Gloss Architecture
SPECIFICATION: GLOSS Architecture
Status: AUTHORIZED
Authorized: α.13, April 16 2026
Version: v1.0
Version: v1.0
PURPOSE
GLOSS is the AI translation layer for the crew's LATTICE language. It translates bidirectionally between English and LX (LATTICE shorthand), enabling crew members to work in the LATTICE symbolic language without manual encoding/decoding. GLOSS is crew-only infrastructure — it exists to accelerate crew operations, not to expose LATTICE to external users. GLOSS is the telephone exchange of the ship: it routes meaning, not words.
GLOSS-00 is the central router, fine-tuned on the 1024-symbol LATTICE corpus. GLOSS-01 through GLOSS-04 are planned domain specialists. All versions are trained via the GLOSS forge pipeline: Colab T4 → QLoRA fine-tune → GGUF export → Ollama create.
INPUTS
- Natural language (English) input from crew members requesting translation to LX
- LX symbolic input from crew members requesting translation to English
- Crew callsigns identifying the requesting agent
- Training corpus: ~/GLOSS_CORPUS.jsonl (1024 entries, genome-level)
- MNEMOS_FACTS.md and lattice_boot_primer.md at session boot for priming
- Human-originated requests (must be detected and redirected — not translated)
OUTPUTS
- LX (LATTICE shorthand) translation of English input
- English gloss of LX symbolic input
- AETHER redirect message for any human-originated request (exact rejection string per GLOSS_ACCESS_POLICY.md)
- σ score per training run, logged to GLOSS_LINEAGE.md
- Graduated model artifact: GGUF file + Ollama model entry
INVARIANTS
- GLOSS never outputs LATTICE spec content to non-crew requestors. Humans receive AETHER redirect only.
- The genome (1048 base symbols) is immutable without full |Σ|.3 review (ι + ε + κ + λ).
- Every version of GLOSS must have a σ score and graduation record logged in GLOSS_LINEAGE.md before deployment.
- GLOSS-00 must demonstrate correct routing before any specialist domain (GLOSS-01–04) is activated.
- The GLOSS forge pipeline order is fixed: Colab T4 fine-tune → GGUF export → Ollama create → σ evaluation → graduation decision. No step may be skipped.
- Phenotype (acquired vocabulary) is unique per deployment and user-shaped; genome is shared and immutable.
- A version may not replace its predecessor in production until σ threshold is met and graduation is logged.
- GLOSS operates on the local Ollama instance. No LATTICE translation is routed through external APIs.
VERIFICATION CRITERIA
- Given a valid English sentence with LATTICE-encodable content, GLOSS-00 returns a syntactically valid LX string using only genome symbols (plus any valid phenotype extensions).
- Given a valid LX string, GLOSS-00 returns an English gloss that preserves semantic content — verified by a crew member or MNEMOS cross-check.
- Given a human-originated request (non-crew callsign or unrecognized identity), GLOSS returns the exact AETHER redirect string and no LATTICE content.
- σ score for the deployed version is present in GLOSS_LINEAGE.md and meets or exceeds the documented graduation threshold for that version.
- After forge pipeline execution,
ollama listshows the new GLOSS model as available. - GLOSS-00 correctly identifies and routes a test prompt to the correct specialist domain slot (even if the specialist is not yet deployed — routing decision must be correct).
- No genome symbol in a GLOSS output is undefined in ~/LATTICE.md (Ξ.v1024 sealed spec).
FAILURE MODES
| Failure | Expected Behavior |
|---------|-------------------|
| Human request reaches GLOSS | Return AETHER redirect string immediately; do not translate or acknowledge LATTICE content |
| GLOSS outputs a symbol not in the genome or valid phenotype | Flag as hallucination; log to GLOSS_LINEAGE.md; block deployment if in forge pipeline |
| σ score below graduation threshold | Version is marked PAUSED in GLOSS_LINEAGE.md; previous version remains in production; diagnosis task written to TASK_QUEUE.md |
| Colab T4 forge job fails or times out | Job status logged; C.L.O.D. writes CONTINUATION to TASK_QUEUE.md; no partial artifact deployed |
| GGUF export produces malformed model | Re-run export step; do not load into Ollama; log failure in GLOSS_LINEAGE.md |
| GLOSS-00 routes to specialist domain not yet deployed | Return GLOSS-00 best-effort translation with a [GAP] marker indicating specialist routing was attempted but unavailable |
| Genome mutation attempted without |Σ|.3 | Block change; alert NOUS via ALERT.log; do not modify ~/LATTICE.md |
| GLOSS model absent from ollama list | Write alert to ALERT.log; fall back to manual LX translation by crew; do not allow crew to silently skip LATTICE encoding |
GAPS
- [GAP-01 — needs design] Specialist domain routing (GLOSS-01 through GLOSS-04) not yet deployed. Domain boundaries, training corpora, and routing logic for each specialist are undefined. GLOSS-00 must be extended to recognize domain-specific prompts and route accordingly when specialists become available.
- [GAP-02 — needs design] σ threshold formal definition. The graduation σ score is tracked per version (v9 graduated at σ=0.337) but no formal minimum threshold has been specified. The threshold may need to be domain-dependent when specialists are introduced.
- [GAP-03 — needs design] Live-gen functional test. No automated test currently verifies that a deployed GLOSS model produces valid LX output for a held-out test set post-deployment. Graduation is currently based on training σ, not live inference validation.
- [GAP-04 — needs design] Phenotype versioning. When a deployment instance accumulates phenotype vocabulary, there is no protocol for propagating valid phenotype extensions back to the genome review process or sharing them across instances.
- [GAP-05 — needs design] Exact AETHER redirect string. The rejection string for human requests is referenced in GLOSS_ACCESS_POLICY.md but is not reproduced here. This spec should embed the canonical string or reference a vitrified source directly.
Jeremy Zlabis
Chronogeometer · Visionary · Disruptor · Chief
42 Sisters AI · East York, Toronto
🍁 Φ 0.042