MMeaning Lab.ai

Independent Research Program · Est. 2025

What a source tag
actually does.

Meaning Lab studies provenance and attribution inside language models — not as policy or metadata, but as a measurable property of the model itself.

Watermarking, content credentials, training-data attribution, and every proposed mitigation for model collapse share one presumption: that carrying a source alongside content does something for the model. That presumption is testable. This program tests it, and reports where it holds, where it fails, and where the author's own earlier predictions were contradicted.

Read the preprints Key findings
152fine-tuned models in the consolidated training program
6,600scored texts across three GPU runs in the anchor study
3current preprints, 2026
3 of 7of the author's own prior falsifiable claims contradicted
01

The research program

Language models are increasingly trained on, and prompted with, text that other language models produced. In that setting a question stops being philosophical and becomes an engineering constraint: when content passes through a model repeatedly, what happens to the record of where it came from?

The standard assumption is that provenance rides along with content — that if the facts survive, the sourcing survives with them. This program treats that as a hypothesis and tests it under controlled recursion, across architectures and scales, with preregistered contrasts.

The answer is more specific, and more uncomfortable, than the assumption. Provenance and factual content come apart sharply. A tag's usefulness is only partly a function of whether it is true. And a monitoring system that tracks information retention can report a system as healthy while its attribution has almost entirely failed.

Systems that monitor only entity or fact persistence can report high information retention while missing severe attribution failure.

The work is deliberately narrow. Claims are bounded to the architectures, scales, and protocols actually run; none of it establishes architecture-general memory behaviour or universal structural laws. Where an experiment contradicted a prediction — including predictions made in this author's own earlier theoretical framework — the disconfirmation is reported in full.

02

Selected findings

61.6percentage points

Provenance is path-dependent under recursion

After ten recursive passes, correctly attributed fact retention was 66.8% under correct self-relay but 5.2% when provenance was corrupted at origin — a preregistered paired difference positive in all 32 records. The corrupted lineage still retained 87.7% of entity–value facts.

Source Relations Under Recursive Transformation (2026)

68%

Most of the protection is structural, not semantic

A true source tag slows recursive drift by 24% relative to no tag. But a bare prefix of five tilde characters — carrying no semantics, no identity, no bracket structure — still captures 68% of that protection. An opaque identifier captures 94%.

Three Jobs of a Source Tag (2026)

13×the noise floor

Position carries the structure, not the source

Take a complete and correct source title and move it from the start of the sentence to the end, and it captures none of the protection — its mean sits inside the noise floor and its sign reverses between runs. Five meaningless characters in the leading position outperform an accurate provenance record in the trailing one by thirteen times that floor.

Three Jobs of a Source Tag (2026)

1.77–2.76×

False attribution costs more than correct attribution buys

Inverting the models by Bayes to recover source from text turns them into attributors without retraining. Across five independent settings the damage-to-benefit ratio never reverses direction. Critically, that damage is invisible to predictive loss: the corrupted model is unremarkable on masked sentence loss while losing 25 points of attribution accuracy.

Three Jobs of a Source Tag (2026)

80% coverage

A compact anchor is a checklist, not a compression

Agent architectures re-inject a compact summary to resist drift, justified by a compression intuition. That intuition is false. Protection is item-wise: an anchor protects the terms it names and not the ones it omits. Naming 20% of tracked terms confers no detectable benefit; naming 80% is statistically equivalent to the full source at 31% of its token cost.

The Anchor Protects What It Names (2026)

26% residual drift

No dose of ground truth arrests a recursive system

Sweeping the fraction of real human text reinjected at each generation produces a smooth, threshold-free curve: 5% of real data removes 24% of drift, 25% removes 62%, and 40% still leaves 26% of the unanchored drift intact. There is no small dose of ground truth that stabilises the loop.

Three Jobs of a Source Tag (2026)

87% / 13%

Attribution slows collapse whether or not it is true

Across thirty GPT-2 fine-tunes, tagged lineages drifted 25% less than untagged. A lineage anchored entirely to false sources resisted collapse nearly as well as one anchored to correct sources: 87% of the protection came from the presence of the slot, 13% from its accuracy.

Attribution Slows Model Collapse Whether or Not It Is True (2026)

AUC 0.539

First-generation synthetic text is near-indistinguishable

Synthetic text separates from human text not by being improbable but by being too probable — and once length is matched, first-generation output is very nearly indistinguishable. Separability accumulates only through repeated recursion, which is a caution for anyone monitoring these systems.

Three Jobs of a Source Tag (2026)

03

Preprints

All experimental work is released as open preprints with code, data-generation scripts, and evaluation harnesses included; figures are regenerated from source result files. The three papers below are the current record. Earlier consolidations they subsume are listed after them rather than removed, so the versioning is visible.

Preprint v1.0 · Aug 2026

Source Relations Under Recursive Transformation: Provenance Path Dependence and External Re-grounding in Causal Language Models

Factual preservation and provenance preservation are not the same problem.

A mechanism ladder narrows a structural source-boundary effect in Pythia-410M to source-token prediction in a first-occurrence regime, then tests recursive provenance in a frozen Qwen2.5-7B-Instruct protocol: 32 records, five conditions, ten recursive passes, 1,600 generations. Initial attribution state strongly constrains later attribution, and re-grounding against an immutable provenance ledger outperforms recursive self-relay by 15.1 points.

Christopher W. Sweeney · 11 August 2026 · Pythia-410M, Qwen2.5-7B-Instruct · Preregistered recursive provenance assay

DOI 10.5281/zenodo.21890267 PDF

Preprint v1 · Aug 2026

The Anchor Protects What It Names

Compact context anchors protect the source content they enumerate and not the content they omit — and refreshing them from working state removes the protection.

Forty technical passages carrying 708 curated terms, rewritten recursively ten times under five conditions, replicated across three independent GPU runs and two architectures. Naming a term raises its ten-generation survival by +0.143. A randomised, yoked assignment settles the direction of causation. A self-refreshed anchor is statistically equivalent to having no anchor at all on the very content it began by naming — a checklist that forgets its items.

Christopher W. Sweeney · August 2026 · Qwen2.5-7B-Instruct (×2), Mistral-7B-Instruct-v0.3 · 6,000 generations, 6,600 scored texts

DOI 10.5281/zenodo.21855188 PDF

Consolidated preprint v1 · Aug 2026

Three Jobs of a Source Tag

Semantic conditioning, attribution integrity and stability under recursive self-training depend on different properties of the tag — and only one of the three requires the source to be true.

Eleven experiments across 152 fine-tuned models separate three outcomes that are routinely conflated. The semantic job requires the tag to be about the text; substituting an author for a title collapses the conditioning benefit to the opaque-identifier floor. The structural job requires only the occupancy of a position. The attribution job is the only one that requires truth — and getting it wrong costs roughly twice what getting it right buys.

Christopher W. Sweeney · 6 August 2026 · GPT-2 124M/355M/774M, Qwen2.5-0.5B · Wikipedia and post-cutoff arXiv · Supersedes Zenodo 21782935, 21796258, 21810517

PDF

Superseded — retained for the record

Both papers below are subsumed by Three Jobs of a Source Tag, which reports the same experiments at greater depth and corrects two claims made here. They remain listed because the corrections are part of the record.

Earlier consolidation · Aug 2026

The Positional Function of Source Attribution in Language Models

Correct attribution lowers loss by 0.22 nats and slows recursive drift by a quarter — but a meaningless placeholder occupying the same position captures two-thirds of that protection.

Eight experiments across approximately seventy fine-tuned models. The conditioning effect is not memorisation: it is undiminished when all test titles are unseen during fine-tuning, it strengthens with model scale, it transfers across architectures, and it is largest on a post-cutoff corpus where pretraining familiarity is impossible. The decisive comparison is the one that fails — what the tag contributes structurally is the occupancy of a position, not the transmission of a source.

Christopher W. Sweeney · 6 August 2026 · ~70 fine-tuned models, a subset of the 152 · Superseded by Three Jobs of a Source Tag, which replicates the suffix arm and narrows this paper's trailing-tag claim

PDF

Earlier preprint · Aug 2026

Attribution Slows Model Collapse Whether or Not It Is True

Structure contributes 87% of the protection; accuracy protects only the attribution chain, and that effect compounds with recursion depth.

Every published mitigation for recursive collapse works by importing information from outside the loop. This paper tests an intervention that imports nothing: a source tag inside the model's own generated text. Thirty GPT-2 models trained across five recursive generations in three lineages. Meanwhile attribution does not collapse under recursion — it inflates: a corpus of 2,073 real sources becomes 6,700–14,044 distinct generated titles, 28.2% of generated sentences carrying a fabricated source by generation five.

Christopher W. Sweeney · 5 August 2026 · 30 GPT-2 fine-tunes · 3 conditions × 5 generations × 2 seeds · This experiment is reproduced and extended within Three Jobs of a Source Tag, which supersedes both companion records

Companion DOI 10.5281/zenodo.21782935 Companion DOI 10.5281/zenodo.21796258 PDF

04

Earlier theoretical work

The experimental program above was undertaken to test a theoretical framework deposited thirteen months earlier. Of its seven falsifiable claims, three survived, one survived in weakened form, and three were contradicted — including a numbered axiom and a preregistered numerical prediction made in the framework's own experimental-validation section. Those papers are retained here as the record of what was predicted, on the view that a framework whose predictions can fail is worth more than one whose cannot.

  • Meaning, Recursion, and the Field

    Retrospective white paper v4.0 · March 2026 · Separates what was validated from what overreached, including an explicit correction to the quantum extension

    PDF ↓

  • The Sweeney Theory of Meaning: A Comprehensive Scientific Framework

    White paper v1.0 · 2025 · The four-axiom framework whose predictions the 2026 experiments were built to test

    DOI ↗

  • Guarding the Signal: A Framework for Identifying and Repairing Semantic Drift in Generative AI

    2025 · Definitions and detection metrics for semantic drift, symbolic degradation, and attribution drift

    DOI ↗

  • The Cognitive Void-Resonance Model: How Emergent Meaning Forms in Human–AI Symbolic Systems

    2 July 2025 · Symbolic substrate of the earlier framework · CC BY-NC 4.0

    DOI ↗

  • Beyond Prompting: How to Speak to AI the Way a Garden Speaks to the Sun

    Book · C. Sweeney · A practitioner's treatment of the meaning-preservation framework, written in documented collaboration with AI systems

    Enquire →
05

About

Meaning Lab is the research arm of Coherence Systems LLC, run by Christopher W. Sweeney. It is independent and self-directed, and the experimental work is funded and run without institutional sponsorship.

The work sits at an unusual intersection. Sweeney spent more than fifteen years in utility regulation and federal energy infrastructure — enforcement casework at NERC, rate-case strategy at Exelon, and national compliance program design for the U.S. Department of Energy under the Bipartisan Infrastructure Law. That career is about evidentiary chains: who asserted what, on what record, and whether the assertion survives adversarial review.

The research applies the same instinct to language models. Provenance in a regulated proceeding is a control-plane property, not a courtesy. The question this program asks is whether it can be one inside a model.

Method matters here, and is stated plainly rather than obscured: the papers are developed in structured collaboration with multiple AI systems in defined roles, with a human author as origin and final witness. That collaboration is documented in each work — not as a disclaimer, but because the argument about attribution applies to this work as much as to anything else.

Christopher W. Sweeney

ORCID
0009-0007-6549-2148
Practice
Founder & Principal, Coherence Systems LLC
Education
M.L.A. Management, Harvard University
B.A., Tulane University
Prior
U.S. Department of Energy
Exelon Corporation (Pepco)
North American Electric Reliability Corporation
Based
Massachusetts, United States
06

Contact

Correspondence on the research, replication requests, and enquiries about collaboration or review are welcome.

cwrsweeney@gmail.com

Code, data-generation scripts, and evaluation harnesses accompany each preprint and are available on request where not already deposited.