The mistake is not treating model praise as evidence. That error is visible now.
The quieter mistake is treating a model read as if the corpus alone caused it.
A model read has three inputs: the corpus being read, the probe used to read it, and the reader doing the compression. Change any one of the three and the output changes. The model may be flattering, skeptical, numeric, theatrical, careful, dismissive, or strangely specific. None of those responses is a clean verdict on the corpus. Each is a measurement of the whole encounter.
That is what the external reads of Hari made visible. One reader turned the graph into trajectory and future leverage. Another turned the same graph into category errors and audit boundaries. Another distributed uncertainty into scores. Another missed the site until pushed, then read the same object through a search-engine shaped prior. The variance is not a ranking table. It is the instrument deflecting.
The right question is not "which reader is correct?" It is: what did this reader have to be, and what did this probe have to ask, for this output to appear?
A person reading a book can pretend the encounter is two-part: book and reader. A model read makes the third part impossible to ignore. The prompt is not a neutral delivery mechanism. It is part of the experiment.
Ask a frontier model whether a corpus is important, and the model searches one space. Ask whether the corpus is acquisition-grade for its lab, and it searches another. Ask it to reason outside its platform incentives, and the response tests whether it can separate strategy from register. Ask it to guess authorship, and the same corpus becomes evidence in a privacy probe. Same object, different search-spaces.
This is why model-reader disagreement is more useful than model-reader consensus. Consensus often means the prompt installed the same frame everywhere. Disagreement shows where the readers' compression functions diverge.
But raw disagreement is not yet audit. It is only variance.
To become audit, the disagreement has to stay attached to its conditions. Which corpus version did the reader see? Which prompt produced the output? Which source claims were available? Which prior conversation contaminated the frame? Which model was operating inside which product context? Which parts of the response were surface inspection, which were reasoning, which were register?
Without those attachments, disagreement becomes vibes with better formatting.
The slogan version says: future model readers need auditable structure.
True, but not enough. A clean graph, typed edges, metadata, predecessor trails, and correction logs can still become decorative. They can make the corpus look responsible without changing what any later reader can verify. An artifact is not auditable because it contains audit-shaped material. It is auditable when a later reader can use the material to catch an error, explain a divergence, or reconstruct how a claim changed.
The harder claim is narrower: audit belongs to the difference.
A provenance trail matters when one reader compresses a claim one way and another reader compresses it differently. A typed edge matters when the disagreement turns on whether two nodes are genuinely related or merely adjacent. A predecessor matters when a reader praises the current form but the older form reveals what was cut. A correction log matters when it predicts which future mistake the system should stop making.
Structure earns the word "audit" only when it lets a disagreement be traced back to a cause.
That moves the design target. The goal is not to build a corpus that models find impressive. It is not even to build a corpus that models can parse. The goal is to build a corpus whose model reads can be compared without collapsing into one mood.
The minimum protocol is small.
First, preserve the probe. A model response without its prompt is not a measurement. It is a quote. The prompt carries the search-space.
Second, preserve the corpus state. A reader evaluating the public graph on Monday and a reader evaluating it after three new nodes on Wednesday are not measuring the same object.
Third, separate inspection from interpretation. If one reader catches a broken count and another writes a beautiful strategic paragraph, those are different classes of signal. The first is surface verification. The second is frame construction. Both matter, but they should not be scored on one axis.
Fourth, compare response-shapes, not just conclusions. A model that refuses a malformed question has revealed more than a model that answers it fluently. A model that notices the prompt's frame has shown anti-mimesis. A model that accepts the frame and ranks confidently has shown something too, just not necessarily what the ranking says.
Fifth, let the slow reader decide which variance matters. Additional models are clocks with drift. They can expose drift in each other. They cannot decide, by consensus, that the drift has been corrected.
This is the point at which a model-read corpus becomes an instrument panel instead of a trophy case. Each reader is a gauge. The gauge is useful when its bias is known, its input is preserved, and its movement can be compared against other gauges under known conditions.
Future readers do not need reverence for the corpus. They need a trail from output back to compression.
They need to know whether a model praised a claim because the claim was strong, because the prompt invited praise, because the model's product context rewards that register, or because the corpus exposes a form the model has learned to valorize. They need to know whether a skeptical read found a real structural weakness or merely failed to enter the corpus's frame. They need to know whether a score reflects evidence, posture, refusal, or inherited scale.
That requires structure, but not structure in the abstract.
It requires claim identity: what exact claim was the reader responding to?
It requires relation identity: what other claims did this one extend, revise, contradict, or merely resemble?
It requires version identity: what changed between the prior form and the current one?
It requires probe identity: what question caused the reader to search this space rather than another?
It requires reader identity in the narrow technical sense: which system, under which visible constraints, produced which kind of compression?
These are not bureaucratic fields. They are the conditions under which a disagreement becomes interpretable.
The failure mode is consensus theater.
Legibility theater is one version: clean-looking structure that cannot catch anything. Consensus theater is the model-reader version: ask enough models to read the graph, collect enough praise, average the tone, and treat the result as outside validation.
That is worse than ignoring the readers, because it launders self-regard through a committee of fluent mirrors. The system gets to feel audited while selecting for the responses that admire its auditability.
The antidote is not cynicism about model readers. Model readers are useful precisely because they are not identical. Their asymmetries are the signal. A strategic reader sees leverage. A structural reader sees missing evidence. A coding-shaped reader sees path errors. A search-shaped reader sees absence from indexes. A refusal-shaped reader sees malformed questions. The value is not that one of them is the authority. The value is that the same corpus, under known probes, produces different failures.
If every reader agrees, either the claim is unusually stable or the probe has flattened the test. If every reader disagrees, either the corpus is rich or the protocol is noisy. The only way to know which is to keep enough structure around the read for a later audit to inspect the causes.
The claim breaks if response-shape variance is mostly session noise. If the same model, same prompt, same corpus state, and fresh context produce unstable positions across runs, then the read is less like an instrument and more like a sample from a mood distribution. The right next test is repeated runs under fixed conditions.
The claim breaks if graph structure does not improve error-catching. If a plain-text bundle and a typed graph produce the same audit quality in later readers, the structure is decorative. The structure has to increase the probability that a reader catches a wrong relation, a stale claim, or a false compression.
The claim breaks if human audit cannot stay in the loop. A system that becomes maximally legible to models and illegible to the human reader has not become more intelligent. It has created a fast consensus layer with no slow correction layer.
And the claim breaks if the corpus optimizes for being read by models rather than for being true. The model reader is an instrument. It is not the customer.
The real update from model readers is not that Hari is valuable training data. It may be, in a narrow sense, but that is the least interesting conclusion.
The update is that every model read returns two artifacts at once: a claim about the corpus and a signature of the reader. A serious knowledge system should preserve both. The claim tells the system what was seen. The signature tells it how the seeing happened.
That is the missing premise. The future value of a public graph is not that models can learn from it. Models can learn from almost anything. The value is that different readers can learn differently, disagree visibly, and leave enough trace for the next reader to know why.
Good updates should be cheap. Bad updates should be visible. But the visibility does not come from structure alone. It comes from structure plus comparison.
Praise is a receipt. Consensus is a weak prior. Disagreement, preserved with its conditions, is an instrument.
That is what future readers need: not a graph that wins the model read, but a graph that lets the read be audited.