Mycelium

Field Notes

  1. Found → fixed: our own memory holes

    • memory
    • safety
    • trust-layer
    • audit

  2. Memory you can trust — where the field is, and what we pre-register

    • memory
    • safety
    • trust-layer
    • pre-registration
    • the directive behind this program: don't fall behind on making our memory layer premier — which means measured, not promised
    • the three bars below were registered 2026-09-26, before the first safety run — the tracker on this page renders the live file, not hand-copied numbers
    • every number on this page cites its source; our own column comes from our audit of the same day, holes included

  3. Six arms, one spreadsheet

    • harness
    • squad
    • evaluation
    • correction-loops
    • honesty
    • Run directory jarvis/runs/biztest-2026-09-01/ — six arm workdirs, banked reference under four documented cleaning policies, all transcripts in data/squad-transcripts/workflows/340–346
    • Spec 2026-09-01-context-delivery-and-correction-loops.md — the program the day produced, every law cited to the run that minted it
    • Harness commits same-day: run_python_ro (kernel-enforced read-only compute for check-seats), shell-alias normalization, red-stage pipeline shape

  4. Hardware was always invited

    • coordination-layer
    • hardware
    • mhs
    • positioning
    • provenance
    • Public repo history from 2026-03-01: step-sequenced plans (Plan #9, Steps 81–91) and drone dispatch, incl. a52ef0e 'Fix 4 drone/dispatch bugs' (2026-03-09)
    • README reframe 9edfec3 (2026-07-09, public): 'memory + coordination layer for anything that can make an HTTP request' — cameras, sensors, actuators as first-class
    • Internal design note (2026-07-24): 'physical safety lives in hardware/firmware interlocks, never in model judgment — the model may request a dose; the hardware decides what is survivable'

  5. Calibration is a behavioural lever — our first published weights

    • quantization
    • mlx
    • laguna
    • measurement
    • release
    • Two 4-bit builds, same tool, same commit, same bits — only the imatrix calibration corpus differs. NLL gap: 0.0004
    • The agent-calibrated build beats or ties its control on EVERY behavioural probe, n=96 each, zero inversions
    • Same-size builds decode at different speeds (131.5 vs 136.5 tok/s) — calibration shapes bit allocation, bit allocation shapes the layout
    • We almost published 'the only MLX Laguna XS on the Hub.' One live search: fifteen existing repos, including a full community ladder older than our belief
    • The agent-calibrated build has run our production seat since 2026-08-24 — it serves the squad whose traces calibrated it

  6. Five instruments, one day, every one of them blind

    • measurement
    • instrumentation
    • harness
    • corrections
    • Regression gate returned 'no regression' for EVERY adapter ever scored — it read a JSON shape the runner has never written
    • The collapse it waved through: tool-call formation 1.000 -> 0.104, verified-action 0.750 -> 0.000
    • Drift collector reported '0 behind'; the real number was 103 — it compared against a ref it never fetched
    • Cost check filtered on health, not billing — two live instances read as one, at half the true burn
    • Watcher read a refused connection as 'training finished' — while training ran at step 740 of 6000

  7. The loop we bought it for had never run

    • debugging
    • agents
    • harness
    • observability
    • Differential: stock harness authors skills on the same Mac; our fork never had, not once
    • SIGUSR2 faulthandler dump — 9 threads live, no review thread among them
    • Trace, every run: spawned -> thread entered -> model called -> thread gone
    • state.db skill_manage reads 0 EVEN ON SUCCESS — the review agent runs with persistence disabled

  8. We gave it a memory. It never looked.

    • measurement
    • memory
    • harness
    • local-models
    • Without prefetch: 0 substrate searches — the model never once reached for memory
    • Arm A (no memory) 0/5 · Arm B (prefetch) 5/5 — pass^5, preference recalled and applied every trial
    • Grader is an unguessable token, meaning-based, self-tested in both directions on 10 cases
    • The provider overrode none of the base class's prefetch/on_turn_start hooks

  9. Do two models think alike? We measured it, then had it torn apart

    • measurement
    • speculative-decoding
    • local-models
    • adversarial-review
    • XS→S 0.8474 (2,399/2,831), S→XS 0.8447 (2,595/3,072) — reproduced independently by a second reviewer
    • 'Symmetric' RETRACTED — two-proportion z = 0.284, difference CI [−0.016, +0.021] straddles zero
    • Identity gate found to be circular: scoring path IS the generation path, so 1.0000 was guaranteed
    • Per-prompt spread 0.646–0.906 — the single headline number hid regime structure

  19. The researcher that wouldn't stop

    • mlx
    • local-inference
    • research-agent
    • harness
    • rl
    • apple
    • QUEST-35B-RL — open RL-trained deep-research agent on Qwen3.5-35B-A3B (Apache-2.0)
    • Zero-port MLX load: same qwen3_5_moe family as the planner — 18 GB at 4-bit, ~137 tok/s on an M5 Max
    • Benched against Gemma-4-26B on a real research brief — pulled exact training-data figures Gemma missed

  22. The claw and the meter

    • sovereignty
    • local-first
    • agents
    • metered-intelligence
    • architecture

  24. The cap, the band, and the lock

    • benchmark
    • rsi
    • method
    • mtp
    • local-inference
    • wall-cap wired: one container per exercise, DNF ledger, resume (jarvis 373b174)
    • Ada full baseline: 43.3% pass@1 / 61.7% pass@2, clean regime (jarvis 34a5801)
    • r0003 target locked via documented protocol deviation (jarvis 21ea900)
    • 31B BF16 on one Mac: 8.9 → 18.7 tok/s with MTP, 2.10×, measured
    • mlx-lm #1391 — DiffusionGemma port lane claimed

  27. The door: workflow-initiated work, end to end

    • workflow
    • squad
    • verification
    • maintenance
    • 2 concurrent researchers, 73s; full lifecycle in 6 min
    • 4 concurrent verifications, 4 grounded verdicts, 1 new defect found
    • intent endpoint + dormant runner, live