The workspace no one sees
Every time you ask an AI a question, it builds a rich private understanding of what you meant, uses it once, and lets it go.
You type a few words of shorthand. Inside the model, those words pass through dozens of layers, and each layer adds context to a shared working space that researchers call the residual stream. By the end of that pass, “skinny keyboard tie, AI interview” has become something much fuller: a question about style, culture fit and self-awareness. The model then speaks its answer one word-piece at a time, choosing each from a spread of possibilities.
That workspace is remarkable, and it is temporary. It is rebuilt from scratch on every step and kept nowhere afterward. It is not a filing cabinet you can reopen next year. When the answer arrives, the path that produced it is already gone: which sources were in view, which readings were close runners-up, where a word quietly changed meaning on its way between fields.
For casual questions, that is fine. For evidence, research and history, it is the whole problem. An answer you cannot trace is an answer you have to take on faith, and fluent prose now costs nothing to produce.
The glass pipeline
We don’t try to turn the model’s private workspace into an archive. We build a transparent layer around it, and we record every handoff that the layer can see.
We call it the glass pipeline because you can look through it at any joint. Sources come in and are fingerprinted. Terms keep the meaning their own field gave them. Links between fields are labeled for what they are: an exact match, a partial overlap, an analogy or a disagreement. The model does its work in the middle, and what goes in and comes out is logged. What leaves is not just an answer but an answer with its evidence, its grade and its open questions.
The model’s inner workspace still shapes the answer. But because the glass pipeline feeds it relevant, labeled evidence on every run, what matters no longer depends on the model remembering it. The durable knowledge lives in the pipeline, where it can be inspected, challenged and corrected.
Lenses
A lens is any point where we can observe and record what happened. Each one shows something real, and each one has limits we state out loud.
| Lens | What it shows | What it can’t tell you |
|---|---|---|
| Source fingerprint | Exactly which document, recording or dataset went in, unaltered | Whether the source itself is accurate |
| Term and relation map | How a word is used in each field, and how fields relate | That two fields mean the same thing, unless evidence shows it |
| Context record | What the model was shown on a given run | What it drew on from its own training |
| Repeated sampling | Whether the model gives consistent answers when asked again | Whether a consistent answer is true |
| Token probabilities | How confident the model was in each next word, where the provider exposes them | The probability that a claim is correct |
| Witness and Adversary | Which objections were raised, and which held | Anything neither side thought to test |
| Human decision | Who accepted, rejected or left a claim open, and why | A final verdict; later evidence can reopen it |
With open-weights models, one more lens becomes possible: looking inside the residual stream itself, using research tools that detect concepts in a model’s activations. We treat that as a promising experiment, not a feature. Even a perfect inner reading would not replace a source you can check.
Two tests, one ledger
A research question and a provenance question look different, but in the glass pipeline they go through the same test and land in the same ledger.
A research stress test asks whether an idea holds up: does the evidence support it, and does anything in another field rule it out? A provenance stress test asks whether a record holds up: where did it come from, what happened to it along the way, and did a copy get mistaken for a confirmation? Both are run the same way. One process builds the strongest honest case, a second attacks it, and a person decides what the evidence warrants.
What makes this powerful is that the results accumulate. Every claim carries a grade from established through contested to speculative, along with the history of how it got there. When new evidence arrives, the grade moves, and the old reasoning stays on the record rather than being overwritten. A source that failed a provenance check lowers the standing of every research claim that leaned on it. A research finding that survives cross-examination strengthens the records it relied on.
Over time, the ledger gets sharper in two directions at once. The evidence gets better sorted, and the system learns how often its own grades held up.
Each test makes the next one more reliable.
One schema, several projects
The glass pipeline is not a single product. It is the shared continuity and enrichment schema underneath Interveil’s flagship work, powered by the Interveil Engine.
- Chronoaesthesis is our open research platform, where specialists, students, enthusiasts and philosophers bring questions into the same room. The glass pipeline lets a physicist and a curious teenager be judged by the same standard: the evidence.
- Glosa applies the same schema to aesthetic enrichment, so creative and interpretive work keeps a record of its sources and influences.
- Forensic and archival work uses it to reconstruct contested records, such as a damaged recording, without hiding where a restoration required a judgment.
Because every project writes to the same schema, what one learns can inform the others. A provenance lesson from archival work sharpens a research grade on Chronoaesthesis, and every link along the way stays labeled and traceable.
Your data
A system built on provenance has to be honest about its own. These are our commitments.
- Opt-in only. We collect personal data only when you choose to share it.
- Never exploited. Your data is not sold, not used for advertising, and not traded for access to anyone’s services.
- Kept in our ecosystem. What you contribute is stored and processed within Interveil’s own systems.
- Named on the receipt. When a request is processed by an outside model provider, the receipt says which one. Our direction is to run open-weights models on infrastructure we control, so that step stays in house too.
- Yours to withdraw. You can ask for your personal data to be removed. Scientific records you chose to publish keep their history, but your identity can be separated from them.
We hold ourselves to the same test as everything else in the pipeline: these commitments should be checkable, not just stated.
Hearing more faithfully
The goal was never a perfectly clean signal. Human records were never made under perfect conditions, and neither are AI answers. The goal is to hear more faithfully: to know which parts of an answer are solid, which are inferences, and where someone can return to the source when an interpretation no longer holds.
The model’s workspace will stay private. Everything around it doesn’t have to. That is the glass pipeline, and it is the direction Interveil is building toward.
— D.
Further reading
- A Mathematical Framework for Transformer Circuits — Anthropic, on the residual stream
- Toy Models of Superposition — Anthropic, on how models pack concepts into shared space
- PROV Overview — W3C’s standard for describing provenance
- C2PA Technical Specification — content credentials for media