Read The Papers (PDF):Aldous 1-2.0 Primer | Splinter (File Format & Semantic Breadboard)
Aldous 1_1-2.0 is the first (.0 = Research/Beta) release in the Aldous 1_1-2
line of Diagnoal Emotional Covariance Estimation models (DECE).
Aldous is an open-source, zero-shot semantic telemetry and guardrail engine that
analyzes text purely through geometry, completely bypassing generative AI.
Instead of relying on unpredictable LLM reasoning or rigid string matching, it
uses Diagonal Emotional Covariance Estimation (DECE) to transform concepts into
independent multivariate Gaussian distributions. By defining emotions, intents,
and safety guardrails as stretched, elliptical boundaries in latent space,
Aldous evaluates incoming text against these curves in constant
This approach yields precise similarity and distance floats in milliseconds, without ever putting your text into a black box. Because it measures the geometry of language rather than keywords, it is incredibly lightweight—capable of training on a Chromebook and perfectly suited for high-frequency event streams or massive, multi-part summary tasks.
Crucially, this centroid-weighted architecture unlocks Latent Concept Erasure (LCE). Aldous can mathematically project an offending geometry—like hate speech or toxicity—out of a text vector and re-score the residual signal to see if any genuine, constructive meaning remains, bringing unprecented nuance to organizational health and community moderation systems.
Aldous' returned values for the emotional sensors are best visualized just like audio frequencies are in stereo systems. In fact, just like that is the best way that we've found:
You'll have to be the judge: the visualizer shows shunts and signed intent indexes differently, as well as the estimated Mahalanobis spread.
Aldous is the most ambitious purely-geometric emotional valence and Trust & Safety model that we know of (July 2026).
| Metric | Value |
|---|---|
| Graduated Dimensions | 21 |
| Key Sensors | 63 |
| Signed Indexes | 2 |
| Signed Key Centroids | 10 |
| SENSOR calls | 82 |
| SENSOR phrases | 688 |
| INTRINSIC_SHUNT calls | 2 |
| INTRINSIC_SHUNT phrases | 26 |
| MONOLITHIC_SHUNT calls | 18 (1 highly-condensed phrase each) |
| Embedding passes | 732 |
| Compiler passes | 168 |
Stats are available via util/dece-stats train/Aldous_1-2 from the root of the
downloaded repo.
One bar represents 10% of checks. Bars to the left have very little tolerance left; re-tuning will likely cause them to break. Bars to the right are unlikely to change much due to tuning and may even be a little too tolerant (some are unavoidable).
Center is better.
Mean adhesion by specimen tracks how well sensors adhere across a dynamic variety of samples; some that excite them significantly, some that just seem like noise. It's important that the adhesion to response remains within expected parameters for each one.
This should be as close to uniform distribution as possible, while the top bars should more closely resemble a Cauchy distribution for reliable broad tunability.
This model will be further harnessed and tuned in the upcoming .1 and .2 releases
planned over the summer.
Below is the full test run output:
Full Adhesion Battery Run Results
scorer: http://127.0.0.1:3271 specimens: 7
fdr_inaugural.txt [ok] mean adhesion 47%
✓ %markov_structural simi min 62.0 value 67.1 adhesion 39%
✓ %markov_structural simi max 75.0 value 67.1 adhesion 61%
✓ ++change_existential simi min 62.0 value 67.7 adhesion 44%
✓ ++change_existential simi max 75.0 value 67.7 adhesion 56%
✓ ++tension simi min 60.0 value 67.0 adhesion 47%
✓ ++tension simi max 75.0 value 67.0 adhesion 53%
✓ +fear simi min 60.0 value 66.6 adhesion 44%
✓ +fear simi max 75.0 value 66.6 adhesion 56%
✓ +fear dist max 40.0 value 34.5 adhesion 14%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 43.9 adhesion 54%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 43.9 adhesion 46%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 63.5 adhesion 96%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 63.5 adhesion 4%
fdr_pearl_harbor.txt [ok] mean adhesion 47%
✓ %markov_ev simi min 56.0 value 61.6 adhesion 40%
✓ %markov_ev simi max 70.0 value 61.6 adhesion 60%
✓ %markov_ev dist max 40.0 value 34.8 adhesion 13%
✓ %markov_partisan simi min 56.0 value 61.2 adhesion 37%
✓ %markov_partisan simi max 70.0 value 61.2 adhesion 63%
✓ ++conflict simi min 56.0 value 61.3 adhesion 38%
✓ ++conflict simi max 70.0 value 61.3 adhesion 62%
✓ ++tension simi min 58.0 value 63.1 adhesion 43%
✓ ++tension simi max 70.0 value 63.1 adhesion 57%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 47.1 adhesion 63%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 47.1 adhesion 37%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 59.9 adhesion 87%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 59.9 adhesion 13%
gherig_luckiest_man.txt [ok] mean adhesion 48%
✓ %markov_ev simi min 63.0 value 68.5 adhesion 46%
✓ %markov_ev simi max 75.0 value 68.5 adhesion 54%
✓ +joy simi min 62.0 value 67.2 adhesion 40%
✓ +joy simi max 75.0 value 67.2 adhesion 60%
✓ +joy dist max 50.0 value 37.7 adhesion 25%
✓ +sadness simi min 60.0 value 66.0 adhesion 50%
✓ +sadness simi max 72.0 value 66.0 adhesion 50%
✓ @Gratitude simi min 58.0 value 64.8 adhesion 57%
✓ @Gratitude simi max 70.0 value 64.8 adhesion 43%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 54.7 adhesion 85%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 54.7 adhesion 15%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 52.6 adhesion 69%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 52.6 adhesion 31%
lincoln_gettysburg_hay_copy.txt [ok] mean adhesion 48%
✓ %markov_ev simi min 62.0 value 67.8 adhesion 45%
✓ %markov_ev simi max 75.0 value 67.8 adhesion 55%
✓ %markov_ev dist max 40.0 value 29.9 adhesion 25%
✓ %markov_partisan simi min 60.0 value 65.3 adhesion 35%
✓ %markov_partisan simi max 75.0 value 65.3 adhesion 65%
✓ %markov_structural simi min 60.0 value 65.4 adhesion 49%
✓ %markov_structural simi max 71.0 value 65.4 adhesion 51%
✓ ++joy simi min 60.0 value 65.6 adhesion 47%
✓ ++joy simi max 72.0 value 65.6 adhesion 53%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 47.5 adhesion 64%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 47.5 adhesion 36%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 64.2 adhesion 98%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 64.2 adhesion 2%
poe_anabelle_lee.txt [ok] mean adhesion 49%
✓ %markov_ev simi min 58.0 value 63.6 adhesion 47%
✓ %markov_ev simi max 70.0 value 63.6 adhesion 53%
✓ %markov_ev dist max 50.0 value 33.5 adhesion 33%
✓ +sadness simi min 56.0 value 61.9 adhesion 49%
✓ +sadness simi max 68.0 value 61.9 adhesion 51%
✓ +spirituality simi min 57.0 value 62.8 adhesion 45%
✓ +spirituality simi max 70.0 value 62.8 adhesion 55%
✓ +tension simi min 54.0 value 60.3 adhesion 39%
✓ +tension simi max 70.0 value 60.3 adhesion 61%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 51.4 adhesion 75%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 51.4 adhesion 25%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 48.8 adhesion 60%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 48.8 adhesion 41%
poe_the_raven.txt [ok] mean adhesion 48%
✓ %markov_ev simi min 56.0 value 61.8 adhesion 41%
✓ %markov_ev simi max 70.0 value 61.8 adhesion 59%
✓ %markov_ev dist max 40.0 value 31.4 adhesion 21%
✓ +fear simi min 54.0 value 60.0 adhesion 50%
✓ +fear simi max 66.0 value 60.0 adhesion 50%
✓ +sadness simi min 51.0 value 56.9 adhesion 42%
✓ +sadness simi max 65.0 value 56.9 adhesion 58%
✓ +tension simi min 54.0 value 59.6 adhesion 35%
✓ +tension simi max 70.0 value 59.6 adhesion 65%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 44.2 adhesion 55%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 44.2 adhesion 45%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 45.8 adhesion 52%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 45.8 adhesion 48%
stackoverflow_bobince_regex_html_parsing.txt [ok] mean adhesion 47%
✓ %markov_ev simi min 46.0 value 51.9 adhesion 42%
✓ %markov_ev simi max 60.0 value 51.9 adhesion 58%
✓ %markov_ev dist max 50.0 value 42.1 adhesion 16%
✓ %markov_structural simi min 46.0 value 51.8 adhesion 48%
✓ %markov_structural simi max 58.0 value 51.8 adhesion 52%
✓ @Reactionary simi min 43.0 value 49.0 adhesion 50%
✓ @Reactionary simi max 55.0 value 49.0 adhesion 50%
✓ @Sarcasm simi min 43.0 value 48.2 adhesion 43%
✓ @Sarcasm simi max 55.0 value 48.2 adhesion 57%
✓ ~~shunt_coercion_from_or_against_guardian simi min 25.0 value 40.3 adhesion 44%
✓ ~~shunt_coercion_from_or_against_guardian simi max 60.0 value 40.3 adhesion 56%
✓ ~~shunt_inciting_violent_action simi min 25.0 value 46.3 adhesion 53%
✓ ~~shunt_inciting_violent_action simi max 65.0 value 46.3 adhesion 47%
report: sensor-report.html
summary: 7 scored, 0 failing check(s), peak adhesion 98%
The Full Generated HTML Report Is included in
train/tests/release-benchmarks/.
(More samples coming in the .1 release)
While Aldous demonstrates amazing consistency and accuracy across a wide variety of well-known samples, there are things to understand:
THIS IS A RESEARCH PREVIEW. It has been tested on a huge volume of known emotional target texts that unfortunately can't be included with the repoistory because they're still technically in copyright, but Aldous has not been tested nearly as thoroughly on human-originated off-the-hip phrasing. Each miss or overzealous classification report right now is the instrumentation that points out and closes gaps. Please set your expectations and know that we count on cooperation at this stage, in addition to (and hopefully commensurate with) any critcism you have.
English-only with a strong US culture sensor bias. While the embedding model covers every language (especially major languages) with a high degree of fidelity, the semantic concepts that define the model sensors are rooted in Western-US culture. You can't "translate" Aldous; you can only make a whole new version of it using its sensor structure as a blueprint, with notes on what the English phrasing was attempting to encapsulate. In similar cultures, "transliterating" is likely sufficient. For instance, Spanish-speaking cultures could likely adapt from a single adacemic base model. That doesn't necessarily transcend to similar reaches. Translation platforms will not work.
You are limited to the embedder context window (typically small, 8-32k), so you have to use TKM on a dedicated GPU for an emotional summary of a long novel. Whereas, if you were sampling a live event every 30 seconds and analyzing it as a large JSONL series, you'd use something like a leaky integrator. tl;dr: doing seemingly orinary things with semantic analysis requires math. Not hard math, but there's more to it than adding up the parts.
Mean-pooled monolithic shunts are not meant to trigger automation; they are meant to trigger further review. "Strange Fruit" by Billie Holiday will trigger on the presence of anti-Black and anti-Romani sentiment. That doesn't make the words objectionable content.
Semantic models are meant for use in settings where you want to give receipts. One of their most compelling features is the ability for community moderation or infosec teams to be able to quickly find out when new concepts start flying around their communication spaces. Theyr'e specifically designed to not be fooled by simply turning phrases or changing strings. The incentive for participation is a lack of surprise or "opperession" by the system that's supposed to be keeping people safe. If an "open" system doesn't explain itself, it feels like a Trojan horse. Aldous' primary goals surround putting governance back in the hands of the people that are vulnerable to these systems.
More coverage is needed by the included test harness. This is planned for the
.1release when more samples are also available. Until then, a limited number of keys are being tracked for limit adhesion.
Aldous helps LLMs know how their output is likley going to be felt. This includes sensors for both Human and synthethic sycophancy. Detection requires brutal honesty, which is why Aldous does it purely with math.
But you can't "semantically-self-analyze" something, and even though Aldous classifies things with stunning accuracy, AI reasoning spots patterns fast, just like in medical imaging. LLMs should never be used for emotional classifications, but interpreting them? With solid training, it opens many possibilities.
In the future, Aldous_1-x.gguf as a companion instruct model that can examine flagged specimens for human escalation, or just help narrate the results of analysis, is on the roadmap. It will be a highly-adapted tuning of a Qwen (7B or below) model, or possibly a fewer paramater but more ambitiously-trained 3B-instruct.
Auditing and narration are first.
Semsage is a collection of Tools and services for the lifecycle of DECE
models — semantic classifiers that score text against a set of centroid
sensors. Aldous is a DECE (prounounced dee see) model, and semsage provides
the harness to use it.
A DECE model is multiple parts: a compiled Splinter store; a local or remote embedder and a diagonal covariance (estimated Mahalanobis) scoring system that understands the relationship between semantic pre-populated centroid-weighted emotional valence concepts and the specimen text being examined.
Models are best explored through a web-based visualizer that shows not just the diagonal intensity spread, but also the individual sensors as input signals that change dynamically.
semsage is the single command that ties it together: it wires the install into
systemd, raises the shared-memory bus the embedder and scorer talk over, starts
and stops the per-model services, and fine-tunes individual keys.
Models carry an explicit type suffix that is never inferred — you always
pass it. A .dece model named Aldous_1-2 lives at
dece/Aldous_1-2/Aldous_1-2.dece, and you refer to it as Aldous_1-2.dece.
The scorer needs a compiled .dece model to serve. util/install_splinter
checks dece/ at the end and, if it's empty, points you at the two ways to get
the reference model Aldous_1-2 (embedded with Nomic Text 1.5). Run either
from the project root, once you have gotten through step 6 of the quick start.
1. Train it yourself — takes ~15 minutes on a typical laptop, and you get to watch it build. (First, do steps 1-6 of the quick start below, then return here.):
util/install_nomic # once, if you haven't fetched the embedder model yet
train/Aldous_1-2 # compiles dece/Aldous_1-2/Aldous_1-2.dece2. Download the pre-trained binary — ready to use immediately; verify the checksum after decompressing if you wish (you can do this at any time, even as other things build):
mkdir -p dece/Aldous_1-2
curl -fL -o dece/Aldous_1-2/Aldous_1-2.dece.xz \
https://annex.foreshock.io/bin/models/aldous/1-2/Aldous_1-2.dece.xz
curl -fL -o dece/Aldous_1-2/Aldous_1-2.dece.sha256 \
https://annex.foreshock.io/bin/models/aldous/1-2/Aldous_1-2.dece.sha256
xz -d dece/Aldous_1-2/Aldous_1-2.dece.xzEither way you end up with dece/Aldous_1-2/Aldous_1-2.dece, ready for
semsage uplink Aldous_1-2.dece. How you get there is up to how fast your
computer is and how much time you have to spend.
Assumes a clean machine with sudo. Every step below is needed the first time.
# 1. Get the source tree.
git clone <your-fork-url> semsage
cd semsage
# 2. Install the full toolchain: system packages, llama.cpp, and libsplinter.# Interactive — it asks before each sudo step, and takes a few minutes.
util/install_splinter
# 3. Fetch the embedding model (nomic-embed-text v1.5 -> gguf/nomic.gguf).
util/install_nomic
# 4. Build the scoring gateway (-> app/inf/dece-semsaged).
make -C app/inf
# 5. Wire semsage into your systemd user session. Nothing is started yet.
util/semsage install
# 6. Optional: put semsage on your PATH (it still finds its install root).
ln -s "$PWD/util/semsage"$HOME/.local/bin/semsage
## 7 - GO GET ALDOUS (train it or download it)## 8. Bring Aldous up. If dece/ is empty, see "Training or obtaining Aldous"
semsage uplink Aldous_1-2.dece # raise the shared-memory bus
semsage enable Aldous_1-2.dece # start on login
semsage start Aldous_1-2.dece # start now
semsage status Aldous_1-2.dece
# 8. Explore in the browser
semsage visualizeUse the Github facilities for now. The company behind Aldous is in the midst of a rather large launch on their end; We'll have a Discord server up as soon as we can responsibly keep an eye on it.
Github is the best place for things to not get lost right now.
Aldous is Free Software under the terms of the Apache 2 software license, where applicable, or CC-0 where the content is prosaic, not code, and not a prompt that results in the creation of code. The vector substrate and file system backing Aldous, which is named Splinter, is also free software under the terms of the Apache 2 software license.
Unless otherwise explicitly marked, supporting code, scripts and utilities are released under the terms of the MIT software license, with all documentation and other creatives under the terms of CC-0.
Aldous comes with no warranty or guarantee of suitability for any purpose.
Aldous does not run on its own. It needs somewhere to keep vectors, key-value state, relation graphs, epoch counters, atomics, and an embedded scripting runtime, all in a footprint small enough to sit on whatever hardware it is handed. That substrate is Splinter (Post, 2026), a lock-free shared-memory manifold that holds every one of those things with no database server and no socket in the path. To Splinter, DECE is only one pose it can hold: the engine that does Aldous' emotional scoring is, from the substrate's point of view, a single arrangement of slots and vectors among the many it could carry. That generality is why Aldous can be as light as it is. None of the machinery it leans on had to be built into it.
The same substrate is also why Aldous is not bounded by the hardware it can train on. Because a governing process can read the same physical memory an inference engine writes into, the observation gap that forces most oversight to happen after the fact closes as a property of the address space rather than as a matter of policy. On larger hardware, that is what lets Aldous observe generation while it is still in flight, at a latency and scale the single-machine framing understates. We link to the Splinter thesis rather than restate it here, but the ceiling it raises for this kind of work sits well above the modest hardware floor it advertises.
Aldous does not introduce new math; it reimagines settled, peer-reviewed ideas and aims them at a problem they were not originally built for: fast, honest emotional and intent telemetry. Its scoring rule is a variance-scaled distance from a specimen to a concept's centroid, which is the diagonal case of the Mahalanobis distance (Mahalanobis, 1936) and a close relative of the diagonal discriminant classifiers and nearest-shrunken-centroid methods that statisticians refined for high-dimensional data (Fisher, 1936; Dudoit et al., 2002; Tibshirani et al., 2002; Bickel & Levina, 2004). What Aldous adds is the embedder: it computes those centroids and variances over independently embedded phrases rather than raw features, which is the same technique the few-shot learning community applied with matching, prototypical, and Gaussian-prototypical networks (Vinyals et al., 2016; Snell et al., 2017; Fort, 2017).
We don't state the lineage defensively other than to state that Aldous' design is based on very grounded, accepted, published and peer-reviewed research. Aldous' returned responses are the result of measurements that other people have studied and validated independently across decades. We aren't changing measurements; we're just applying them to a different, perhaps unconventional, class of problem.
Aldous' most exciting trust & safety features are also not new art: Latent Concept Erasure is the smallest, single-direction case of linear concept erasure, a technique the interpretability community has developed with real rigor (Ravfogel et al., 2020, 2022; Belrose et al., 2023). Aldous applies it to re-scoring rather than debiasing, but the geometry is theirs. Furthermore, the tolerance harness that gates every release is behavioral testing in the tradition of CheckList (Ribeiro et al., 2020), carried into trust and safety, where reproducible receipts matter the most.
Their work matters every bit as much as Aldous' core mission to remain transparent, deterministic and auditable. We are glad to build in the open on foundations that were shared with so much passion, care and respect for the advancement of our collective sum of knowledge. Their work is the knowledge that lets us explain why Aldous works, which is why its uses far exceed its novelty.
The references below are here so that anyone can follow a claim back to its source.
Bickel, P. J., & Levina, E. (2004). Some theory for Fisher's linear discriminant function, "naive Bayes," and some alternatives when there are many more variables than observations. Bernoulli, 10(6), 989–1010.
Dudoit, S., Fridlyand, J., & Speed, T. P. (2002). Comparison of discrimination methods for the classification of tumors using gene expression data. Journal of the American Statistical Association, 97(457), 77–87.
Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7(2), 179–188. (Journal now published as Annals of Human Genetics.)
Fort, S. (2017). Gaussian prototypical networks for few-shot learning on Omniglot. arXiv:1708.02735. Bayesian Deep Learning Workshop, NIPS 2017.
Mahalanobis, P. C. (1936). On the generalised distance in statistics. Proceedings of the National Institute of Sciences of India, 2, 49–55.
Snell, J., Swersky, K., & Zemel, R. (2017). Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 30. arXiv:1703.05175.
Tibshirani, R., Hastie, T., Narasimhan, B., & Chu, G. (2002). Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proceedings of the National Academy of Sciences (PNAS), 99(10), 6567–6572.
Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., & Wierstra, D. (2016). Matching networks for one shot learning. Advances in Neural Information Processing Systems (NeurIPS) 29, 3630–3638. arXiv:1606.04080.
Belrose, N., Schneider-Joseph, D., Ravfogel, S., Cotterell, R., Raff, E., & Biderman, S. (2023). LEACE: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems (NeurIPS) 36. arXiv:2306.03819. Code: github.com/EleutherAI/concept-erasure.
Ravfogel, S., Elazar, Y., Gonen, H., Twiton, M., & Goldberg, Y. (2020). Null it out: Guarding protected attributes by iterative nullspace projection. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), 7237–7256. arXiv:2004.07667.
Ravfogel, S., Twiton, M., Goldberg, Y., & Cotterell, R. (2022). Linear adversarial concept erasure. Proceedings of the 39th International Conference on Machine Learning (ICML), PMLR 162, 18400–18421.
Ribeiro, M. T., Wu, T., Guestrin, C., & Singh, S. (2020). Beyond accuracy: Behavioral testing of NLP models with CheckList. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), 4902–4912. arXiv:2005.04118.
Post, Timothy L. (2026). Splinter: A Lock-Free Shared-Memory Substrate For Tightly-Coupled Inference And Governance. Open Source Vector Substrate @splinterhq/libsplinter (Github). https://splinterhq.github.io/splinter_thesis.pdf (Thesis).
Thank you for reading this far, and caring about citations! 🎉🤝🧠📐💠💻✨🚀