Provenance

What It Is

Provenance is the generating path behind a value — where a price, a conclusion, a confidence, a need, or a task actually came from — and the law is that the path determines the evidential weight. Two identical surface readings can warrant completely different posteriors depending on how they were generated, which means a number is never self-certifying, and neither is a belief. The audit is one question, and it is the anti-hallucination move: how do I know this?

The founding catch is documented in coherence is not evidence: a confident claim ("there are no papers about agent convergence") arrived mid-thought wearing the uniform of knowledge, and the only thing that exposed it was the provenance audit —

"How do I know this? I haven't checked arxiv. This is a hallucination coming from lack of evidence. I have not sampled."

That article owns the promotion policy built on the catch; this one owns the general law. "Generated or retrieved" is one instance of the provenance question. The full question applies to any value from any source — market prices, your own cached conclusions, your confidence, your product ideas, the artifacts your agents produce — and in every domain it does the same work: it recovers the generating process that the surface reading has projected away.

A Price Is a Projection

A market price looks like a property of the asset. It is a lossy projection of a generating process — who traded, under what constraints, with what information — and the number alone does not tell you which process produced it.

"if you look at the market price of something and try to bet against it, a naive thing is to think that market price is just one number, but a more stat-mech / computational approach hints at the reality-contact distance and provenance of that number. Almost like the Monty Hall problem."

Monty Hall is the canonical proof that provenance moves posteriors while the observation stays fixed. After the host opens a door you see the same thing no matter how the door was chosen: one open door, one goat. But if the host was constrained to reveal a goat, switching wins two-thirds of the time, and if he opened a random door that merely happened to hide a goat, switching is even odds. Identical observation, different generating process, different correct belief. The information is not in the number; it is in the path.

Provenance Edge

The betting consequence follows directly: edge is provenance edge. Every number has a reality-contact distance — the hops between it and hard settlement, the measurement or transaction that would pin it to the territory. Find values whose distance is worse than it looks, where the market treats a many-hop projection as a settled fact and you have a shorter path to the underlying reality, and the mispricing is yours. The naive bettor prices the number; the provenance bettor prices the process that emitted it.

A Conclusion Is a Cached Value

The same law applies inside your own head, because human memory is a lossy cache with a specific bias: it prunes derivations and keeps conclusions. Gist survives, verbatim decays. The reasoning is the function; the conclusion is one memoized return value.

"it is weird that I care so much about the provenance of my thoughts / ideas and I always try to write the idea because I find myself pruning the reasoning and keeping only the conclusion (is this normal in humans??!?)"

For a mind that rederives, cache-only memory is a real defect. Keep only the cache and you cannot invalidate it when inputs change, cannot recompute it, and cannot even tell whether it was ever computed or merely absorbed. The fix is architectural: write at the moment of thought, which is the only interception point before the lossy write happens, and relocate the lossless derivation to an external substrate so the biological cache can compress freely. Memory is the substrate gives the storage side the same instruction from the other end — store raw, defer compression — because write-time extraction is compressing against a question you don't know yet, and "the detail you need was deemed non-salient months ago." A derivation is exactly such a detail.

Authenticating Your Own Thoughts

The cache has no source column, and that is a security hole. Epistemic contamination establishes that there is no read-only mode: every exposure is a write, and integration happens before awareness. An absorbed framing therefore enters memory through the same door as a generated insight, and once the derivation is pruned, the two are indistinguishable. A viral soundbite can sit in your head wearing your voice.

Provenance-of-thought is the authentication. In signal theory's vocabulary, the alpha/beta distinction is itself a provenance taxonomy — alpha signals are defined by their generating source (your core system) and beta signals by theirs (the market, the defense) — so tracking where a thought came from is how alpha gets authenticated in practice. An un-sourced conclusion in your own head is a price with bad provenance, and it deserves the same distrust as the number. The golden orb is the generative source this authentication protects: provenance is how you verify a thought actually came from the orb rather than from something that flowed past in the feed.

Confidence Has Provenance

"Instead of confidence being a mystical feeling or identity claim, I can point to why I think I'm right ... When confidence is real, it has provenance."

The move takes confidence exterior. As an identity claim ("I'm usually right about this") confidence is unauditable and unfalsifiable. As a pointable object it decomposes into evidence, mechanism, causal model, and calibration history — and once it decomposes, high confidence can be kept without being suppressed or moralized, because it has become testable. The audit is the same single question in a new domain: this feeling of being right — can I point?

Chain of Provenance for Ideas

Product ideas are values too, and their provenance is the need they were generated from. Will's documented failure mode is the idea with a dirty chain — the excitement to build something that came from no need, exposed by a single question:

"For that thing to not come from any need — so when I ask myself 'would I actually use this,' I realize I just wanted to build it."

The doctrine that came out of it makes clean provenance a build precondition:

"It's better to clear out the chain of provenance such that it is clean… If you think about a specific person, you can think of a specific need that they will want from you and the specific agreement you can get from them."

A clean chain runs person → need → agreement → build, and every link is checkable, because the person at the head of the chain is real and askable. (The simulated other covers why "a specific person" is load-bearing: "the market" is a rendering, while a named person can be asked.) A dirty chain runs excitement → idea → retrofitted need, and it produces the same pitch deck. Identical surface artifact, different generating path, different survival odds — the fantasy collapses at exactly the first contact the chain never contained.

Provenance Rules for Agents

Working alongside generators makes the law operational, because a human-AI system produces artifacts continuously and the single most important bit about each one is which side generated it. Two standing rules from one documented human-AI system, both installed after failures, are provenance engineering:

• Captures-only boards. The first attempt at a life-management board dumped dozens of AI-extracted issues from transcript mining onto it, and every one was rejected — "fake / not mine / over-eager." The rule that survived: AI-extracted items go to a review surface, never straight onto the board, and the board holds only the operator's own captures. The board's authority over attention depends on every item carrying the provenance "I put this here."
• Exact-text voice approval. An agent shipped rewritten site copy and drafted an essay in AI register under the operator's name, and both read as embarrassing. The standing rule since: nothing publishes in a person's name or voice without their explicit approval of the exact text. A sentence carrying a byline is a value whose provenance readers price whether they know it or not.

The general form: erase the generated-or-captured bit and the substrate's trust collapses, because a record that cannot be audited for provenance has to be re-verified from scratch, which defeats the reason for keeping records. Provenance is the type system of human-AI collaboration — it is what lets downstream consumers, including future-you, know what weight each artifact can bear.

Failure Modes

Integration with the Mechanistic Framework

Connection to Coherence Is Not Evidence

The founding instance. The provenance audit is Trigger 1 of that article's promotion gate — generated or retrieved — applied to belief promotion. This article generalizes the audit to any value in any domain.

Connection to Epistemic Contamination

The ingestion mechanism. Contamination explains how bad-provenance content gets in: no read-only mode, integration before awareness, no source column on the write. Quarantine is provenance protection at the input; authentication is the after-the-fact audit.

Connection to Signal Theory

Alpha/beta is a provenance taxonomy over signals: the classification is entirely about generating source, not content. Provenance-of-thought is how alpha gets authenticated in practice.

Connection to Memory Is the Substrate

The substrate is what makes provenance recoverable. Store raw and the generating path survives to be audited later; compress at write time and provenance is the first thing deleted.

Connection to The Simulated Other

The sibling page. A rendering is a belief about a person with generated provenance, and the tripwire sentence "I already know what they're going to say" is a provenance flag.

See Also

• Coherence Is Not Evidence — the founding instance: the audit as a belief-promotion gate
• Epistemic Contamination — how bad-provenance content gets ingested: no read-only mode
• Signal Theory — alpha/beta as a provenance taxonomy; authentication of your own signal
• Memory Is the Substrate — store raw so the generating path survives
• Golden Orb — the generative source that provenance-of-thought authenticates
• The Simulated Other — the sibling law: renderings of people as generated beliefs
• Reality Contact — settlement: the territory end of every provenance chain
• Legibility — sibling property: a value with visible provenance is one form of legible state

Core Principle: The generating path behind a value determines how much to trust it — identical surface readings warrant different posteriors depending on where they came from, and Monty Hall is the proof that the path moves the odds while the observation stays fixed. The law runs everywhere: a market number is a projection whose reality-contact distance can be priced (edge is provenance edge), a conclusion is a cached value that cannot be invalidated once its derivation is pruned, an un-sourced conviction is a price with bad provenance that may be an absorbed soundbite wearing your voice, real confidence is pointable rather than mystical, a product idea needs a clean chain from a real person's real need, and a human-AI substrate stays trustworthy only if every artifact records which side generated it. The audit costs one question: how do I know this?

Every value you hold arrived by some path. Skipping the audit means pricing a measurement and a hallucination the same.