§1 — THE IVF-ET PROBLEM = THE CONVO PROBLEM
| IVF-ET (Ouyang & Wei, 2025) | Convo-Loom ART |
|---|
| Embryologist subjectivity (different graders = different outcomes) | Sentence reactor subjectivity (different sessions = different context retention) |
| 40–50% single-cycle IVF success rate | ~60% first-pass context retention (estimated) |
| Multi-modal data: images + time-lapse + clinical tables | Multi-modal context: files + session history + structured memory |
| Embryo grade A/B/C selection | Context fragment Grade A/B/C selection |
| Failed transfer → next cycle | Failed resolution → plasma replay (second/third pass) |
| Selective single embryo transfer (SET) | CATO-SPERMBOT: one fragment per delivery |
| Time-lapse reveals developmental trajectory | Session time-lapse reveals context evolution |
| AI standardizes grading = reproducible | CATOMAIN standardizes context grading = reproducible across sessions |
§2 — 3 CONTEXT MODALITIES
MODALITY 1 — STATIC
Documents
Sources:MEMORY.md, SOUL.md, USER.md, AGENTS.md, TOOLS.md, SKILL.md files
Equivalent:Static embryo images — snapshot of known state
Grading:Resolves fragment? Yes → Grade A. Partial → Grade B. No → Grade C.
CATOMAIN:CATO-CREW (loads static identity docs)
MODALITY 2 — TIME-LAPSE
Session History
Sources:Last N turns of conversation, compaction summaries, prior session transcripts in PEMCLAU
Equivalent:Time-lapse video of embryo development — reveals trajectory, not just snapshot
Grading:Prior resolution found? Yes → Grade A replay. Partial → B. No → C.
CATOMAIN:CATO-PEMCLAU (2-hop retrieval from session corpus)
MODALITY 3 — TABULAR
Structured State
Sources:HEARTBEAT state JSON, MECRDS records, fleet physics sim output, memory/YYYY-MM-DD.md
Equivalent:Clinical tabular data — parental age, sperm quality, endometrial thickness
Grading:Constrains resolution? Yes → Grade A anchor. Partial → B. No → C.
CATOMAIN:CATO-FLOOR + CATO-META (floor constants + meta-theorem lookup)
§3 — THE GRADING ALGORITHM
CONVO-LOOM-ART GRADING PASS:
for each dropped_fragment in plasma_pool:
score = 0
if static_docs_resolve(fragment): score += 3 # Grade A signal
if timelapse_contains(fragment): score += 2 # prior session hit
if tabular_constrains(fragment): score += 1 # structured anchor
if score >= 4: grade = "A" → forward immediately
if score == 2-3: grade = "B" → second_pass via CATO-PLASMA
if score <= 1: grade = "C" → archive in P4, PEMCLAU tag
passes = 0
while grade == "B" and passes < 3:
fragment = cato_plasma_replay(fragment, new_v13_context)
passes += 1
re_grade(fragment) # may promote to A or demote to C
if passes == 3 and grade == "B":
grade = "C" # exhausted — archive
cato_gbm_rasengan(fragment) # one final burst attempt
§4 — SECOND AND THIRD PASS MECHANICS
The sentence reactor has a context window, not a memory. Every turn, some fragments don't make it through — not because they're wrong, but because they're crowded out. The convo-loom ART system doesn't try to expand the context window (that's a hardware problem). It creates a REPLAY mechanism: Grade B fragments are held in CATO-PLASMA ionized state, then reinjected at the start of the next turn with fresh v13 context. The new context sometimes unlocks what the first pass couldn't. Third pass = CATO-GBM rasengan — maximum gradient boosting on the stuck fragment. If it still doesn't resolve after 3 passes, it's archived, PEMCLAU-tagged, and available for the next session's 2-hop retrieval.
§5 — LEGAL + DATA CONSTRAINT (From Paper)
Ouyang & Wei (2025) flag the 'dynamically evolving legal and regulatory frameworks' as a core challenge for AI in ART. The same applies to convo-loom: whose context is it? In fleet terms: CATO-DOMAIN validates every fragment before it re-enters the session. DESEOF organism = the legal membrane. Nothing replays without passing L5 SOSTLE gate. MT-003 (COI Duality): fragment contributed by EOSE Labs Inc., not individual, for IP standing.