MAX3 · Session Work Summary

Three deliverables, one session

GitHub-tracking retirement · Brain connectionist addendum · Tech-stack evaluation

2026-06-14 PRs #767 · #768 · #769 all CI-green

00

What this session delivered

This document collects the three pieces of work completed in one session on 2026-06-14, each shipped as its own pull request against the MAX3 repo. The first is an infrastructure change that retires GitHub Issues and the GitHub Project from all automation; the second folds the last 24 hours of brain work into the field guide with two new diagrams; the third is a full evaluation of the technology stack against open-source, free, and paid alternatives for the current hardware.

Note on landing. All three are CI-green and queued to merge; the final mark-ready + squash-merge step was waiting on GitHub's GraphQL rate-limit window to reset (the very API pressure this session's #767 work removes from the headless tasks). A 10-minute check-in is armed to land them.

01

Retiring GitHub Issues + Project from all automation (PR #767)

Two directives drove this. First: “we’re not using github issues or project anymore — just the repo and functionality; the local wall should be the only thing being updated for my tracking every two minutes … setup, pulls, check-in, merge agents, gates should not touch the github api for project or issues lists.” Second: “the every two minute scheduled tasks are still popping up visually — I want them to run without seeing them.”

WHAT CHANGED

■ board_sync_tick.py is now local-only. The GitHub Project sweep (github_tick, the backfill_project_fields call, the sentinel + min-interval machinery) is removed; it now only reconciles merged items from git log and renders the local wall — zero GitHub API.

■ board-audit.bat retired to a stub. The weekly Project correction pass no longer runs; the scheduled task is deleted by the one-time script.

■ backend_watchdog.py drops the claude-fix issue relay. The watchdog still self-relaunches the backend locally — it just no longer files a GitHub issue on relaunch.

■ The merge gate stops firing github-projects-maintainer. The only post-merge propagator is the doc sweep; board state rides in-PR as a fragment + the local git-log reconcile.

■ CLAUDE.md standing rules flipped. A new top-level rule makes the local wall the single tracker; the old “GitHub Project is the source of truth” section is marked superseded.

RUN HIDDEN (NO POP-UP)

■ tools/watchdog-hidden.vbs (new). Runs the watchdog via wscript window-style 0 — no console window. The watchdog was launching python.exe directly, which is what flashed up every two minutes.

■ One-time script. scripts/one-time/2026-06-14-retire-board-audit-and-hide-tasks.py deletes MAX3-board-audit and re-registers MAX3-backend-watchdog + MAX3-auto-pull hidden every 2 min. Applies once per machine via run_pending.

YOUR ONE MANUAL STEP

After PR #767 merges, run a single pull-latest. run_pending chains the one-time script — it deletes the board-audit task and switches the two 2-minute tasks to hidden. Nothing else.

02

Inside MAX’s Brain — the connectionist addendum (PR #768)

The field guide Inside MAX’s Brain 2.0 was cut on 2026-06-12. In the 24 hours after, the brain gained its biggest conceptual addition yet: a design that re-reads the whole graph brain through Geoffrey Hinton’s connectionist lens, plus the first working slice of code that proves it — a small Forward-Forward network that learns what a week is by learning the edges between its days. Section 11 was added to the document; the two figures below are the new diagrams.

WHAT WAS

Before, “tuesday” and “wednesday” were tokens joined by hand-named edges. The relations were real, but the concept of a week — an ordered, wrapping cycle the days belong to — was nowhere in the linguistics. The graph knew facts; it did not learn features.

WHAT CHANGED

BRAIN_NEURAL_REFRAME.md re-frames every node as a neuron and every edge as a synapse, with four commitments: meaning is relational (a word’s meaning is its pattern of edges), prediction is the engine (features of one word predict the next — inference and learning in one act), truth is an input, not a veto (perception is a probability to ground against provenance), and language is one medium among many (the concept is the unit). The first code slice (backend/max3/brain/) is a numpy-only Forward-Forward net that recovers the week as a wrapping cycle and learns Hinton’s kinship tree, discovering that “grandfather” is a composed traversal rather than a stored fact.

WHY

Meaning is a word’s relationships, and a relational task — give me a sentence and I predict the last word — forces a system to learn features instead of memorising rules, searching a continuous space of connection strengths rather than a discrete space of symbols. Forward-Forward is the right learner because it models the mean of words locally, without a brittle global pass. The longer arc is meta-learning — MAX learning how to write the code that learns — fenced behind the reversibility and self-improvement rules, and a teaching arc: a brain that grows past its maker and teaches back, through evolution.

03

Tech-stack evaluation vs OSS / free / paid (PR #769)

Evaluated for the actual box: RTX 3090 Ti (24 GB VRAM), 64 GB RAM, 8 cores / 16 threads, local-first behind OPNsense. Every option is filtered through four binding constraints — local-first, open-source dependencies, a single core process, and no heavyweight agent frameworks — so an alternative that wins a benchmark but breaks one of those is rejected regardless.

Headline verdict: keep ~90 % of the stack. Pilot three low-effort model swaps, plan one consolidation, reject the rule-breakers, and hold on new GPU hardware.

THE THREE PILOTS WORTH DOING

■ TTS — pilot Kokoro-82M (Apache). Far more natural than Piper, still local + permissively licensed; keep Piper’s six-axis DSP persona chain regardless.

■ Vision — pilot RT-DETR (Apache). Real-time replacement for DETR with the same no-AGPL posture; plays nicer with the shared-GPU voice budget.

■ Embeddings — pilot bge-m3 / mxbai vs nomic. Better retrieval under GraphRAG + memory; cost is one background re-embed pass.

PLAN ONE CONSOLIDATION

■ Vector store — converge sqlite-vec → duckdb-vss when its WAL recovery is proven (two engines collapse into one).

THE SCOREBOARD (SELECTED LAYERS)

EXPLICITLY REJECTED (AND WHY)

■ vLLM / TGI / NIM — optimise multi-user concurrency we don’t have, at the cost of the single-process VRAM headroom Whisper + ArcFace + TTS need to coexist.

■ Ultralytics YOLO — AGPL-3.0; forbidden by the OSS-deps rule.

■ XTTS / cloud TTS / cloud STT / cloud vector — send audio, frames, or embeddings off-box; break local-first.

■ LM Studio / Neo4j — closed-source / licence + server shape that don’t fit a single-box local-first resident.

HARDWARE: HOLD

24 GB VRAM is enough for the design as specified — the only thing it can’t do is run a 70B locally at interactive speed, which the cloud fallback already covers for the rare hard turn. The cheapest real latency win is a fast NVMe for the graph + model files, not a new GPU. RAM at 64 GB is ample. The network posture (OPNsense, local-first, LAN opt-in) is exactly right.

04

References

Pull requests. #767 (GitHub-tracking retirement + hidden tasks), #768 (brain §11 addendum), #769 (tech-stack evaluation).

Design + decision docs. BRAIN_NEURAL_REFRAME.md, TECH_STACK_EVALUATION.md, STT_SWAP_DECISION.md, SPEAKER_ID_SWAP_DECISION.md, DUCKDB_MIGRATION.md, VECTOR_STORAGE_HYBRID.md, VISION_CONTENTION_DESIGN.md, and the binding hard rules in CLAUDE.md.

Brain code + grounding. backend/max3/brain/ (relational_net.py, proving_data.py) shipped across PRs #761–#766; Hinton’s 1986 distributed-representations / family-tree work and the 2022 Forward-Forward Algorithm.