TOM World & Query Kernel 0.6.0

TOMK is a research implementation of a deterministic finite symbolic learner and its publication pipeline. Kernel 0.6.0 / TOM Learner 0.2 evaluates 121 literal hypotheses from four exact families against sixteen supplied data sets. It records ambiguity, counterexamples, contradiction and regression evidence, then materializes accepted or rejected outcomes through the TOMAGI runtime shipped in this repository.

The Hugging Face repository is c0mbatduckzz/TOMK. Start with the learner specification, seeded compiler specification, and formal language specification. The code and literal definitions are released under the MIT license.

The exact authoritative seed is TOM_seed_genome_2026-09-01.txt: 244 ASCII bytes, no terminal newline, SHA-256 d1417a3136772c0cf3eddcd4962ce07d42cbf87616f7b5bae09fc652d9b807b5. Its requester-supplied attribution is Tom Klootwijk; NL200678942; 10-07-1990, as recorded in NOTICE.md. The identifiers are supplied attribution and were not independently verified.

TOM Learner 0.2 — finite typed hypothesis-family authority

WQK 0.6 continues the corrected authority roadmap rather than restarting it. The exact canonical TOM seed remains the root. The fixed TOMAGI 1.0 ABI remains 128-byte header, 64-byte State64, 48-byte Cell48, and sixteen opcodes.

The release incorporates and regression-tests the CODEX 0.5.2 kernel repairs before broadening the learner:

defined C arithmetic with explicit wrap32 lowering;
rejection of all nonzero reserved TOMAGI header words in Python and C;
same-host thread/process publication locking from expected-HEAD read through atomic replacement;
recursive formal-value and canonical-byte limits at every intermediate node;
reproducible package construction from pinned authority sources; and
public audit-store SOURCE STORE argument order.

Above that repaired boundary, WQK 0.6 implements a finite content-addressed registry of four exact hypothesis families:

Family	Candidates
Rational polynomial, degree at most two	34
One-breakpoint piecewise affine	21
Complete finite transition table	27
Depth-two expression tree	39
Total	121

The domain decision is made by static content-addressed formal definitions. The generic Python formal evaluator evaluates the declared learner and promotion operations during seeded compilation. The compiler canonically encodes those results and lowers them into a Cell48 graph that emits the result bytes. Python and C then execute the compiled .tmg program; authenticated ordered EMIT records reproduce the materialized bytes. This release's learner search therefore runs at compilation time, while TOMAGI execution replays its compiled output.

Host Python also provides strict validation, independent falsification, trace authentication and generic immutable storage. The separately implemented fractions.Fraction oracle compares the learner outcomes and cannot select or publish an authoritative result.

Authority chain

canonical 244-byte TOM seed + pinned literal definitions and exact data sets
-> verified content-addressed dependency graph
-> bounded formal learner and promotion evaluation during compilation
-> canonical result bytes
-> deterministic Cell48 EMIT lowering and compiled .tmg
-> equal Python/C TOMAGI execution
-> replay-authenticated EMIT materialization
-> same-host locked immutable publication
-> reconstructed terminal HEAD

Multiple exact train survivors produce an explicit ambiguity record and rejection. There is no hidden stable-sort winner. Supersession proposals must pass regression cases for all pinned prior definitions. Accepted and rejected sessions both continue through the parent-bound 0.5.2 publication profile.

Canonical result

families:          4
candidates:      121
data sets:        16
accepted:          9
rejected:          7
ambiguities:       3
false promotions:  0

The independent oracle agrees with every data-set result.

Learner chain

formal program:
sha256:a07d27c1fe88b75b56f19d1e623a170da6ee3271c3638836f7badf079ec170c3

compiled .tmg:
32,880 cells; 1,578,368 bytes
sha256:5feac19609ed9577688990e1e5adeb7caa81c05d9e26da559ec93be45899c3cf

materialized result:
131,517 bytes
sha256:4e59666a7ccdc2505d94fe760f5d317f5302f8766a6bd8f1b022912890e5844a

Promotion chain

initial repaired 0.5.2 HEAD:
sha256:a3bd8ecd8578b28158b96a3dce814910beb3d627068159dc668a682c85b85448

publication plan:
sha256:335b3349591e489af6c67c16b563547997f5e7cb29d4a1e685476b1cff69510c

compiled .tmg:
157,014 cells; 7,536,800 bytes
sha256:fe32d60b54a8fc38e0bf07f3ad7311af01d485ccf342488a880d69bc455a0b6b

materialized value:
628,055 bytes
sha256:9e1d55a17bf45db48cc22588a4a7168ae2dc10720f2edb557130c3ad80318663

terminal HEAD:
sha256:f52198541544eff90df272327236af75c4dd729b77cdf75628b0bad0bf17502e

The promotion store has 176 files, 597,515 bytes, and deterministic tree hash:

sha256:d125c28b7570cd2edae109747557cdc07573ec28ac141fe50f9059250eaa4787

Installation, build and validation

Requirements:

Python 3.10+
GNU Make and a POSIX shell for the build recipes
a C99 compiler available as cc, or selected through Make's CC variable
jsonschema 4+ for fixture generation, schemas and validation; included in the dev extra

Download and extract the complete release archive. From its repository root, use this Linux/WSL setup recipe:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e ".[dev]"
tomagi --help
tom-learner06 --help

The installed entry points expose the Python runtime and CLI. Keep the repository checkout available for the seed, formal programs, fixtures, C sources and Make recipes. On Windows, the Python package can be installed from PowerShell using a virtual environment; run the full Make recipe inside a POSIX environment such as WSL. Keep store writers within one supported host environment: mixed native Windows/WSL writers to the same store are unsupported.

The following commands regenerate local outputs. They are the intended release-validation recipe; the recorded reports described below identify the evidence already shipped. Use CC=gcc if the compiler is named gcc instead of cc.

# Generate formal sources, execute the learner and promotion chains,
# compare the independent oracle, publish and audit the immutable store.
make learner06 PYTHON=python

# Run the complete inherited, repair, and Learner 0.2 suite.
make test-learner06 PYTHON=python

# Add core validation, two-build clean replay, and final validation.
make validate-learner06 PYTHON=python

# Build the deterministic ZIP and replay it from a clean extraction.
make package-learner06 PYTHON=python

The core report records 283 passing tests, eighteen passing validation checks, twenty passing rejection cases, and no failures. The final report records nineteen passing checks, including two equal clean builds of the declared boundaries and the promotion-store tree. The publication revision of 2 October 2026 rebuilds these records on native Windows with Python 3.12 and a Zig C99 toolchain. The reports certify their recorded host and declared boundaries; inherited historical artifacts remain package inputs.

The publication manifest and document validation additionally bind twelve publication documents, including both use-case reports in PDF and Markdown, to executable literal definitions. Each document is compiled and checked through equal Python/C traces, ordered EMIT records and exact materialized bytes. For the imported reports this proves archival replay. Their proposed applications remain proposals. The full source, generated programs, traces and inventories are available in the complete release archive; its external audit records complete ZIP equality across two independent builds and a finished-archive replay.

The package command additionally requires deterministic ZIP construction, an internal manifest and checksums, and replay from a clean extraction. Preserve the complete archive and its matching manifests for reproduction. A mathematical determinism argument in the specifications and the recorded conformance tests are different forms of evidence; this release provides no machine-checked proof of the whole implementation.

CLI

# Independent oracle over one or more data sets.
PYTHONPATH=src/python python3 -m tom_learner06 oracle \
  examples/learner06/family_registry.json \
  examples/learner06/prior_authority.json \
  examples/learner06/datasets/*.json

# Validate the formal publication plan.
PYTHONPATH=src/python python3 -m tom_learner06 validate-plan \
  validation/learner06/promotion_authority.direct.json

# Apply the plan to a new store.
PYTHONPATH=src/python python3 -m tom_learner06 apply-plan \
  validation/learner06/promotion_authority.direct.json \
  TOM_seed_genome_2026-09-01.txt \
  /path/to/store

# Audit argument order is SOURCE then STORE.
PYTHONPATH=src/python python3 -m tom_learner06 audit-store \
  validation/learner06/promotion_authority.direct.json \
  examples/learner06/promotion_store

Key files

TOMAGI machine and binary format — runtime equations, ABI and fixed-width records.
Seeded compilation — literal grammar, hashing, dependency evaluation, lowering and materialization.
Formal language — executable formal operations and bounded evaluation semantics.
WQK 0.6 learner — finite families, selection, ambiguity and promotion requirements.
CODEX repair handoff and artifact proof — correction source and provenance.
Validation handoff — validation-repair disposition and replay handoff.
Release overview and roadmap.
Family registry, formal learner, and formal promotion — executable literal authority.
Independent oracle and immutable store — falsification and generic publication mechanics.
Canonical outcomes, final recorded validation, and package inventory.

Proposed use cases

The supplied use-case documents describe application architectures and proposed future development. Their implementation status is proposal: the designs extend beyond the implemented and validated capabilities of kernel 0.6.0. The kernel's current evidence remains the finite four-family learner, deterministic compilation and execution, and authenticated publication pipeline described above.

Use-case overview and implementation boundary.
TOM Commons / Human Needs Network: PDF and Markdown.
TOM Manufacturing Horizons / Frontier Biological Life: PDF and Markdown.

Evidence boundary

This release demonstrates exact finite search over four literal hypothesis families. It does not establish noisy learning, open-domain induction, cognitive memory, planning, perception, autonomous action, distributed consensus, physical GPU execution, general intelligence, or AGI.

The next milestone is TOM Learner 0.3 / WQK 0.7: interval-valued observations and finite noise families with explicit calibration, coverage, distribution-shift, ambiguity, supersession, and regression evidence. Confidence must remain a typed record rather than an invisible TOMAGI control path.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support