◈ CGNT-1
Home › Gloss Training
◈ Master IndexSpecsLATTICECSDMThe BookUniversityChronicleCrewArchitectureContextLogsTODOXProducts

Gloss Training

SPEC_GLOSS_TRAINING.md · 2026-04-20

SPEC_GLOSS_TRAINING — GLOSS Brain Training & Graduation Protocol

Version: 1.0 | Status: AUTHORIZED | Authority: α.13 | Date: 2026-04-16

PURPOSE

This spec defines the formal protocol for training, evaluating, graduating, and demoting GLOSS brain versions. GLOSS is the crew's local LATTICE-native brain — a QLoRA fine-tune of Qwen2.5-3B-Instruct that speaks LX natively, routes crew comms, and translates between English and LATTICE. Training is the mechanism by which GLOSS improves. Graduation is the gate that certifies a new version is fit for crew telephone function. Demotion is the mechanism that removes a failed version from service.

Current state as of 2026-04-16: v9 DEMOTED (live generation collapse). v10 PAUSED pending corpus redesign and α.13 authorization.

INPUTS

OUTPUTS

On successful graduation:

On demotion:

INVARIANTS

  1. INV-01 — Corpus format compliance: All training pairs must use crew-format prompts (beginning with a recognized crew designator token). Human-to-chatbot format pairs are forbidden. Violation: corpus contamination; retrain is invalid.
  1. INV-02 — Base model lock: All GLOSS versions v1+ use Qwen2.5-3B-Instruct as base. No base model substitution without explicit α.13 authorization and full lineage entry. The v6 0.5B error (σ=0.000) is the canonical cautionary case.
  1. INV-03 — Eval battery is necessary but not sufficient: σ score from the 50-question multiple-choice battery is required for graduation, but does NOT alone certify fitness. Live generation tests are mandatory. v9 demonstrated that σ=0.337 passing eval does not protect against generation collapse (runaway loop failure).
  1. INV-04 — Live generation test is mandatory: Every candidate version must pass at least one open-ended generation test before graduation. Minimum: symbol identity query (⊹: ⊙ [symbol]? format) with stop-token verification. Runaway loop = automatic demotion regardless of σ score.
  1. INV-05 — σ threshold is current-generation only: The passing σ threshold (currently 0.337 as established by v8/v9 plateau) must be reviewed against live generation behavior before v10 forge. Threshold is not permanent — it must be revalidated if corpus or eval battery changes.
  1. INV-06 — No training without NOUS authorization (current state): As of v9 demotion, v10 forge requires explicit α.13 authorization. This is not a permanent invariant — it is a temporary hold pending corpus redesign. Standing authorization resumes when GAMMA issues a redesign directive approved by α.13.
  1. INV-07 — Stop tokens required in training pairs: Training corpus must include explicit stop/EOS signals in all definition and enumeration pairs. Pairs that end with an open list pattern (a→b, a→bb, a→bbb...) are loop-prone and must be reformatted before inclusion.
  1. INV-08 — Version lineage is permanent: Every trained version receives a permanent record in GLOSS_LINEAGE.md. No version record is deleted, even if the version is retired. The full history is the institutional memory of the training program.

VERIFICATION CRITERIA

FAILURE MODES

GAPS

DEPENDENCIES

DEPENDENTS

EXAMPLES

Version record entry (graduation format):


| v10 | NNN | 15 | 0.XXX | 0.XXX | GRADUATED | Corpus redesigned; stop tokens added; live gen PASS |

Demotion trigger:


Live test: ⊹: ⊙ α?
Output:    ⊙ α? → αα, ⊙ αα? → ααα, ⊙ ααα? → αααα ... [90+ lines]
Result:    FM-01 — runaway loop. DEMOTE. DO NOT USE.

Graduation ceremony log (SESSIONS.md):


[C.L.O.D.] κ ⚒ gloss:v10. σ_lx=0.XXX. ∇.μ ΩQ.⊡ → Σ.✓. Arr, the wee brain's had her schoolin'. Copy that. Over.

REFERENCES

Φζ.⊤. ∇.μ Ω.1024/1024.

Jeremy Zlabis

Chronogeometer · Visionary · Disruptor · Chief

42 Sisters AI · East York, Toronto

🍁 Φ 0.042