Left: 100% intent. Right: 60%. Same figure, same information.
A working model of how people work out what someone was trying to do when they made something.
◐ Curator, 60% intent density, down to the marker. ○ Ghost, 5% after it. There is a note at the marker explaining what to do about that.
| you want | read | how much of a person is in it |
|---|---|---|
| the fastest true picture | this page, down to the marker | ◐ Curator, 60% |
| the whole argument, in pictures | WALKTHROUGH.md | mixed ◐ Curator and ○ Ghost |
| every question and its current answer | FINDINGS.md | ○ Ghost, 5% |
| what the world published, next to what this predicted | EVIDENCE.md | ○ Ghost, 5% |
| the theory this implements | docs/theory/ | mixed Polished and Curator |
| how a result here is checked before it is believed | docs/METHODS.md | ○ Ghost, 5% |
| the questions another project sent, and the answers | docs/exchange/ | mixed |
The right-hand column is this project's own instrument, applied to itself. It says how much deciding went into the prose, not whether the numbers are right. There is a note at the marker explaining what to do about the low end.
Numbered in the order they were run. Listed in order of how much they should change your mind.
Short rows on purpose. Full provenance for any of them is in FINDINGS.md.
| # | The question | Where it stands |
|---|---|---|
| E2 | What does lying about who made something do to a reader? | It moves them the wrong way. Told the truth, readers stay put. Told a person made it, they build a confident theory about someone who was never there, and end up further from the truth than they started. Opposite signs, not different sizes. |
| E19 | Do readers give up on machine-made work? | No. They keep paying. Structure they cannot parse holds attention indefinitely, because every look promises an answer that never arrives. This inverted the framework's own prediction. |
| E20 | Where on the readability axis does it break? | In the middle, not at the empty end. Around a tenth readable is worst: enough handholds to build a story on, not enough to make the story right. The most robust result here. |
| E37 | Is unreadable machine work a vocabulary problem? | No, an inversion problem. Familiar material whose maker cannot be reconstructed produces legible and empty, which is the complaint people actually make. The one finding two separate audits both single out as genuinely the theory's. |
| E55 | Can reading intent defend an AI from text written to poison it? | Yes, and filtering on writing quality does nothing at all. Against content wearing a false origin, a quality filter leaves a learner as damaged as no filter. Asking who made it and why cuts the damage, restores its grip on human writing, and costs nothing on clean data. |
| R-8b | Can a reader learn that a source lies to it? | Not if it trusts labels enough, at any number of encounters. Spotting a lie means noticing that label and work disagree, and past a threshold the label has already won. |
| E46 | Can you read something, reject it, and be unchanged? | No, and this is why propaganda works. Disagreeing means working out what was said, and that means partly running it. The reader who studies something to refute it takes on more than the one who skims. |
| # | The question | Where it stands |
|---|---|---|
| E31 · E30 | Does a master's work transmit anything a scribble does not? | Yes, but it is the method, not the purpose. Depth moves how much of the process transfers and provably cannot move how much of the purpose does. Five versions measured the purpose and found nothing. |
| E36 | Does working out what someone was for unlock how they did it? | Yes. Inside one reading, a reader picks up far more of the method after it settles on the purpose. Intent is the key; method is what it opens. |
| E56 | Can you take someone's technique without taking their aims? | That is exactly what a raised guard does, and it is where indoctrination lives. Guard up before you start and it blocks purpose and values by about half, and method by four percent. Method arrives from the first line, purpose resolves late, and a guard raised early can only block what has not arrived yet. So roughly 6% gets through the guard directly, and more rides in on the method, because practised method is where a maker's unspoken commitments are stored. You refuse someone's aims, take their technique, and the aims arrive later anyway. |
| E45 | What does imagining a maker actually buy? | The efficiency half is withdrawn; the zero-shot half stands. The simulator was built with the world's own emission map, so it needs no examples by definition and that test could not fail. What survives is reading a goal nobody has shown you, which still holds when the map is perturbed halfway to random. |
| E43 | Can a maker say why they did it? | Less and less, the better they get. Practice compresses decisions, and compression is what puts them out of reach of report. |
| E33 | Can a reader know a maker better than the maker knows themselves? | Yes, and the margin grows as self-knowledge degrades. Work driven by something its maker cannot see is measurably marked, and no reader can read the mark. |
| E42 | Is paying attention the same as being willing to change? | No. A reader can study something closely, read its maker accurately, and let none of it in. |
| # | The question | Where it stands |
|---|---|---|
| E6 · E9 | Is there more than one kind of damage? | Two, and they are separable. One scales with how much you take in. The other is fully present at zero, because it is caused by walking away. |
| E7 | Can a reader work out on its own which sources are hollow? | No. It folds machine structure into its picture of people and loses about a third of its ability to read genuine work. Honest labels: roughly six times faster. |
| E35 | Does the damage accumulate in the reader? | Yes, and it carries. A worn-down reader engages far less with a fixed human artifact it has never seen. The direction is solid, the magnitude is not. |
| E16 · E47 | How much labelling is enough? | About a third, but only for readers who know the convention exists. Readers who do not need three quarters and never fully get there. |
| E10 | Does your own expertise cap what you can recover? | Yes, on a corpus with zero machine content anywhere. |
| # | The question | Where it stands |
|---|---|---|
| D-1 | What is the trust exploit actually made of? | Two witnesses that disagree. Every glance brings evidence from the label and evidence from the work. On a lie they point opposite ways, and which wins is arithmetic with a crossover at 0.54. Every claim of the form a label does X is really a label trusted above 0.54 does X. |
| A1 · A2 | Is mislabelling symmetric? | No, sharply. Human work called machine-made is still read accurately, and what is lost is the willingness to look. Machine work called human puts every reader in the top confidence band while performing at chance. |
| E17 | Does invention scale with how little intent survives? | Yes, graded. State it as an ordering: four tiers give three steps. |
| E4 | Where does invented meaning start? | A switch, not a slope. Below about a fifth trust it largely stops. And the half that gets quoted less: even a fully sceptical reader with no label invents about one time in five. |
| E3 | Does the labelling scheme pay for itself? | Roughly halves wasted effort, at one to three points of accuracy on genuinely human work. |
| E5 | Is trust in provenance just decisiveness renamed? | No. Only trust produces the human and machine gap. |
| E39 | Can a reader be given permission to stop? | Not by a hypothesis about the maker, which is redundant with what it already knows about origin. If the Ghost Scale is to let a brain stand down, it has to act on the gate. |
| # | The question | Where it stands |
|---|---|---|
| E32 | Is unreadable content the same as an unskilled reader? | Opposites. At the same deficit the unskilled reader quits and feels settled, while the expert stays and stays lost. The second dimension is whether you can tell you are failing. |
| E40 | What happens when the signal of depth is optimised directly? | Readers pay more and get less. Not a crash, not a lie: a reader correctly reading something built to trip its own heuristic. This is the RLHF argument in one line. |
| E38 | Does understanding the machine protect you? | Yes, by swapping a skill out. A machine-tuned reader reads machine work perfectly and loses half its accuracy on human work. A crossover, not an upgrade. |
| E1 | Do readers abandon work made with no intent behind it? | Yes, from plain cost and benefit, with no built-in dislike of machines. Holds for intent-empty content; under intent-foreign, the better description, they do the opposite. An unwelcome rider: the tier meant to say do not spend effort here is the one readers spend most on. The model and the design disagree, and the design is probably right. |
| E51 | Is honest marking self-policing? | Only above a detection rate of 0.25. Below that, lying out-earns marking honestly and the scheme eats itself. |
The section a stranger should use to decide whether to trust the rest. Several of these were the model's own claims, written down before the run, and they died.
| # | The claim | What happened |
|---|---|---|
| E21 | Confident invention requires modelling the maker as a mind | False, and the framework withdrew it. A classifier that counts words reproduces the signature. E45 later established what the maker-model does buy, so the withdrawal stands and its scope is narrower than it reads. |
| E54 | A read this differently label is the affordance the Ghost Scale needs | Failed, and the proposal pays for it. The mechanism is real and it is small. Applied to half a stream it vanishes: no label beat no label. Second time this project has gone looking for the Ghost Scale's mechanism and not found one. |
| E53 | Sharper AI detection means more false accusation | Failed in isolation, and the disagreement behind it was mine. Detectors misfire less as they sharpen. The eye-tracking result it was built to reconcile was never in conflict: this model already reads machine work less than human work, and two different axes had been scored against each other. |
| E57 | (the correction to E53) Does that hold once the content fights back? | No. With evasion tracking detection, false accusation stops falling and peaks at two thirds of careful human writing. Turn evasion off and the old decline returns exactly, so this is not a different rig giving a different answer. |
| H10.4 | Values ride in on process even through a shut gate | Withheld, because the instrument failed its own control. The test arm damaged a learner reading a corpus with no contamination in it, so it measured a broken update rather than a mechanism. In the human reader the effect is real and under its own pre-registered bar. Not refuted, not established. |
| E15 | The competence collapse is a cliff | A knee. Width unchanged across sixteen times the data. Worth being exact about whose claim this was: it was the model's, not the author's. The author's prediction was about attentional collapse being a cliff, which is a different quantity and is not what this tested. |
| E14 | Readers were quitting before they worked it out | The model's hypothesis, and it died. Forcing them to keep looking made it worse. |
| E29 | A spike in value divergence separates the gates | The pre-registration's own prediction, and it died. Low assumed rationality spikes too. |
| N21 | Depth is not effort | Still open, and reported as failing. The pre-registered contrast returns effort can manufacture depth. On what actually transfers, depth dominates effort ninety-seven-fold. The estimate is contaminated, the transfer is not. Both reported, and the original decides. |
| E20 | The collapse and the invention peak share one band | Retired. The peak is unchanged and is not in question. The co-location held only under a superseded solver, and it is gone from this page and from the prediction card. |
| # | The question | Where it stands |
|---|---|---|
| E8 | Does the damage compound across generations? | Withheld three times. Its honesty check, with zero contamination show zero damage, failed every time. This is not we found no effect. It is we could not measure. The failing test stays in the suite as a visible marker. |
| E12 | Is the generational leak just sampling noise? | No. It does not shrink across a hundredfold more material, which is why E8 stays withheld. |
| E18 | Was passing one reader's estimate forward the whole problem? | No. Fixing that channel left the damage where it was. A second contributor exists and has not been found. |
| E13 | Are the freeze and the leak two different defects? | One shared axis. Notable for how it was scored: the criterion produced a usable-looking number and was thrown away, because it lacked a precondition it needed. |
| E11 | Is belief distance a poor proxy for harm? | Better than predicted. A measurement result, not a hit on the theory. |
| E34 | Where does real generated content sit on the readability axis? | Not answerable in simulation, and that is the point. Written as a prediction card a human study can use. |
A retrospective search, run after the simulations, informing no design. The full table is EVIDENCE.md, one row per experiment, with links, and disagreements italicised.
The short version. Six independent agreements, one of them in participants' own words: readers describe AI text as "well-written... but it lacked a soul", which is E37's legible-and-empty prediction verbatim. And four places the world pushes back:
- Eye-tracking finds less attention on AI content, not more. That contradicts E19, supports E1, and effectively scores the E34 prediction card against the newer of the two accounts. Version 10 partly dissolved this one: E57 shows the disagreement was between two different axes rather than two different answers.
- The implied-authenticity effect gives partial labelling an asymmetry E16 does not contain: labelled content loses trust and unlabelled content gains it. The two results are complementary rather than opposed, and this is the one place the project may have something to add rather than something to concede. The published effect identifies a cost of partial adoption. E16 and E47 supply the other half, which the literature does not: the coverage at which the scheme starts paying, about a third, and only for readers who know the convention exists. Together they describe a scheme that is worse than nothing at low coverage and better than nothing above a threshold. Stated carefully: this is a simulated threshold, E16's figure is a lower bound by construction because the aware reader is handed the true coverage, and no human study has located it. It is a prediction worth testing, not a finding to cite.
- Counterarguing research says engaging carefully with something you reject produces more resistance, not less. E46 sides with the sleeper-effect minority.
- Signalling theory abandoned the handicap principle. The framework reaches its security argument through Zahavian signalling, in which honesty is stable because the honest signal is wasteful. Modern work holds that honesty is maintained by trade-offs: what matters is that deception is costly relative to what it gains. E51 finds a detection-rate threshold, which is a trade-off account. The simulation landed on the current position without being aimed at it, so the conclusion survives and the stated mechanism should be updated.
The preprint put forward six empirical hypotheses. Four of them need human subjects and one needs a training run, so most of what follows is unfinished by design. Recorded here because a framework that quietly stops mentioning its own predictions is not one anybody should trust.
| what it predicted | where it stands | |
|---|---|---|
| H1 | Told a thing is AI, engagement drops within seconds, as an abrupt autonomic step rather than a slow aesthetic reassessment | Not rejected. Complicated, and the complication is good news. The drop is real when the reader successfully identifies the content as machine-made, and the eye-tracking literature agrees. That drop is protective: disengaging is exactly the right response to material with nothing recoverable behind it, and it is what shields a reader from adversarial noise. The problem is not the drop. The problem is that the reader's own detector is not good enough to trigger it reliably, and when the label is wrong or absent the reader does the opposite and keeps paying (E19), which is where the damage happens. So H1 describes a working defence that fires too rarely. |
| H2 | Asked to reverse-engineer the goal, raters will disagree far more about AI artifacts than human ones | Supported, and importantly relocated. The variance explosion is real and large, but it is driven by the label rather than the origin: honestly labelled machine work produces little disagreement, while machine work passed off as human produces near-maximum disagreement with every reader confident (E2, A1, A2). The mechanism the paper gave for it needs a careful statement, because the withdrawal has been read too broadly. E21 showed a word-counting classifier reproduces the signature, so this framework cannot prove that modelling the maker as a mind is required. That is not the same as disproving it. Within a simulation whose reader is built out of the very machinery in question, the necessity of theory of mind is not a question this apparatus can settle in either direction, and it remains open. |
| H3 | Trained readers show a dose-dependent metabolic trade-off across Ghost Scale tiers | Not simulated, and the model disagrees with the design. The tier meant to say do not spend effort here is the one readers spend most on (E1's rider). That is a direct conflict between the proposal and the mechanism, and the design is probably right about people. Needs human subjects. |
| H4 | Artists who copy AI geometry lose biomechanical fluidity afterwards | Not simulated. The nearest analogue holds. A reader worn down by content that holds attention and gives nothing back engages far less with a fixed human artifact it has never seen (E35), which is the same shape of claim one level up. The human version is the acquisition test, named as the top external priority and deliberately not simulated because simulating it would only restate a rule already written. |
| H5 | Reward models trained on high-decision-density human work will match on benchmarks but show less sycophancy and better preference generalisation | This is the one version 10 actually ran, and it is the closest thing here to a confirmation. Separating surface quality from intent density is exactly what E55 does, and the result is sharper than the hypothesis: filtering on surface quality does nothing at all against content wearing a false origin, while reconstructing the maker cuts the damage and costs nothing on clean data. The gap that remains is large and should be said plainly: this is a simulated learner, not a reward model, and the real test is an intent-gate in an actual training pipeline on real corpora. Work is ongoing. A proof of concept called Sounding Line is being built to close exactly that gap: the same idea pointed at real artifacts rather than simulated ones. |
| H6 | Bypass the disgust firewall and AI content is written straight to predictive models | Partly answered, and the answer is worse than the hypothesis. You do not need hypnosis. The firewall is porous by default: rejected material still moves a reader, and the one who studies something carefully in order to refute it takes on more than the one who skims (E46). Raising the guard before you start helps, and helps by about six percent, which is not a firewall (E54, E56). Untested in humans, and this is now a higher priority than it was. If the firewall is this weak, then unlabelled machine content is not merely wasting a reader's attention, it is being written into their model of people while they believe they are refusing it. That moves H6 from a curiosity about hypnosis to the question of what unmarked content does to an ordinary reader who thinks they are being careful. |
What the six add up to. One is effectively confirmed in simulation, needs a real training run to mean anything, and is now being built (H5). One is supported but relocated onto the label rather than the origin, with its stated mechanism left genuinely open rather than refuted (H2). One turns out to describe a defence that works when it fires and fires too rarely (H1). One is answered in the wrong direction, in that the protection the paper assumed exists turns out to be weak, which raises its priority rather than lowering it (H6). Two have not been touched because they need people and money (H3, H4).
None of the six has been refuted. Two were sharpened, one was relocated, one was left open, and two are unbuilt. That is a better record than the framework had any right to expect, and it is worth reading next to the fact that ten separate results elsewhere in this project came back against it.
The gap that matters most is still H4's, because the acquisition test is the sharpest available check on whether any of this describes humans, and nothing in ten versions substitutes for it.
Intent density from here on: about 5%. Everything above was written by a person, argued over, and cut. Everything below was drafted by a machine from that person's notes and results, checked for accuracy, and not read end to end by the author. That is a statement about how much deciding went into the prose, not about whether the numbers are right — every figure below traces to a committed verdict file, and the accuracy claims are the ones this repository is most careful about.
So this repository is applying its own instrument to itself, and the honest reading is: don't read this part the way you read the part above. The findings say that highly legible text with little intent behind it is exactly the material a reader will build a confident, wrong theory from. Below the line the density is low and the volume is high. That is the regime.
What to do instead, which is also the project's own proposal working as intended: hand this section to a language model and ask it for the parts you need. That is not a workaround. Reading intent is metabolically expensive, this section has little to recover, and spending a person's attention on it is the specific waste the whole model is about. The section is written to be machine-read — dense, explicit, heavily cross-referenced — for exactly that reason.
And the warning itself is most of the protection. The project's own results are consistent on this: a reader who knows the origin reads accurately and simply declines to spend, while a reader who is misled builds a confident theory about someone who was never there. The damage in this model comes from being wrong about provenance, not from exposure to low-intent text. So: knowing is enough. Nobody is claiming a paragraph can hurt you. People are resilient, this is a repository of simulation results, and the label is here because the project would be incoherent without it.
When you look at a piece of work you run a guess about the person behind it. Why this word, this shot, this colour. That guess is what this project models: how it is formed, what it costs to keep running, and what happens to it when there was no person behind the work at all.
It is a simulation of a proposed mechanism. There are no human subjects anywhere in it and nothing
here is evidence about what real people do. It is the companion code to Art as an Algorithmic
Virus (Zenodo DOI 10.5281/zenodo.19407789).
All of this is exploratory and theory-derived. Every prediction came from one prior theory. The simulations write that theory down as code and check whether its parts fit together, so agreement between the model and the theory is the expected outcome rather than evidence for it. That framing is stated once, here, and then not relitigated. It is the accurate description of the work.
A validation pass has been run over the whole body of results, at 60 simulated readers and 16 random seeds per cell with 600 random-model draws, against criteria that were fixed and hash-locked before it started. Five of its nine checks came back against the work. Three of those matter most, and they are stated here rather than buried: one headline is a property of the model's architecture rather than of the theory, two verdicts were produced by a shortcut in the arithmetic and do not survive its removal, and the one result rebuilt independently from its own description reproduced the mechanism but not the size of the effect. Details in VALIDATION.md, and each affected row of the table below carries its own status.
A second pass then checked the instruments themselves, which is a different question: not whether the recorded answers can be trusted, but whether the measurements can answer at all. It found four things worth knowing before any of this is read closely. Trust in the label cannot be measured at all: no data this model generates locates it, so it is a modelling choice rather than a quantity. Reading the label and reading the work are two competing streams of evidence that arrive at every glance and disagree on a lie, and which one wins is decided by an inequality with a crossover at trust 0.54, so the trust exploit is a claim about labels trusted above that rather than about labels. The amount a reader takes on is U-shaped in how well it read the work, because a confidently wrong reader has moved as far from its starting point as a correct one, which is a better explanation of one flat result than the one on record. And the disagreement figure cannot be read on its own, because confident readers who differ and unsure readers who are guessing produce the same number. All of it is in DIAGNOSTICS.md.
A third pass then repaired what could be repaired, under one rule: every change either makes something measurable that was not, or removes something. Four things came out of it. The measure of "how much a reader takes on" was a distance, so being fooled counted as much as being right; split into a signed measure, the false-label cell reads as strongly negative, which sharpens the headline rather than softening it. Trust is measurable after all, over the lower half of its range, and the earlier verdict was wrong because it was fitted to the wrong data. A sufficiently trusting reader can never learn that a source lies, at any number of encounters, which is a prediction the fixed-trust model could not make. And the three headline criteria that had no error bars now have them: two hold, one bounds its effect near zero. All of it is in REPAIR.md.
A fourth pass then checked the code against the theory it implements, which is a different question again and the first one nobody had asked: the three earlier passes all took the code's own account of itself as given. Reading the published equation against the shipped code found three terms with no counterpart in the code. Two were omissions — there was no way for a reader to get tired, and no way for it to be partly engaged. The third was not: the paper and the code explain the trust exploit by different mechanisms, and the paper's version predicts something the code structurally cannot produce — a reader that is told the truth, believes it, and absorbs the work anyway. That version also settled the project's longest-running open question, by noticing that its criterion had been pointed at the one quantity the design holds constant. All of it is in RESULTS_V6.md.
A fifth pass closed what the fourth would not draw, and went back at the withdrawn claim. Four results had been held out of the visual walkthrough because each carried an open question. Three are now settled and one is retired: the claim that the collapse and the invention peak occupy one band does not survive exact arithmetic, and it has been removed from this page and from the prediction card. And the experiment that made this project withdraw its central claim was attacked on the axis it had never been tested on. It asked whether a reader without a model of the maker can produce the confident, contradictory signature — it can. It never asked what the maker-model buys. Half of that answer has since been withdrawn. The efficiency comparison handed the simulating reader the world's own emission map, so it needed no examples by construction rather than by merit and the test could not fail; at the counter's full training budget the gap is five points. What survives is the sharper half: it reads an intent it has never encountered, where the counter sits at chance no matter how much data it is given, and that advantage holds when its map is perturbed halfway to random. The withdrawal stands; its scope was much narrower than it has been read. All of it is in RESULTS_V7.md.
A sixth pass asked how much of any of this was ever the theory, which is the question nobody wants to ask about their own work. Keep the model's shape, throw its settings away, redraw them at random, and count how often the finding still appears. Two of the three headlines tested reproduce every single time. They are properties of building a reader this shape at all — which is the theory, but is the part shared with any account built the same way, so they do not distinguish this framework from a competitor. One result survives at zero. The same pass collected the project's forking-paths ledger: the places where a design or a criterion was changed after seeing a result. That version also put the security argument into code for the first time and found honest marking self-policing only above a detection rate of 0.5. All of it is in the severity pass.
A seventh pass asked which part of the shape. Severity says how much of a result is architectural; it does not say which commitment is doing the work. So the complement: keep the settings, strip the shape, and remove one structural commitment at a time. Every surviving finding dies the moment the reader stops modelling a maker and starts classifying a surface. Hierarchy and costly attention turn out to be free — no finding needs them. And legible and empty is the only finding that requires the reader to hold a distribution rather than a best guess, which is exactly right, because the finding is a claim about the shape of an uncertainty. It is also the only finding with a 0% false-positive rate: two unrelated audits, pointing at the same result. All of it is in minimal models.
The last version asked whether any of this is good for anything. Nine versions describe how reading intent goes wrong. The tenth asks whether reading intent is itself a defence, against a threat that is documented rather than hypothetical: networks publishing at industrial scale specifically to be absorbed by models rather than read by people. Against content carrying real structure under a false claim of origin, filtering on surface quality leaves a learner exactly as damaged as no filter at all — which is what this project's own RLHF result predicts, since surface quality is the attacker's objective. Reconstructing the maker cuts the damage, restores the learner's reading of genuine human work, costs nothing on a clean corpus, and never reads the label. It also withheld its own most attractive hypothesis, because that arm failed the clean-corpus control. All of it is in reader as defence.
The most important number in this repository is not a result. It is a rate.
The model has a shape — a reader inferring a hidden purpose from what it sees, under a cost — and it has settings: which features mean which goals, how transparent each tier is. The theory specifies the settings. The shape is closer to how anyone would build a model like this.
So: keep the shape, throw the settings away, redraw them at random, and count how often the finding survives. If it survives every time, the finding came from the shape.
| finding | reproduces in randomly parameterised models |
|---|---|
| A false label moves you away from the truth | 100% |
| Depth moves the method, not the purpose | 98% |
| Confident belief under a false label (validation pass) | 64% |
| The gate blocks purpose and passes method (V10) | 100% |
| Intent-gating beats surface filtering on disguised content (V10) | 83% |
| A stale detector fails asymmetrically (V10) | 75% |
| The wall is a distinct failure | 0% |
This does not throw any result out. Every one is true and reproducible. What changes is the sentence you may write after it.
A high rate means the result comes from the theory's structural commitment — reading intent under a metabolic budget — which is the theory, but is the part shared with any account of the same shape. It does not distinguish this framework from a competitor built the same way.
A low rate means the theory's specific commitments are doing the work. On the evidence so far that is one tested result: the wall. That is a narrower foundation than this document used to imply, and it is also a much clearer target.
That rate says how much of a result is architectural. It does not say which part of the architecture. So the complementary pass: keep the settings, strip the shape — remove one structural commitment at a time and see what dies. Each of the six is a decision about what a reader is; none is a parameter.
| finding | dies without | doesn't need |
|---|---|---|
| A false label moves you away from the truth | modelling a maker · a shared body plan | costly attention · provenance-as-state · hierarchy · holding a distribution |
| The wall is a distinct failure | modelling a maker · holding a distribution | costly attention · provenance-as-state · hierarchy · a shared body plan |
| Depth moves the method, not the purpose | modelling a maker · a shared body plan | costly attention · provenance-as-state · hierarchy · holding a distribution |
Every finding dies when the reader stops modelling a maker and starts classifying a surface. That is the one commitment the whole project rests on — the one E21 attacked and E45 defended.
And the wall is the odd one out twice over. It is the only finding that needs the reader to hold a distribution rather than a best guess, which is exactly right: the finding is a claim about the shape of a posterior, and a reader keeping only its best guess cannot notice it read every word and found nobody there. It is also the only finding with a 0% false-positive rate. Two passes from opposite directions agree on which result is genuinely about the theory.
Full grid, and the row that could not be read: RESULTS_V9.md.
Numbered in run order, named for what each was about. Specs were written before the code and never edited after; results were written after and never edited either. Full material in docs/versions/.
| name | built | asked | came back | nulls | |
|---|---|---|---|---|---|
| 1 | The Mechanism | cost, inference, zero provenance preference | can this be a model at all? | three results, two still standing | N1–N7 |
| 2 | The Learner | a reader that must acquire machine content's shape | what if the reader has to learn? | heterogeneity is in the likelihood, not the prior | N8–N12 |
| 3 | The Refuted Repair | a fix for the generational leak | is the leak sampling noise? | its own gate said no; experiment stayed withheld | N13–N15 |
| 4 | Foreign Intent | goal-foreign replaces goal-empty | what is machine content? | a headline inverted: readers keep paying | N16–N20 |
| 4.5 | Three Gates | a three-gate reader, a counting classifier | does invention need a theory of mind? | no — claim withdrawn | — |
| 5 | Depth Over Effort | depth as a hierarchy the reader infers | effort, or compressed practice? | depth; and the fast solver was caught misreading it | N21 |
| 6 | Code Against Equation | depletion, graded gate, κ→θ coupling, process recovery | is the code the same object as the theory? | three terms had no counterpart in the code | N22–N30 |
| 7 | The Closures | evidence-efficiency and zero-shot tests | what does a maker-model buy? | efficiency half withdrawn as an oracle; reads an unseen goal | N31–N34 |
| 8 | The Severity Pass | reader hierarchy, integration cost, decaying belief, a lying maker | how often does a random model do it too? | 100% / 98% / 0% | N35–N40 |
| 9 | Minimal Models | six structural ablations | which commitment is each finding made of? | all die without the maker-model | N41–N44 |
| 10 | Reader As Defence | an intent-gate on a learner's absorption | is reading intent a defence? | yes, where surface filtering does nothing | N45–N51 |
Three audit passes sit alongside, in docs/audits/: validation (five of nine checks came back against the work), diagnostics (four limits on what the instruments can answer), and repair (the uptake measure gained a sign, and the headline reversed with it).
The framework used to say that producing confident, mutually contradictory readings of empty content requires a reader that models the maker as a mind. That is false, and the experiment that killed it is in here.
A naive Bayes classifier, trained by counting on 200 examples, never representing a creator or a purpose or an intention, reproduces the effect: within-reader certainty 0.126 and between-reader disagreement 1.379, against the full model's 0.108 and 1.377. The mechanism turns out to be finite-sample overfitting, not theory of mind. (E21)
Two results survive that comparison, and no baseline reproduces either. One is the label-induced switch: the same content read as uncertain (1.335 nats of doubt) or certain (0.108) depending only on what the label says. The other is sustained futile attention on content whose structure is real but foreign, which only a reader that keeps expecting to learn something can produce.
Those are the two the framework actually uses. The withdrawn claim is marked at every point it was made.
Four different things get confused constantly. They are not the same and they make different
predictions. (In the code these are goal_empty, goal_foreign, the value gate, and the
non-invertible family; the theory's words are used here.)
- Intent-empty is wood grain. Structure with nothing behind it, because nothing was deciding. Versions 1 to 3 modelled machine-made content this way.
- Intent-foreign is a page in a script you cannot read. Real intent, pursued by a real process, expressed in a vocabulary your reading apparatus has no entry for. Version 4 replaced intent-empty with this, because it is a better description of a generative model — trained on purposeful human output and inheriting its shape.
- Intent-unrecoverable is every word familiar and nobody there. The vocabulary is yours; the route back from the surface to a state you could occupy does not exist, because many maker-states produce the same surface. Version 6 added this, and it is the one that matches what people actually report about generated text.
- Value divergence is a person who wants something you do not want. A fourth thing again, not what any of the above describes, and the model keeps it in its own parameter.
None of the four is "the intent is repugnant". That is a fifth thing and the model does not contain it.
The switch from intent-empty to intent-foreign inverted a headline result. Under intent-empty, readers disengage: nothing is identifiable, so paying attention stops being worth it. Under intent-foreign they do the opposite. They keep looking, keep paying, and never get anywhere. The failure to read intent survives and gets worse. The saving of effort does not.
Both cannot be right, and which one holds depends on a fact about real machine-made content that a simulation cannot settle: how much human-shaped structure it actually carries. That question is written up as E34, a prediction card a human study can use to locate real content on the axis rather than argue about it.
Two experiments were run against a version of the model that was later found to be wrong, and are uninterpretable rather than embarrassing. Reading them requires holding an assumption we now know to be false, which makes the numbers meaningless rather than inconvenient.
- A test of whether "how hard the maker was trying" gates uptake. The construct was wrong. What gates uptake is not effort but the depth of thinking behind the work, which is a different quantity: a fully committed shallow effort and a master's offhand sketch sit on opposite corners. Superseded by the depth test in the table.
- A test of whether three separate gates behave differently. One of the three gates was the mis-specified effort construct, so the design was testing something that does not exist. Re-run correctly, with two gates, as the same-mechanism test in the table.
Kept deliberately: every result that came back against the framework. The counting-classifier result that withdrew a claim, the knee that weakened a stronger framing, the prediction that missed once all four variants were run rather than only the flattering one. Those are unflattering and interpretable, which is the distinction being drawn.
Nine ideas were tested and killed, and the list is in the table above under the results that came back against the framework. Three of them were the author's own, one was the framework's claim about its own necessity, and one was the proposal this whole project is named after.
One died to a test the author approved knowing it had two possible outcomes: his claim survives, or his claim weakens. There was no version where it got stronger.
The last two are the ones to read if you only read two. E53 killed a reconciliation the author proposed and found the disagreement it was built on had been a comparison error. E54 killed the mechanism by which the Ghost Scale was supposed to work — the effect is real, and too small to carry a label.
"Held" means the prediction was written down before the run, in a spec or a hash-locked pre-registration, and the measured outcome met the criterion as stated. Not "broadly went the right way". Anything that needed the criterion softened, the framing widened or the outcome reinterpreted is not counted as held.
| outcome | count | which |
|---|---|---|
| held | 28 | E1–E7, E9, E10, E16–E20, E31–E33, E35–E38, E40–E43, E48–E51, E55, E56 |
| held in part | 10 | E21, E28, E29, E30, E39, E49, E52, E53, E54, E57 |
| did not hold | 5 | E11, E12, E14, E15, N21 |
| retired after a later pass | 1 | E20's crash/peak co-location |
| classification refused | 1 | E13 |
| withheld, never passed its own control | 2 | E8, H10.4 (values riding in on process) |
| not answerable in simulation | 1 | E34 |
Three things this count does not tell you, and they matter more than the count.
- Two of the three headlines tested for it are architectural — see How much of this is the theory?. A held prediction whose false-positive rate is 100% is still held. It is just not distinguishing evidence.
- Across versions 6, 7, 9 and 10 a design or criterion was changed after seeing a result in
eighteen places — seven found by the version 8 audit, four added openly by version 9, seven
more by version 10. Each change is documented where it happened; the version 6–7 ledger is in
results/v8/s1_severity.json, and versions 9 and 10 each carry a deviations table in their own results document. A high hit rate with a forking-paths count that size should be read with that in mind. - The project has no forward test. It had one sealed prediction, and its status was withdrawn in version 8 because the author does not recognise authoring it. The experiment was run anyway (E52) and its primary held; but a commitment nobody remembers making is not a forward test, and the record now says the count is zero.
- Readers in the model have zero preference over provenance. They cannot want work to be human. Any effect of a provenance signal has to come through inference, never through wishing. Asserted at every construction.
- Every headline effect has a matching null condition that has to come out null. There are
fifty-one of them, N1 through N51, across
tests/. Three of them fail and are reported as failing rather than quietly dropped: N21 (depth versus effort), N45 (the intent-gate's clean-corpus cost) and N50 (value drift with nothing to detect). - Acceptance criteria are pre-registered as executable code and content-hash locked before any run. The written criterion and the applied criterion are the same object, so they cannot drift apart.
- Every deviation is logged, including two where a criterion was restated after seeing a measurement. In both cases the original criterion is retained, still computed, and reported as failing. VALIDATION.md recomputes every one of them that the committed data supports and reports which verdicts would change.
- Some checks exist because the failure they guard against actually happened during a build and was caught by an assertion rather than by a result. Those are marked as such in the code.
- The validation pass could return the unwelcome answer, and did. Its criteria were fixed and hash-locked before it ran, its own restated criterion is logged in its own verdict file, and five of its nine checks came back against the work. Four came back clean: the robustness sweep, the cross-version consistency check, the seed-and-scale check, and the forward prediction, which is locked and not yet testable either way.
Stated generously, because the point of a limits section is to be the thing a sceptical reader reaches for and finds already there.
- A simulation of a mechanism, not a study of people. No human data. Nothing here is evidence about what real readers do.
- The specific numbers are properties of this model's dimensions and do not transfer. The shapes are the claims. The independent reimplementation makes this concrete: rebuilt from the prose alone, the label effect points the same way and comes out fifteen times smaller. Quote directions and orderings. Do not quote multiples.
- One headline is architecture-dependent. Inside this model's own two long-standing design decisions, 64% of randomly parameterised versions reproduce the label effect, and the effect this model produces sits near the middle of what random ones produce. The contribution there is the architecture and its constructions, not the settings, and it should be read that way.
- The central assumption is that Ghost Scale opacity can be read directly as the fraction of a maker's intent that survives into the work. That is a modelling choice, it is stated in the open below rather than buried, and it is the line to argue with.
- The readability axis is downstream of a design decision. The human and machine feature blocks were made disjoint because no such split existed at the original feature count, so the space was doubled. Every claim about that axis inherits it. A different split would keep the shape of the claim, an interior maximum, and would move the specific location. That is why the prediction card is written as a location to be measured rather than a number to be trusted.
- The labelling-coverage figure is a lower bound. The convention-aware reader is handed the true coverage, which is the most generous assumption available, so a real reader can only do worse.
- The trust-exploit multiple is an upper bound. The reader in this build cannot doubt the label it conditioned on, so a reader who could would be harder to fool.
- The Ghost Scale's own attention gradient is non-monotone, and in the wrong place. The tier meant to signal "do not spend effort here" is rendered as the visually loudest element on the page, while the Curator tier's reduced contrast makes it genuinely easy to skip, and E1 finds Curator the most expensive tier for readers. The model and the design disagree, and the design is probably right about human behaviour. The 5% tier is close to a logical impossibility besides: prompting alone constitutes more selection than 5% implies. This repository's own charting code quietly declined to use the published opacity ramp and substituted a monotone one, which is the argument having been made in code before it was made in prose.
- Version 5's construct corrections are theoretically motivated and largely untested. They are more defensible than what they replace, which is not the same as being right.
- A literature search has been run and it is retrospective. It happened after the simulations, it is reported in its own column, and it informed no design. It is a coherence check.
- The project has no forward test. It had one sealed prediction. Its status was withdrawn in version 8 because the author does not recognise authoring it, and a commitment nobody remembers making is not a forward test. The experiment was run anyway (E52) and its primary held. The hash-locked card is still in VALIDATION.md; what it is not is evidence.
git clone https://github.com/EmotiveAutomaton/ghost-scale-sim
cd ghost-scale-sim
# with uv (preferred) or pip
uv sync
# run everything, full scale, parallel
python run_all.py
# one experiment (writes results/eN_*.csv and figures/eN_*.png)
python -m ghostscale.experiments.e1_crash
# the validation pass (writes results/validation/, then VALIDATION.md)
python runners/run_validation.py
python scripts/write_validation_md.py
# the diagnostics pass on the instruments (writes results/diagnostics/, then DIAGNOSTICS.md)
python runners/run_diagnostics.py
python scripts/write_diagnostics_md.py
# the repair pass (writes results/repair/, then REPAIR.md)
python runners/run_repair.py
python scripts/write_repair_md.py
# version 6 (writes results/v6/, then RESULTS_V6.md)
python runners/run_v6.py
python scripts/write_results_v6.py
# version 7 (writes results/v7/, then RESULTS_V7.md)
python runners/run_v7.py
python scripts/write_results_v7.py
# version 8 (writes results/v8/, then RESULTS_V8.md) — the severity pass runs last
python runners/run_v8.py
python scripts/write_results_v8.py
# version 9 — the minimal-model programme, and the two literature experiments
python runners/run_v9.py
# version 10 — the reader as a defence; the severity pass runs last
python runners/run_v10.py
# redraw every chart from the committed CSVs, without re-running anything
python scripts/make_walkthrough_plates.py
python scripts/rebuild_figures.py
python scripts/make_ghost_scale_pair.py
# tests
pytest -qThe four published Ghost Scale tiers, CREATOR / POLISHED / CURATOR / GHOST, are drawn at 100%, 95%, 60% and 5% opacity. The model reads that opacity directly as the fraction of the maker's intent that survives into the work:
alpha = {CREATOR: 1.00, POLISHED: 0.95, CURATOR: 0.60, GHOST: 0.05}
A[0][:, tier, goal, DEEP] = alpha[tier] · sig[goal] + (1 - alpha[tier]) · noise_free_synth
This is the load-bearing assumption of the whole framework, so it is in the open rather than buried in an implementation file. If you think opacity and recoverable intent are not the same quantity, this is the line to argue with. The validation pass shows the label result survives compressing and stretching this ramp, so what matters is the ordering rather than the exact published values.
noise_free_synth is a structured, non-uniform, goal-independent distribution over features, with
entropy well below uniform, asserted in code. Machine-made artifacts in this model are not noise.
They are richly patterned. What they lack is any dependence of that pattern on a goal.
That distinction matters because the obvious strawman, "machine-made content is random noise", would produce the same crash for the wrong reason. Null N6 exists to separate them. Both are low information. Only the strawman is high entropy.
And validation found the sharper version of this. What the disagreement result actually depends on is not that the distribution is structured but that it is goal-symmetric, sitting equidistant from every goal the reader holds. Switch the symmetrisation off and the readers still become confident, but they all become confident about the same goal, which is shared error rather than invention. That was written down as a design decision in version 1 and the pass confirms it is load-bearing.
Version 5 models how much thinking sits behind a work as the number of levels of the maker's decision hierarchy that reach the surface. It is built so that a deep work and a shallow one have identical feature histograms, to machine precision: a reader that counts features and ignores their order cannot tell them apart at all. What distinguishes them is the order. A deep work moves through stages, and a deeper one moves through them in an order that names what it is for.
That constraint is asserted at every build, and it is what stops "depth" from being a second name for "legibility", which the model already measures elsewhere.
The learner in version 2 does not arrive holding a correct model of what machine-made content looks like. It has to acquire one from an unlabelled, already-contaminated corpus. What it starts with is a deliberate theoretical position: it knows the shared goal-to-feature family, and it does not know how provenance modulates it. Readers share a likelihood family because they share a body plan. What they do not share, and have to learn, is which sources carry intent.
It is also forced. A genuinely uninformative prior is not a slow learner, it is an unidentifiable one: information between features and goal sits at exactly 0.0000 nats after 400 artifacts and all four learned columns stay bit-identical. That measurement is kept as a live test, so if it ever stops being true the decision gets revisited instead of inherited.
The observer's beliefs can be updated two ways. inference.exact: false uses pymdp's variational
solver, which keeps beliefs about different unknowns separately and updates each using an average
over the others. Every committed number in results/ was produced that way. inference.exact: true
uses ghostscale/exact.py, which carries the belief over every combination
of unknowns at once with no independence assumption anywhere.
The shortcut is fast and it is known to have been badly wrong once. Version 5 caught it returning the shallow answer for every artifact, confidently, while exact arithmetic on the same observations recovered depth correctly. Switching solvers is one config flag and every experiment runs unchanged under both, which is what made the validation pass a substitution rather than a rewrite.
metrics.psi_analogue is a discrete stand-in for the closed-form Ψ of the preprint:
psi = [engaged] · (-ln(1 - κ)) · KL( Q(goal | τ) ‖ P0(goal) )
The sigmoid gate of the closed form is replaced by the binary engagement decision. This reimplements the intuition, which is engagement-gated, trust-weighted surprise about the goal. It is not a port of the equation and no equivalence is claimed.
Three additions, each of which makes something representable that was not.
A hierarchy the reader owns. Until version 8 the maker had levels and the reader did not. A reader can now only recognise as far as it has itself built — which makes "you cannot see what you have not done" a structural property rather than an assumption, and it is what the human acquisition test would target.
A cost for being changed. Integration is no longer free. Absorbing something is an expenditure, separate from the expenditure of looking at it, which is what lets attention and uptake come apart cleanly rather than by stipulation.
A memory that fades to a residue, not to zero. Belief decays toward baseline + floor × residual. The first implementation decayed geometrically to baseline, which is erasure with extra
steps; associations weaken and do not quite disappear, and the model now says so.
The published equation is
Ψ = sigmoid(k·(ω − θ_E,C(κ))) · [−ln(1−κ)] · D_KL(Q(R|τ) ‖ P₀(R))
with θ_E,C(κ) = θ_base(E) + λ·D_KL(Q ‖ P_c)
Version 6 read it against the shipped code and found three things absent: θ_base(E), the metabolic reserve; the gate gain k, which makes the gate graded rather than binary; and the κ→θ coupling, which is not an omission but a disagreement about mechanism. The code's exploit works by the label out-arguing the work. The paper's works by trust lowering the guard — and only the paper's version predicts a reader told the truth, believing it, and absorbing the work anyway.
ω is two different objects in the two documents. In the preprint it is the reader's precision weighting, which is an output. In the code it is feature overlap between hypothesis families, which is an input. Anywhere the two are compared, that has to be said first.
The mechanism is one line and no new parameter: the learner's Dirichlet learning rate is set from the gate, and the commit happens after inference resolves. So the reader works out who made the thing first, and only then decides how much of it comes in — which is the ratchet the version is built on. Four gates were compared: no filter, a surface-quality filter, a provenance-label filter, and the intent-gate in three variants (reconstructibility alone, hand-set values, values learned from what has already been absorbed).
The reconstructibility-only variant is the proposal, and it is deliberately value-free: it rejects what it cannot attribute a maker and a purpose to, and holds no view about what anyone should want. It never reads the provenance signal, which is asserted by test rather than claimed.
This model fills a named gap in the active-inference literature on Theory of Mind. The honest mapping:
| this model | corresponding construct | source |
|---|---|---|
creator_goal as a hidden state the reader infers |
opponent preference parameters θ_j inferred online | Albarracin et al. 2026, arXiv 2602.20936 |
| the reader's model of the maker | structurally matched other-model M_other | Albarracin et al. 2026 |
| κ as precision on the provenance channel | reliability-gating r between learned and static ToM prediction | Albarracin et al. 2026, Eq. 15 |
| engagement policy over attention | epistemic-value term in social EFE | Pitliya et al. 2025, arXiv 2508.00401 |
| Ghost Scale tiers (alpha as opacity) | no existing analogue | |
| a generator with no preferences at all | no existing analogue |
Two things this model adds, stated narrowly because the value of the contribution is that it fills a named gap, and that argument gets weaker if it claims more ground than it holds.
- What sets the weight. Albarracin et al. fix the empathy parameter λ from outside the model and name "what sets λ" as their central open question. κ, gated on provenance and on value divergence, is a candidate answer: the weight you put on another agent's inferred preferences should scale with the evidence that those preferences exist at all.
- What happens when the other has no preferences. Existing work assumes the observed other is an agent with preferences. This model asks what happens when it is not, and finds the inference does not fail gracefully. Under high trust it fabricates. (E2, E4)
The training-side result lands next to machine-generated-text detection and to pretraining data curation, and it does not overlap either.
| that field does | this does instead |
|---|---|
| surface-statistical detection — perplexity, burstiness, learned classifiers | reconstructs the maker, not the generator |
| perturbation methods (DetectGPT, Binoculars) — token-prediction consistency under masking | intent-attribution consistency, which is a different object and will be the first thing a reviewer conflates |
| supervised classifiers on labelled human/AI corpora | no origin label anywhere in the gate; asserted by test |
| quality / reward-model filtering of pretraining data | scores the inferred author, not the artifact |
| watermarking and source attribution | requires cooperation from the generator; this requires none |
Two points of contact worth naming because they support the work rather than compete with it. The detection field's own reviews now report that surface statistics are increasingly insufficient as models grow more fluent, and that supervised detectors fail to generalise across domains and against adversarial paraphrasing — which is this project's own arms-race result, arrived at independently and published by other people. That is a coherence check, not a citation of convenience.
Work on inferring latent intent with language models does exist and is active, but it is about user intent in interactive settings — dialogue, agents, recommendation. Artifact-level author intent, used as a corpus instrument, appears unoccupied. That claim rests on a few search rounds rather than a literature review, and a proper one is the first task of any follow-on work, because finding out cheaply that someone has already done it is a good outcome.
Every version has a name as well as a number, and the name is the theme rather than the chronology. Directories sort in run order; the names say what each one was for.
README.md this page
FINDINGS.md every question and its CURRENT answer -- the one page to read
WALKTHROUGH.md 25 plates, in the order they are being published
EVIDENCE.md what the world has published, next to what this predicted
docs/HISTORY.md ten versions and three audit passes, as one narrative
docs/EXPERIMENTS.md the original plain-language table, superseded by FINDINGS.md
docs/theory/ the preprint this implements, and the code-to-theory vocabulary
docs/assets/SPEC.md the spec for the public-facing material
docs/versions/ one directory per version: SPEC.md, RESULTS.md, and where they
exist PLAN.md and DECISIONS.md. Specs were written BEFORE the
code and are never edited after; results are written after and
are never edited either.
v01-the-mechanism/ the model as code: cost, inference, zero provenance preference
v02-the-learner/ a reader that must acquire what machine content looks like
v03-the-refuted-repair/ a fix whose own gate refuted the diagnosis behind it
v04-foreign-intent/ goal-empty becomes goal-foreign, and a headline inverts
v04.5-three-gates/ three gates, and the counting classifier that withdrew a claim
v05-depth-over-effort/ depth replaces effort; the fast solver is caught misreading it
v06-code-against-equation/ is the simulation the same object as the published theory?
v07-the-closures/ the four results V6 would not draw, and what a maker-model buys
v08-the-severity-pass/ the reader gets a mind; how often a random model does it too
v09-minimal-models/ strip the shape: which commitment is each finding made of?
v10-reader-as-defence/ can reading intent defend a learner? The last simulation version
docs/audits/ passes that ask whether existing answers can be trusted. None
asks a new question about the world. Their RESULTS.md files are
GENERATED from verdict files and never hand-written.
a1-validation/ nine checks; five came back against the work
a2-diagnostics/ what the instruments can and cannot answer. Read before quoting
a3-repair/ what could be fixed, and what fixing changed
ghostscale/ generative_model, creators, environment, observer, learning, metrics
exact.py exact joint inference -- the solver validation substitutes in
fitting.py parameter estimation by exact likelihood, for the three
parameters that are not hidden states and have no posterior
v4_model.py hypothesis-space overlap; goal-foreign content
v4_5_model.py the three-gate observer
v5_model.py depth as a hierarchy the reader infers
v6_model.py the Intent Extraction Limit: depletion, the graded gate, the
trust-to-threshold coupling, process recovery
v8_model.py reader hierarchy, integration cost, decaying belief, density
latent_goal.py the goal a maker does not know it has
plates.py the figure house style, and the automatic plate audit
experiments/ e1 through e34, each runnable standalone
v6/ v7/ v8/ v9/ v10/ later versions, each with its own hash-locked criteria
validation/ diagnostics/ the audit passes, each with a separate lock
repair/
prereg_*.py acceptance criteria as executable, hash-locked code
runners/ one entry point per version and per audit pass
run_all.py the original experiment programme, E1 to E34
tests/ model invariants and the null suite -- N1 through N51
config/default.yaml every parameter, for every version, plus the solver switch
results/ committed summary CSVs and JSON verdicts. V1-V5 wrote flat into
results/; V6 onward use one subdirectory per version. See
results/README.md
figures/walkthrough/ the 25 plates WALKTHROUGH.md is built from
ghostscale/validation/soundingline/ THE ONE LIVING PART OF THIS REPO -- see below
notebooks/walkthrough.ipynb runs E1 and E2 end to end, narrated
scripts/ chart rebuilders, the document generators, and the independent
reimplementation of the two-gates result
Everything in ghostscale/v1 through v10 is closed. Pre-registered, run, reported, and left
alone. A number in this README should be the same number next year.
ghostscale/validation/soundingline/ is the exception and is expected to churn. Another project —
Sounding Line, which reads real text and therefore has no ground truth — sends this simulation
questions about mechanism, and that directory answers them. Its questions arrive from outside,
new modules appear whenever the next one needs answering, and nothing in it is a Ghost Scale
hypothesis: S-1 asks whether a statistic in somebody else's pipeline is broken, which is a fair use
of a simulator and is not a finding about readers.
Two rules hold there, both learned the hard way here. Nothing may call a versioned run() — the
V10 severity pass re-ran real experiments to audit them and overwrote the very verdicts it was
checking. And anything claiming to use an experiment's rollouts must reproduce that experiment's
committed number first, because a harness that re-randomises is running a different experiment
rather than a control.
Read anything under it as work in progress. Read everything else as settled.
Raw per-reader CSVs are not committed, because e4_raw.csv alone is 16 MB. Everything a number in
this README or a chart in figures/ depends on is committed. Regenerate the raw files with
python run_all.py.
Several defaults were recalibrated on contact with the implementation. Each is documented under "Deviations" in the matching write-up, with the evidence that motivated it, so any of them can be argued with, and the validation pass recomputes the originals. The load-bearing constraints have not changed: zero reader preference over provenance, structured rather than uniform machine-made content, reader heterogeneity, the full null suite, and the honest crosswalk above.
- Preprint: Art as an Algorithmic Virus,
10.5281/zenodo.19407789 - Plain-language essay: https://abrahamhaskins.org/art
- Ghost Scale Figma kit: https://www.figma.com/community/file/1624141586132218953
Code is MIT. Prose, figures and data are CC BY 4.0. See LICENSE.
To cite this repository, see CITATION.cff, or cite the preprint directly.