Three filed patents, a 25-paper working-paper series on the substrate architecture, and a five-paper preprint series on interpretable MoE routing, RoPE schedules, training curricula, activation-space intervention, and decoding strategy — documenting what this platform is built on, what's measured, and what isn't yet.
Each preprint is self-contained with abstract, math, related work, proposed evaluation, and honest limitations. The five papers are individually contributory and collectively form a stack for interpretable, auditable, and grounded language models. This is the public evidence for what an engagement through Work With AsAManThinks actually draws on.
TERA
Provisional 64/040,504
4-axis lexical/state decomposition — Truth, Emotion, Reasoning, Action read simultaneously from a single interaction, rather than a single sentiment scalar.
Breath Cycle
Provisional 64/040,509
Convergence/expansion phase classification for interaction state — a cyclical model instead of a flat trend line.
Sign-Inversion + Multiplicative RAG
Provisional 64/040,513
Detects when stated content and underlying signal point in opposite directions, and routes retrieval multiplicatively rather than by simple concatenation.
Six mechanisms from a longer internal survey, chosen because each has a specific, checkable technical claim — not because they sound impressive. Status is reported honestly: some are measured against real data, one is disclosed as validated on synthetic data only. That distinction is the point of publishing this at all.
maiiam-alchemist/packages/harmonic-engine — coverage/odu_coverage_map.py, decoding/geodesic_steer.py
Problem: Long-form LLM generation collapsing into repetition or narrow modes, with no zero-parameter way to detect and correct it live.
Mechanism: Quantizes each generated chunk into one of 256 discrete semantic cells, tracks occupancy and streaming entropy, and — when a cell over-fills — computes the exact closed-form Fisher-Rao geodesic distance to the nearest under-visited cell, turning that direction into a direct logit bias. Intervenes at the logits level, not via a text instruction or a frequency-penalty heuristic.
Status: Live and tested.
packages/voxel-runtime/src/shaders — aamt_gemv_mr*.wgsl, fused_gate_up_gemm.wgsl
Problem: Sub-8-bit quantized inference in-browser needs custom kernels; nearly all existing mixed-bitwidth kernel work targets CUDA, not WebGPU.
Mechanism: Branchless inline unpacking of int4/ternary/1-bit weights with no dequant-to-dense step, fused gate+up GEMV in one dispatch with per-row boundary resolution inside a shared tile, and runtime capability-probed dual-module compilation so unsupported WebGPU adapters degrade gracefully instead of failing the whole module.
Status: Measured: ~32× fewer redundant activation reads, 2.5–3.8× kernel speedup, numerically verified to 7e-7 parity against the reference implementation.
maiiam-alchemist/scripts/patch_aamt_chat_contract.py
Problem: Metadata-only edits to a multi-gigabyte model file (e.g. a chat-template fix) normally require rewriting the entire file.
Mechanism: A [blobs][manifest JSON][32-byte footer] container layout lets a metadata-only edit rewrite only the trailing manifest region — O(manifest size), not O(file size) — with SHA-256 weight-region verification before and after the patch.
Status: Validated against a real 3.2GB production checkpoint: ~4 seconds, idempotent on repeat runs.
A canonical chat contract embedded in the model manifest, checked identically across three independently-implemented runtimes
Problem: The same model served from three different runtimes (TS/WebGPU, Python/MLX, Unity C#) silently drifting out of agreement on how a conversation should be formatted.
Mechanism: One contract, machine-verified across all three runtimes rather than trusted by convention.
Status: Already caught two real production bugs before this was even a formal survey — Unity had zero check in place; the MLX sidecar was not chat-formatting at all.
packages/voxel-runtime/host/echo-refocus-prompt-opt.ts
Problem: Correcting apparent drift in a session's interaction trajectory without assuming every session actually has coherent, correctable drift.
Mechanism: Classifies a session's 4-D trajectory (coherent / incoherent / alternating) via a corrected single-trajectory permutation test, and only applies a point-inversion correction — reprojected through a closed-form log-Euclidean manifold projection — when the session is classified coherent, because naive mean-drift subtraction is proven elsewhere to fail on roughly 11.7% of inputs.
Status: Validated on synthetic data only — disclosed candidly: the source paper's real-session test has not yet found a coherent-drift example to confirm on live data.
apps/api/services/reasoning_ledger.py
Problem: Building a longitudinal signal for reinforcement learning without storing raw conversation content or exposing small-cohort user data.
Mechanism: Stores structural metadata only — never raw content, just a truncated SHA-256 hash for deduplication. RL candidacy is flagged from live polarity-classifier confidence, and the reward label is derived from implicit next-turn behavior (a rephrase reads as negative signal, an affirmation reads as positive). The public aggregate endpoint buckets by tenure and suppresses any cohort below a minimum distinct-user count, returning proportions only.
Status: Live in production.
This is a curated excerpt of a longer internal novelty survey. An engineering-level scan is not legal advice or a prior-art search — treat these as worth an attorney's time, not guaranteed grants.
Separate from the five-paper preprint series above: a 25-paper working-paper series documenting the substrate architecture underneath this platform — from generation control, through geometric addressing and provenance, through memory consolidation, to a synthesis of how bounded agent states can be created and merged. Each paper states plainly what is measured, what is proven by construction, and what remains unimplemented — several editions of the Self-Mirrored newsletter are public readings of specific papers in this series.
Generation & platform substrate
Agent reasoning & memory primitives
Geometric addressing & provenance
Memory consolidation & living-state synthesis
World genesis & branch operators
Mixture-of-Experts (MoE) language models have become the dominant architecture for cost-efficient scaling, but their routers — typically learned softmax gates — are opaque. The selection of which expert processes a given token is the result of training dynamics over millions of s…
Rotary Positional Embeddings (RoPE) [Su et al. 2021] encode token position by rotating successive pairs of head-dimension components through angles $m\theta_i$, where $m$ is the position and the $\{\theta_i\}$ form a geometric progression $\theta_i = \mathrm{base}^{-2i/d}$. Subse…
Curriculum learning [Bengio 2009] and structured evaluation taxonomies [Hendrycks 2021; Liang 2023; BIG-Bench 2023] are two of the more robust levers for improving the sample efficiency and the interpretability of language-model training. Curricula need an ordering; evaluation su…
Abliteration [Arditi et al. 2024] is a weight-editing technique that identifies a single "refusal direction" in residual-stream activation space and removes it from selected projection matrices, producing a model that no longer refuses harmful requests. The mechanism is remarkabl…
We describe **Multiverse Superposition Inference (MSI)**: a decoding strategy in which a language model maintains *K* parallel hypothesis trajectories ("multiverse") during generation, assigns each trajectory an evolving scalar credence ("superposition" weight), and collapses the…