language-forge/word-lab-core.js`
- `https://darkoza.com/asi/language-forge/word-lab.css`

Result: **7/7 exact hash matches**. `https://darkoza.com/asi-api/healthz` returned HTTP 200 with `{"status":"ok","service":"asi-public-adapter"}`. Five observed services remained active: `darkoza-asi-gateway`, `darkoza-autonomy-loop`, `darkoza-cognitive`, `darkoza-oersi-pilot`, and `forgejo-production-runner`; `systemctl --failed` reported zero failed units.

## 5. Model-backed word invention: tried, not accepted

Three distinct inference approaches were attempted against local models; no generated term set was accepted or published:

1. The cognitive gateway's `/v1/chat` imagination route did not return within a 75.22-second client budget. The gateway metrics recorded one failed request. Inspection showed the imagination system prompt forces a fixed six-line English invention format, which is not a compatible contract for Persian lexical proposals.
2. Direct `darkoza-coder:3b-darkoza` inference with a 180-token output budget timed out at 38.04 seconds.
3. A 64-token 3B trial returned HTTP 200 in 19.74 seconds but emitted malformed/truncated numeric placeholders rather than six usable word candidates; rejected as unusable.
4. Direct `qwen2.5-coder:7b-darkoza` inference timed out at 30.03 seconds.

The attempted 3B model was unloaded through Ollama's local API after testing. Post-test checks found no loaded model reported by `/api/ps` on either `127.0.0.1:11434` or `127.0.0.1:11435`, and about 13 GiB RAM available. No model weights or service configuration were changed. This is a model-route/latency/output-quality blocker, not evidence that ASI word invention works.

## 6. Strict claim boundaries

This deployment is a useful, functioning **experimental lexical-candidate tool**. It does not demonstrate superintelligence, human-level creative language invention, semantic accuracy, or global lexical novelty. A local collision check is not a dictionary, Internet, trademark or legal-name search. No candidate is marked `human-reviewed` or `verified`.

The existing Darkoza/OERSI public beta is deployed, but there is no evidence from this task that OERSI is superintelligent. Production public memory and independent answer verification were disabled/false in the release metadata observed before this task; do not imply otherwise.

## 7. Next high-value gates

1. Add a dedicated bounded `word_generation` task contract instead of reusing the incompatible imagination prompt.
2. Diagnose model inference latency with short reproducible requests, actual prompt/completion token counts and per-model time budgets; retain a timeout/cancellation check so no orphaned runner consumes resources.
3. Require valid output parsing and candidate-quality gates; keep model proposals isolated from the official lexicon.
4. Compare model proposals with the deterministic baseline and blinded native-speaker judgments; measure usable-proposal rate, review agreement, latency and resource cost.
5. Run browser E2E after a safe, approved resolution for the missing Chrome dependency.