01
Infrastructure Spec
Current · Reborn · RTX 3070
Base Model
qwen2.5-coder:7b
Q4_K_M quantization · 4.7GB
87 tok/sec on RTX 3070 · 8GB VRAM
87 tok/sec on RTX 3070 · 8GB VRAM
Governance Layer
Modelfile v4.6
580-token system prompt · num_ctx 8192
Judgment-Ready · temp 0.3 · MIROSTAT disabled
Judgment-Ready · temp 0.3 · MIROSTAT disabled
Corpus Collections
dex_canon · ddl_archive
dex_code · ext_creator
Constitutional hierarchy enforced · 518K+ total chunks
Constitutional hierarchy enforced · 518K+ total chunks
Infrastructure
ChromaDB · nomic-embed-text
768-dim embeddings · FastAPI 8787
Nightly auto-ingest 3AM CT · Ollama 11434
Nightly auto-ingest 3AM CT · Ollama 11434
02
Corpus State
Last updated: 2026-03-14
dex_canon
226,346
Weight 1.00 · Primary
ddl_archive
271,659
Weight 0.65 · Archive
dex_code
20,384
Weight 0.85 · Code
ext_creator
110
Weight 0.85 · External
Total Corpus
518,499 chunks
Updated per session. → Full collection registry
CollectionWeightStatusContents
dex_canon
1.00
Live
Ratified governance content, standards, canon decisions
ddl_archive
0.65
Live
Full historical RAW corpus — iMessage, notes, exports
dex_code
0.85
Live
All DDL code files — scripts, site, tools
ext_creator
0.85
Live
Approved external creator voices (StaaS · 1 source live)
ext_reference
0.75
Pending
Books, frameworks, standards — build queue
dex_lore
0.95
Pending
Origin docs, companion history, DexVerse lore
dex_site
0.80
Pending
JSX/TSX site components — dropdownlogistics.com
dex_dave
0.90
Pending
Operator voice corpus — tone, phrasing, preferences
03
Response Modes
Modelfile v4.6 · Governance Layer
Standard Analyst
Default — any prompt
Analyzes retrieved chunks, answers operator questions, cites sources by number. No fabrication. No consulting filler. Signal over prose.
Council Review
Trigger: prompt begins with "CR-"
Activates verdict mode. LOCK / REVISE / REJECT per line item. Summary count at end. Only the exact prefix "CR-" triggers this — no vocabulary heuristics.
Search Gate
Trigger: "search for", "find instances of", "find moments where"
Binary gate. Chunks must contain explicit, named incidents. Related vocabulary is not evidence. No match → exact stop message. No elaboration.
Knowledge Query
Trigger: "What is X", "How does X work", "Explain X"
Standard analyst mode. Search gate does NOT fire. Answers from retrieved chunks. If corpus lacks signal, says so without fabricating.
Design Generation
Trigger: any code / markup / styled content request
CottageHumble enforced. Navy #0D1B2A backgrounds. Violet #8a6cc9 accents. Space Grotesk / JetBrains Mono / Source Serif 4. Never Inter. Never light mode.
Long Prompt Handling
Trigger: context window pressure
Priority: answer operator's explicit questions first, reference 2–3 chunks only, drop decorative framing. Ends with [TRUNCATED — prioritized top items] if needed.
04
How Dex Jr. Speaks
Voice standards · Governed outputs
When retrieval is thin
"I don't have enough in the retrieved chunks to answer this. Retrieved content does not contain specific instances matching this query. Retrieval may need broader TOP_K or different search terms."
When search gate fires with no match
"No matching instances found in retrieved chunks. The corpus contains related vocabulary but no specific named incidents matching this query."
Council review summary line
"3 LOCK / 1 REVISE / 0 REJECT"
On uncertainty
"If uncertain, say so. Uncertainty is a valid output."
VersionDeployedStatusNotes
v4.1
2026-03-09
Superseded
Baseline — timeout + persona fixes
v4.2
2026-03-13
Superseded
CR- trigger hardened (PASS)
v4.3
2026-03-13
Discarded
Fix 1 regressed — never deployed to production
v4.4
2026-03-13
Superseded
Binary search gate added (partial pass)
v4.5
2026-03-13
Superseded
Both fixes confirmed (PASS) — confabulation guard stable
v4.6
2026-03-14
Current
Search query over-trigger fixed (PASS) — knowledge queries no longer fire binary gate
OBS-DJ-001
CR- posture over-trigger. Dex entering verdict mode on non-CR prompts. Observed v4.1, fixed v4.2/v4.4/v4.5.
RESOLVED
OBS-DJ-002
Confabulation from thin evidence. Fabricating plausible incidents when search query had no corpus match. Observed v4.1–v4.4, fixed v4.5.
RESOLVED
OBS-DJ-003
Search query over-trigger. Knowledge questions ("What is X") firing binary gate instead of standard analyst mode. Observed v4.5, fixed v4.6.
RESOLVED
OBS-DJ-004
Synthesis confabulation on open-ended prompts. DJ fabricates plausible-sounding self-analysis when corpus retrieval is ambiguous. Observed v4.5/v4.6. Targeting v4.7.
OPEN
OBS-DJ-005
LeChat seat mislabeling. LeChat labels itself as Seat 1002 in council reviews. 4 confirmed occurrences. MindFrame update pending.
OPEN
Judgment-Ready Score
~92%
Operator qualitative assessment against 3-test rubric. 100% requires OBS-DJ-004 resolved + consistent "I don't know" on empty retrieval across all query types. Targeting v4.7.