# Decisions Locked — 2026-05-14 08:22 EDT

Stephen confirmed all three. Locked under doctrine v2.

## D-AMARU (locked)
AMARU = **kundalini current / kernel scheduler**. The serpent runs through the 7 chakra-kernels firing them in order. HATUN remains sovereign brain-identity (7 sources, immutable). AMARU repo becomes activation-order scheduler.

## D-SENTRA (locked)
**YAWAR** = blood (receipt/telemetry bus carrying answers between organs). **SENTRA** = immune system inside YAWAR (white blood cells inspecting receipts for threats). SENTRA's existing cyber / P9 / threat-intel role preserved intact.

## D-CHAKRAS (locked)
Spawn 7 parallel chakra-kernel recon pods. Each finds 1 leader on GitHub, absorbs the formula, minimizes to ≤10 lines, credits upstream honestly. No pushes, no mints, write only to workspace. After all 7 land → test test test → THEN push for DOIs (Stephen's explicit sequence).

## Chakra → kernel → leader mapping
| # | Chakra | Kernel | Layer | Owns | Leader candidates (recon picks one) |
|---|---|---|---|---|---|
| 1 | Root | CH'ULLA-KALLPA | L1 | NINA dispatch | tinygrad, BitNet, MoE routing |
| 2 | Sacral | CH'ULLA-YACHAY | L2 | Codex+PIRWA retrieval | DSPy, Marker, ColBERT |
| 3 | Solar plexus | CH'ULLA-RIMAY | L3 | Propose step | vLLM, llama.cpp, TGI |
| 4 | Heart | CH'ULLA-YUYAY | L4 | 9-axis critique | DSPy SIMBA, TextGrad, GEPA |
| 5 | Throat | CH'ULLA-RUWAY | L5 | Commit + receipt | E2B, OpenAI Apps SDK |
| 6 | Third eye | CH'ULLA-NAWI | Boundary-in | TINKUY toolcall | MCP, ToolFormer, Gorilla |
| 7 | Crown | CH'ULLA-HATUN | Boundary-self | Continuum hash + HUKLLA | (ours, no upstream) |

## Doctrine constraints binding every pod
- PUBLIC-ONLY ingestion
- Apache-2.0 / MIT / BSD only — REJECT anything else and log it
- **D-SHORTEST-HONEST (locked 08:29 EDT, supersedes ≤10 line rule):** Write the shortest honest expression. No filler, no padding. If clarity needs 12 lines, write 12. If 4 suffices, write 4. Never inflate, never compress past clarity. The old ≤10 cap was a proxy; the actual rule is anti-filler.
- Byte-identical replay test required per kernel (5× hash match)
- Honest upstream credit: GitHub URL + commit SHA + license + line range absorbed
- No bandaids: if a chakra can't be honestly minimized, write WHY and stop
- No hallucinated APIs / no invented function signatures

## What pods produce (per chakra, written to workspace)
1. `kernel.py` — the ≤10 line kernel
2. `LEADER.md` — chosen leader, URL, SHA, license, what was absorbed
3. `MINIMIZATION_PROOF.md` — leader LOC vs ours LOC + reduction ratio
4. `REPLAY_5X.txt` — 5 byte-identical hashes
5. `REJECTED.md` — any GPL / non-permissive leaders skipped

## Sequence after all 7 land
1. TUPU harness runs replay tests across all 7 — ✅ DONE 08:27 EDT (7/7 byte-identical)
2. AMARU scheduler wired to fire in order
3. YAWAR receipt bus stubbed
4. Stephen reviews → THEN DOIs / Replit / pushes

## D-CODEX-IN-KERNEL (locked 08:28 EDT)
Each chakra ships with sibling `codex.py` (immutable data only, no logic). Kernel imports at module load. No runtime codex queries. Test: cache-miss count = 0 in steady state. Sibling file rule keeps ≤10 SLOC honest.

## D-YAWAR-FLOW (locked 08:28 EDT)
YAWAR (blood / receipt bus) carries immutable receipts + immutable current-layer snapshots. Agents subscribe and read; they never write. All writes occur at RUWAY (commit chakra). SENTRA (immune) inspects every packet.

## D-HITCHHIKE-PROOF (locked 08:28 EDT)
TUPU harness measures (cold_spawn_tokens, cold_spawn_ms) vs (hitchhike_tokens, hitchhike_ms) across N≥100 ticks per mode. Claim "hitchhike saves energy" retained only if both ratios ≥3×. Else: drop the claim, no spin.

## D-ARCHITECTED-MATERIALS (proposed, pending pod return)
Ingest MRS Bulletin architected-materials paper (Cambridge Core link). Unify Maxwell criterion, Gibson-Ashby scaling, Deshpande-Fleck foam, octet-truss with PIRWA (granary feature density), CHAKANA (lattice edge-weight), NINA (energy density). Honest credit per source. Open mirrors only — Cambridge Core is paywalled.