2026-07-22 PASS
Radical simplification
The entire prior research program was archived immutably; the active branch was reduced to one governed vertical.
Receipt · archive/pre-simplification-2026-07-22
The receipts culture
We publish our failures
Most AI marketing shows you the demos that worked. The MARS program publishes every gate — passed, rejected, or pending — as SHA-256-bound JSON receipts in the open repository.
Claims need custody chains
A capability claim without an executable receipt is an anecdote. Every MARS claim binds an exact artifact hash, a sealed evaluation suite, and a deterministic decision — so anyone can check whether the claim still holds.
Rejections are preserved as durable negative evidence. They are how the system earns the right to be believed when it says PASS.
Promotion standard
The bar an adapter must clear
Training loss is not evidence. Promotion requires every gate below — and if a correct channel does not beat its shuffled and masked controls, that channel has not been shown to be causal.
▸ correct-channel typed accuracy >= 95% per channel
▸ masked abstention >= 99% per channel
▸ shuffled conflict behavior >= 95% per channel
▸ coordination >= 90%
▸ non-finite adapter values = 0
▸ natural grounded + multimodal controls
Receipt timeline
The program, as it actually happened
-
- 2026-07-24 ADMITTED
Multichannel corpus admitted
73,248 rows across 265 source units, zero train/held-out overlap, complete correct/shuffled/masked control groups.
Receipt · mars-hyphae-multichannel-datasheet
- 2026-07-26 REJECTED
Natural generalization gap — published
The multichannel adapter scored 4/6 on natural controls: schema drift and a missed conflict retry. Not promoted. Remediation defined.
Receipt · mars-hyphae-multichannel-natural-evaluation
- 2026-07-27 PROMOTED
Nine-channel component promoted
100% typed accuracy on every channel, 100% shuffle-following, 491/491 exact parent preservation — promoted as a governed component with mandatory authority override.
Receipt · mars-hyphae-full-promotion-decision
- 2026-07-28 PASS
Consumer hardware gate passed
MARS-1 Q4 measured on an RTX 5050 laptop GPU: 50.2 tok/s generation at 3,098 MiB peak VRAM of an 8,151 MiB budget.
Receipt · mars1-q4-rtx5050-benchmark
- 2026-07-28 REJECTED
Q4 full-body promotion rejected
Source-pointer accuracy 86.1% against a ≥99% gate. Q4 remains an experimental parent; the gate was not weakened.
Receipt · mars1-q4-promotion-decision
- 2026-07-28 REJECTED
Physical experts smoke rejected
Causal imbalance detected: shuffle-following at 0.2%. The scaling setup was rejected and redesigned — publicly.
Receipt · mars-gemma-physical-nine-channel-smoke
- 2026-07-28 ADMITTED
MARS-T bounded slice admitted
A tiny ternary {-1,0,+1} policy head passed its bounded bilingual gates. The receipt explicitly bounds the claim: capability slice only.
Receipt · mars-t-bounded-student
- 2026-07-28 PASS
Governed local API green
The loopback-only backend passed all 20 runtime test suites: ingest → ask → receipt → remove, schema drift fail-closed, origin rejection.
Receipt · hivemind-runtime test suites 20/20
Receipt anatomy
What a receipt binds
Receipts are small, versioned JSON documents. Four bindings make them auditable:
01
schema — a versioned identifier, so receipts are machine-checkable forever
02
status — PASS, REJECTED, ADMITTED… decisions are explicit, never implied
03
artifact SHA-256 — the exact model / dataset / suite the claim is about
04
decision — what changes because of this receipt, and what remains blocked
Intelligence you can audit.
MARS, HiveMind and Hyphae are being built in the open by Celiums Solutions — local-first, fail-closed, and receipted end to end.