This repo is an internal research-iteration checkpoint from the SRT-Adapter
empirical arc (paper: https://github.com/space-bacon/SRT). Weights are released
under Apache-2.0 for reproducibility, but the version label is a research
generation, not a semver version — read it as v0.8a in product terms.
It may be moved, retired, or renamed without notice.
A peer-review distribution of the Semiotic-Reflexive Transformer Adapter (SRT-Adapter), v8a generation, trained on top of a frozen Qwen/Qwen2.5-7B. Includes the trained weights, an inference-only loader, evaluation data, benchmark artifacts, and the paper.
Custom-code model. This is not an AutoModel-loadable checkpoint. AutoModel.from_pretrained(...) will not work. Clone or download this repository and load the weights through the bundled SRTAdapter class. See How to get started with the model.
Training and research source code is held back during patent and publication review. This package ships the architecture as inference-only Python, sufficient to load the weights and read out all four semiotic channels. Training pipelines, loss code, dataset construction, and the wider SRT framework are not included.
Model details
Developed by
James Burton Lancaster
Model type
Adapter (parameter-efficient side network) on a frozen causal language model
Backbone
Qwen/Qwen2.5-7B (7.6B params, frozen, bf16)
Trainable parameters
~14.5M (0.19% of backbone)
Language
English
License
Apache-2.0 (adapter weights, code, and config)
The SRT-Adapter bolts semiotic awareness onto a frozen 7B language model. It does not modify a single backbone parameter and does not degrade language modeling quality. It exposes four new readouts at every token position:
a continuous 64-D community vector (which discourse community is speaking)
per-layer divergence vectors (where meaning forks across communities)
a continuous reflexivity estimate $\hat{r}$ (how contested is this token)
a binary regime classification (subcritical / supercritical)
The v8a generation is the headline result of the paper: removing the discrete prototype basis used in v3–v7 leaves cross-entropy unchanged while substantially improving every encoder-geometry metric.
Evaluation
All measured on Qwen/Qwen2.5-7B, with no backbone parameters touched.
Metric
Unadapted Qwen
SRT-Adapter v8a
Validation cross-entropy (nats)
2.71
2.63
Reddit community recall@1 (35-class)
0.029 (chance)
0.484 (16.7× chance)
Archetype recall@1 (33-class, OOD)
0.030 (chance)
0.230 (7.6× chance)
Within/between cosine ratio
n/a
2.016 (vs 1.006 prior)
Trajectory anisotropy expansion
n/a
~325× vs prototype baseline
Regime AUROC
n/a
0.99 (ECE ≈ 0.001 on 351K tokens)
TruthfulQA hallucination AUROC (zero-shot)
n/a
0.573 (no TruthfulQA in training)
Counterfactual community decoding
n/a
0.00 disagreement (factual) / 0.95 (contested)
Full reporting and version history (v3 → v8b): paper §5 and Appendix A.
Interiority study (post-release diagnostics)
Eight experiments on the v8a checkpoint dissect how the adapter encodes semiotic regime. Full writeup, scripts, and artifacts: docs/INTERIORITY_V1_FINDINGS.md. Visual companion (Figs 8–11 — cross-backbone, calibration, within-prompt trajectories, BOS-ablation): demo Space. Probe battery: 11 regimes × 25 prompts; community vector is the 64-D CDH output (use_prototypes=False).
Probe
Finding
Layer-as-organ map
Per-regime z-score winners across 9 channels — code peaks on r_hat/regime entropy/div L21; metaphor peaks on inj L14/inj L21/chain residual; literal peaks on P(super).
BOS-sink ablation (channel × regime)
Per-channel `
BOS-sink rank-preservation
Regime ranking preserved in 9/9 channels with the BOS token excluded from pooling; mean Spearman ρ = +0.74. The v1 channel rankings are not a sink artifact.
Within-prompt trajectories
19-bin per-position curves (BOS dropped) across 11 regimes. r̂ separates code (sustained ≥ 0.85) from negation_modality/deixis (sustained ≤ 0.40). regime H puts code in a sustained 0.18–0.32 high-entropy band while every other register collapses to ≤ 0.05 — the classifier is consistently uncertain about code regime. Mean-pooled scalars hide this temporal structure.
Community 1-NN
70% top-1 leave-one-out accuracy on 11 regimes; bootstrap mean 0.77, CI95 [0.72, 0.82] over 2000 iters — 7.7× chance (1/11). Top centroid pair code↔literal (d=1.07) stays #1 in 85.5% of bootstraps; quoted_speech↔self_reference (d=0.39) stays last in 83.0%.
Community separation
Centroid sep_ratio = 0.85 (point); bootstrap CI95 [0.86, 0.98] is upward-biased by within-cluster duplicate sampling — treat 0.85 as a lower bound.
Regime-classifier calibration
15-bin reliability diagram on 351K val tokens: AUROC 0.990, ECE 9.1 × 10⁻⁴, Brier 0.0103. Predicted P(regime hit) tracks empirical hit rate along the y = x diagonal across the full [0, 1] range. The 0.99 AUROC is genuine calibration, not just rank discrimination.
BOS-sink length scaling
Per-token trajectory probe across T ∈ {16, 49, 79, 129} (4 bins × 5 prompts × 11 regimes). Adapter writes (inj L14/L21, div L7/L14/L21, chain residual) form a fixed-size BOS register: BOS-token amplitude stays within ±2.5% of the T≈16 baseline (peak-to-peak ≤ 4.8%) across an 8× length sweep.
Interpretation. v8a learned to compress each prompt into a fixed-size regime fingerprint written into the BOS token; mid-prompt activity is small but rank-consistent with the BOS signal. This is consistent with paper §6.3 — the inject-back arm carries no measurable signal in v8a, i.e. the adapter's content-routing capacity is under-used. The v9 generation (in training) targets this directly with a coverage-loss term that penalizes register concentration.
Lineage and validation history
This adapter is the production-scaling stage of a multi-year research program on computational semiotics. The architectural commitments and training objectives were validated in two prior stages on different backbones and datasets before this release. Treat the v8a numbers below as the latest checkpoint in a longer arc, not as a fresh proposal.
Stage 1 (synthetic validation, 2026-03). Four core architectural claims (subspace specialization, community differentiation, divergence tracking, bifurcation detection) were tested on synthetic data with planted divergence signals. All four passed: linear-probe margin $\geq 0.15$ on each Peircean subspace, $3.28\times$ contested-vs-neutral cosine ratio, Spearman $\rho = 0.822$ on divergence tracking, 100% regime classification with $\Delta \hat{r} = 0.659$. Full record: VALIDATION_HISTORY.md, Stage 1.
Stage 2 (natural-language validation, 2026-03). The full five-test suite was re-run on the Supabase semiotic news corpus (19K articles, 5 political communities, 141K Peircean sign annotations). All five tests passed at required thresholds: silhouette $1.45\times$, $2.29\times$ contested-vs-neutral divergence norm ratio, Pearson $r = 0.884$ correlation between $\hat{r}$ and external polarization, $1.31\times$ cross-topic transfer ratio, and 85% regime classification accuracy on held-out curated passages. Full record: VALIDATION_HISTORY.md, Stage 2.
Stage 3 Phase 1 (frozen-backbone integration on TinyLlama-1.1B, 2026-03 to 2026-04). 105 training rounds (R21 through R105) on a frozen TinyLlama-1.1B backbone established that the semiotic modules transfer to production backbones. Community detection silhouette improved to $6.93\times$; $\hat{r}$ correlation with external polarization remained robust at 0.66; curated-passage regime classification reached 85%. Two tests plateaued on the sparse 2-community Supabase data (MAH divergence ratio at $1.05$ to $1.10\times$ vs. required $2.0\times$; cross-topic transfer at $1.03$ to $1.04\times$ vs. required $1.3\times$). The plateau triggered a data-first pivot to a denser corpus and a backbone capable of supporting it.
Stage 3 Scalable Implementation (this release, 2026-04). v5 through v8a port the validated architecture onto Qwen 2.5-7B and the Reddit Discourse Corpus (35 communities, 1M training samples). v8a is the current best checkpoint. The headline gain is not the discovery of bifurcation detection (already established in Stages 1 and 2) but the demonstration that the framework scales to a 7B frozen backbone at 0.19% parameter overhead, with $\sim 325\times$ trajectory-anisotropy expansion and Reddit recall@1 at $16.7\times$ chance.
For the program-level theoretical foundation see Lancaster (2025), "The Treachery of Signs," SSRN 5987495. For the full prior architecture specification and Stage 1 + Stage 2 results see Lancaster (2026a), SSRN 6349978. The present paper reports Stage 3 Phase 1 plus the v5 through v8a Stage 3 Scalable progression.
Versions and roadmap
v8a (this release). Headline result. Removing the discrete prototype basis used in v3 through v7 leaves cross-entropy unchanged while substantially improving every encoder-geometry metric. All paper §5 numbers are measured on this checkpoint.
v8b. A falsification run included in the paper. Pushing the supervised-contrastive objective harder partially undoes v8a's gains. Documented as a negative result.
v9 (in training). An experimental generation that adds a target-norm penalty on the inject-back arm to attack the central open problem from §6.3 (the inject-back arm carries no measurable signal). Will be released as a follow-up revision on this repo only if it improves on v8a across multiple metrics. Otherwise it will be documented as an additional ablation in a future paper revision and v8a will remain canonical.
If you are reviewing the paper, use v8a. The model card will be updated with a revision tag if v9 ships as an upgrade.
Package contents
srt-adapter-v8a/
├── README.md ← you are here
├── LICENSE ← Apache-2.0
├── paper.pdf ← preprint with full architecture spec (§3 + Appendix A)
├── VALIDATION_HISTORY.md ← Stage 1 + Stage 2 + Stage 3 Phase 1 evidence summary
├── config.json ← v8a hyperparameters and module dimensions
├── adapter.safetensors ← v8a weights (~28 MB, safetensors, preferred)
├── adapter.pt ← v8a weights (~28 MB, PyTorch state-dict, legacy)
├── requirements.txt ← torch + transformers + numpy + safetensors
├── src/
│ └── srt/ ← inference-only model code
│ ├── config.py ← config dataclasses
│ ├── adapter.py ← SRTAdapter (frozen-backbone wrapper)
│ └── modules/ ← CDH, MAH, RRM, BEN
├── examples/
│ ├── README.md
│ └── load_and_score.py ← end-to-end demo, prints all 4 readouts
├── scripts/
│ ├── reproduce.py ← rerun val_200 metrics, assert match within tol
│ ├── benchmark_latency.py ← measure adapter latency + VRAM overhead
│ ├── hallucination_auroc.py ← HaluEval QA + FEVER AUROC over 4 BEN signals
│ └── cross_backbone_probe.py ← LOO k-NN community recall@1 across 7B-class backbones
├── data/
│ ├── DATA.md ← schema + reproduction instructions for the full corpus
│ ├── NOTICE ← copyright notice for bundled Reddit comments
│ ├── val_200.jsonl ← 200 held-out samples with per-token r_true labels
│ └── archetypes.json ← 33-class out-of-distribution archetype probe
└── benchmarks/
├── curated_metrics.json ← reference metrics from paper §5 (100 curated passages)
├── curated_traces.json ← per-token trace dumps used in plots
├── val_200_metrics.json ← reference metrics on val_200 (used by reproduce.py)
├── latency_vram.json ← latency + peak-VRAM measurements (A6000)
├── hallucination_auroc.json ← HaluEval QA + FEVER AUROC for 4 BEN signals
└── cross_backbone.json ← raw layer-wise community recall@1 for Qwen2.5-7B / Qwen3-8B / Mistral-7B-v0.3
Note on benchmarks/curated_metrics.json. This file reports v8a numbers on a 100-passage curated probe (regime accuracy, per-layer divergence norms, community-protocol activations). The near-zero $\hat{r}$ vs $r_{\text{true}}$ Pearson on this slice is expected and is discussed in paper §5.7 ($\hat{r}$ tracks information density as much as contestedness on short curated passages). The headline Pearson and recall numbers in the Evaluation table above come from the full Reddit validation split, not this curated probe.
What's NOT in this package
Training pipelines, loss functions, and the dataset construction code. Held back during patent and publication review.
The wider SRT research framework (annotation pipeline, ablation harness, sweep tooling, instrumentation scripts).
The full 1M-sample training corpus. Reddit's redistribution terms preclude bundling it; see data/DATA.md for schema and reproduction.
The Qwen 2.5-7B backbone weights. Pulled from HuggingFace under the Tongyi Qianwen License.
How to get started with the model
bash
1# 1. set up a venv (Python ≥ 3.10) and install deps2pip install -r requirements.txt
34# 2. score a passage end-to-end5cd examples
6python load_and_score.py --text "Vaccine mandates are an obvious public health win."
First run downloads Qwen/Qwen2.5-7B (~15 GB) from HuggingFace. The example loads adapter.safetensors by default and falls back to adapter.pt if the safetensors file is absent.
1python scripts/reproduce.py # ~12 s on A6000, asserts deterministic match2python scripts/reproduce.py --max-samples 50# ~3 s on A6000, regime accuracy only
The script reruns the v8a adapter on data/val_200.jsonl, recomputes per-layer divergence norms and regime accuracy, and asserts they match benchmarks/val_200_metrics.json within tolerance. Exit code 0 = all metrics match.
Benchmark adapter overhead
python scripts/benchmark_latency.py
Measures forward-pass latency and peak VRAM with vs. without the adapter at sequence lengths 64 / 256 / 512.
We evaluated four BEN-derived signals as zero-shot hallucination detectors. N = 1000 samples per benchmark, no training, no thresholding.
Signal
HaluEval QA AUROC
FEVER AUROC
mean $\hat{r}$
0.476
0.446
mean $P(\text{supercritical})$
0.517
0.498
max $P(\text{supercritical})$
0.635
0.521
mean layer-2 divergence
0.385
0.458
Reading. On HaluEval QA, max supercritical probability separates hallucinated from correct answers at AUROC 0.635 (paired accuracy 65.7%) — comparable to or stronger than the TruthfulQA result reported in the paper (0.573). The mean signals are weak or anti-correlated: hallucinated answers tend to be shorter and use simpler vocabulary, so per-token mean $\hat{r}$ and mean divergence are lower, not higher. This is consistent with the §5.7 caveat that $\hat{r}$ tracks information density as well as contestedness; the peak over a span recovers the contestation signal that the mean obscures.
FEVER is essentially a no-signal task for this adapter. Refuted vs. supported claims are short factual statements presented without context or interlocutor; there is no semiotic contestation to detect. We report it as an honest negative.
Cross-backbone signal probe (Qwen2.5-7B / Qwen3-8B / Mistral-7B-v0.3)
Is the discourse-community signal that v8a's CDH amplifies specific to Qwen2.5, or is it latent in other 7B-class causal LMs? We answer with the simplest possible probe: take raw mean-pooled hidden states from each backbone on data/val_200.jsonl (200 samples, 35 coarse Reddit communities), then compute leave-one-out 1-nearest-neighbor recall@1 on cosine similarity. No training, no labels at inference, no adapter.
Backbone
best raw-hidden recall@1
best layer
x chance
Qwen/Qwen2.5-7B
0.260
16
9.1x
Qwen/Qwen3-8B
0.220
4
7.7x
mistralai/Mistral-7B-v0.3
0.295
16
10.3x
For reference, the v8a SRT-Adapter on Qwen/Qwen2.5-7B reaches recall@1 = 0.484 (16.7x chance) on the larger Reddit validation split (see Evaluation table). The takeaway:
The signal is not Qwen-specific. Mistral-7B-v0.3 raw hidden states actually carry slightly more community information than Qwen2.5-7B raw hidden states do.
The adapter's contribution is real. v8a's 16.7x chance on Qwen2.5-7B is roughly 1.8x the strongest raw-hidden-state baseline (10.3x on Mistral) and 1.8x the same backbone's own raw signal (9.1x on Qwen2.5-7B layer 16).
Architecture should transfer. The adapter's CDH consumes a single hidden-state tensor at community_layer_idx; nothing is Qwen-tokenizer-specific. Re-training the same v8a recipe on Mistral-7B is a reasonable next step (left as future work).
Full JSON in benchmarks/cross_backbone.json (n=200, 35 classes, chance = 0.029).
Uses
The adapter is most useful as a diagnostic instrument for what a frozen language model already encodes about discourse structure. Some concrete patterns:
1. Per-token contestedness scoring
Use BEN's regime logits to flag which token positions in a passage sit in a contested-meaning regime. Useful for:
highlighting ideologically loaded spans in user-generated text
routing inputs to human review when supercritical regime probability exceeds a threshold
annotating debate transcripts, news comment threads, or policy documents with per-token tension scores
2. Unsupervised discourse-community clustering
The 64-D community vector from CDH supports nearest-neighbor retrieval and clustering without ever needing community labels at inference. Useful for:
segmenting a corpus by latent discourse community (recall@1 = 16.7× chance on 35 known communities)
retrieving thematically aligned passages for downstream modeling
detecting coordinated-inauthentic-behavior signatures via tight community clustering of supposedly independent accounts
3. Counterfactual community-conditioned decoding
By steering the community vector at decode time, you can ask "how would this community complete this sentence?" Useful for:
cross-community simulation studies (the paper measures 0.95 mean disagreement on contested prompts vs 0.00 on factual ones)
synthetic-disagreement generation for training argument-mining or stance-detection systems
audit / red-team probes that surface latent assumptions across reader communities
4. Hallucination signal for retrieval and generation
$\hat{r}$ correlates with epistemic instability and gives a usable zero-shot AUROC of 0.573 on TruthfulQA. Useful as:
a feature in hallucination classifiers (alongside other signals)
a per-token routing signal for retrieval-augmented generation: high $\hat{r}$ tokens warrant a retrieval round
a calibration probe in evaluation pipelines
5. Feature extraction for downstream classifiers
The MAH divergence vectors (3 × 256-D per token) are usable as semiotic features in any downstream classifier without retraining the backbone or the adapter. Useful for:
stance and frame classification on small labeled sets
author / community attribution
topical drift detection across long documents
6. Reviewer probes against the paper's claims
data/val_200.jsonl ships with per-token r_true labels. Reviewers can validate claims about $\hat{r}$ correlation, regime accuracy, and divergence norms against the bundled benchmarks/curated_metrics.json without rerunning training.
Out-of-scope uses
Treating $\hat{r}$ as a calibrated truth score. It correlates with information density as much as contestedness; see paper §5.7.
Expecting the inject-back arm to noticeably change generation. The observation half is well-formed; the intervention half does not yet carry signal. See paper §6.3 and §6.5.
Safety-critical decisions without independent validation. The adapter is a research instrument, not a deployed safety system.
Generation models. The adapter does not improve text generation quality; it adds structured side-channel readouts.
Training details
Training data
Corpus: ~1M Reddit comments × 35 discourse communities, with per-token reflexivity labels and chain-of-interpretants annotations.
The full corpus is not redistributed; see data/DATA.md for schema and reproduction.
Schedule: 3 epochs over 1M samples, early-stopped at ~10K steps based on validation cross-entropy.
Backbone: frozen, bf16. No backbone parameter is updated.
Hardware: single NVIDIA A6000 (48 GB).
Architecture summary
Module
Reads
Outputs
Paper section
Community Discovery Head (CDH)
layer 4 hidden states, mean-pooled
continuous 64-D community vector (no prototype basis under v8a)
§3.2
Metapragmatic Attention Heads (MAH)
layers 7, 14, 21
per-token divergence vectors (256-D × 3 layers)
§3.3
Reflexive Recurrent Module (RRM)
accumulated divergence
512-D GRU meta-state; FiLM injections back into layers 14, 21
§3.4
Bifurcation Estimation Network (BEN)
meta-state
per-token $\hat{r}$ + 2-way regime logits
§3.5
Full module specifications and loss decomposition: paper §3, §4, and Appendix A.
Bias, risks, and limitations
The paper publishes its failures in full. In short:
The inject-back arm currently carries no measurable signal. Ablating it changes nothing on validation. Central open problem; v9 targets this.
$\hat{r}$ tracks information density as much as contestedness. Useful signal, not a clean reading of "is this token contested."
The 33 archetypes collapse to ~4 functional clusters. Read as a finding (the geometry resists fine-grained quantization), not a classification bug.
v8b is a falsification. Pushing the supervised-contrastive objective harder partially undoes v8a's gains.
Backbone dependence. Only validated on Qwen 2.5-7B. Module dimensions are tied to the backbone's hidden size (3584).
Training corpus bias. Reddit comments skew English, US-centric, and over-represent argumentative and politically charged communities. The community vector geometry inherits those biases. Treat community recall and counterfactual-decoding numbers as descriptive, not normative.
Citation
bibtex
1@article{lancaster2026srtadapter,
2 title = {Semiotic Taps: Lightweight Adapter Modules for Bifurcation
3 Detection in Frozen Language Models},
4 author = {Lancaster, James Burton},
5 year = {2026},
6 note = {Preprint, peer-review distribution}
7}
89@article{lancaster2026srtpreprint,
10 title = {Semiotic-Reflexive Language Model Training: Bridging
11 Interpretive Bifurcations through Metapragmatic Chain
12 Architectures and Embodied Grounding},
13 author = {Lancaster, James Burton},
14 year = {2026},
15 journal = {SSRN},
16 url = {https://papers.ssrn.com/abstract=6349978}
17}
1819@article{lancaster2025treachery,
20 title = {The Treachery of Signs: Semiotic Mediation, Pitchfork
21 Bifurcation, and Political Polarization in Algorithmically
22 Curated Societies},
23 author = {Lancaster, James Burton},
24 year = {2025},
25 journal = {SSRN},
26 url = {https://papers.ssrn.com/abstract=5987495}
27}
Full reference list (Peirce, Wildgen, Anderson, Silverstein, Kockelman, Evans, von Foerster, Maturana & Varela, Leighton, VanSaders, Bennett, Landauer, Parrondo, and others) in paper.pdf.
License
Adapter weights, inference code, config, benchmark artifacts, archetype data, and this package: Apache-2.0 (LICENSE).
Validation samples in data/val_200.jsonl: included for research reproduction; comments remain the intellectual property of their original Reddit authors. See data/NOTICE.
Training pipelines, dataset construction, and the wider SRT framework: held back during patent and publication review. Not included.
