Strengthen empirical controls and offline reproduction - #1
Closed
bc1cindy wants to merge 224 commits into
Closed
Conversation
The clusterer declines a doubted merge that would fuse two established clusters, cuts views along cluster-collapse regions, and orders a clustering by how little each merge contradicts itself. The shared detectors go with it: equal- amount coinjoins are recognised by their denomination, and the UIH axis is labelled by the heuristic it actually computes.
…tion Streaming transactions with a dense-int union-find lets a slice larger than memory cluster; the chunked downloader collects a month at a time without billing-side limits. Epoch splitting and backend profiling come with them.
…tion Seeds bootstrap from vertices whose local degree signature is unique and identical across both views, so propagation no longer needs an externally supplied seed, and the link-prediction fallback abstains by default to keep precision high. Measured on a real 2016 slice, where the attack's precondition fails and is reported as failing.
Run against real transactions for the first time, the channel was cutting every coin it could not measure: the oracle spells unreachable as negative infinity while the guard tested for None, which is 98.2% of input coins. The whole-transaction count is bounded at the width where it stops returning, and the link matrix supplies a fee-tolerant per-coin reading the exact-hit counts cannot.
Subset-sum multiplicity does not enter a refuse-only bound. In the sub-transaction literature the mapping count is the denominator of a link probability, never a multiplier on one, and the column being weighted is already renormalized, so the factor reintroduced exactly what the matrix had divided out. The paper moves with the contract: each claim the tree no longer supports is corrected at its source rather than only where it was retracted, and the guard matches the claim rather than three phrasings of it.
…a direction A manifest records a result's source identity and the population invariants a byte digest cannot see. separable is three-valued, so a decisive inversion is not reported as noise, and it gates the exact binomial on min(a, b) rather than a + b: the exact sum costs min(a, b) terms, so the lopsided shape a real direction claim takes stays exact at any sample size, and only near-ties reach the approximation.
…olicy The persistence curve and the slice gate reproduce their published tables from committed query output, so both move to state 1. Change-id validation moves the other way: the window it names holds 739,889 transactions and cannot be a fixture. Twenty-five documents that carried no evidence level now name one; the two that fit no state say which and why, and one of them is why the policy needs a state for a measurement whose data is gone. slice_gate.sql did not run: all three of its queries closed a CTE they never opened.
The rule is the one already applied here: bold marks a claim, not a term. Applied to running prose it leaves twenty independent clauses bold and strips a hundred and twenty-one names, values and single emphasis words. Run-in headings, table cells and quoted emphasis are untouched — the emphasis inside a quotation belongs to whoever wrote it.
Every evidence channel had a canonical run and the engine that fuses them had none: it ran on stubs in unit tests and on data this repository does not ship, so the object the argument rests on was the one nothing published executed. The reason was structural — the engine read transactions through a cache that is not committed — so fetch becomes an argument, the way the subset-sum hook already was. The ladder is the result. Topology corroborates merges the fingerprint refused and provenance cuts pairs it kept, which is what fusing them is for; the amount channel moves nothing here because it is refuse-only and gated behind fingerprint disagreement.
… carry The weekly export is 977 MB in five files, each past the 100 MB a git host accepts, so the results measured on it could not be reproduced by anyone. What the pipeline reads is narrower than what the export stores: the transaction id is never read, and an address only has to be distinguishable, not itself. Sixteen times smaller, 5.9 million transactions, every file under 14 MB. That an address may become an integer is a claim about the engine, so it is tested against the engine: the same partition under an order-preserving and an order-reversing renaming. Addresses a detector names itself in keep their text — 94 of 1,788,311 on the first epoch.
Numbering each file from zero passed every check the encoder had — counts, heights, coincidence graph — because each of them read one file. The damage was between files: `a000000000` stood for a different address in each window, so the clusterer read a coincidence that does not exist and the rejoinable population fell from 17,431 to 4,972 on the committed epochs. What caught it was not the suite, which stayed green, but running the same pipeline over the export the derived files came from. With one namespace the two agree exactly, and two tests now cover the seam the round-trip could not see.
The population was first measured on a 1.1 GB slice that no longer exists, so the figure could neither be checked nor taken again. On the committed weekly graph it lands in the same place: 17,431 clusters straddling the boundary at a mean internal degree of 2.866, and one correct rejoin out of the 15,688 pairs left once a tenth are seeded, with every shuffled-seed arm at zero. The null is asserted as a bound rather than left as an absence, so a run that does ignite fails the test instead of passing quietly. assortativity sorted vertices that a contracted view makes incomparable, which is why no contracted view had ever been measured for shape.
Half of them were reachable from no claim that named a source, and five of those cited nothing in the file either — no identifier, no year. A reader could not tell which paper, or which edition of it, the code was meant to agree with. The other half were bound through the claim catalogue, which does not help anyone who opens the module. A gate now requires the citation and that the entry exists, which is the weakest rule that fixes the problem: whether a reimplementation is faithful is a separate question this cannot answer.
Each was measured on a source that no longer exists, so nobody can check the numbers and nobody can take them again. Keeping them behind a label would give a figure nobody can verify the standing this work argues a vendor should not get: the burden of proof does not reverse because the measurement was ours. The paper's witness-drift paragraph went with them, and five declared-unbacked numbers with it. Git keeps the record; the published surface no longer offers it as evidence. REPRODUCIBILITY.md names the condition with the vocabulary the corpus already uses rather than inventing a sixth state.
Removing the results page took the script with it, and a test pins that script: the mechanism is unit-tested even though the sample its headline numbers came from is gone. Two checks disagreed about whether anything imported it and the narrower one was believed.
…y over Both of the framework's positive properties are the size of a minimum cut — how many coins a separation has to take before an output stops reaching its origins — and nothing here computed one. path_count accumulates route mass and says in its own docstring that this is a different number; on the committed cache the two disagree for three coins in every five. A tenth of the coins are separated from their whole ancestry by removing one coin. The boundary the walk stopped at is recorded beside that, because almost every origin it reaches is the edge of the slice rather than a coinbase, and a reader given the fragility without the sample limit would take a floor for a ceiling. Routes are counted, not weighted: the value half stays unimplemented and the capability contract now says so rather than denying the whole metric.
The specification prunes before it counts — a path worth less than the amount in question may have been removed — and leaving that out made the measurement overstate redundancy in the direction that matters. Routes that could carry the coin take the single-coin cut from 10% to 21%, and 38% of coins have no such route at all. This is the same error as counting origins without their overlap, made twice: a route that shares a coin does not add redundancy, and neither does one too small to explain the coin. Value enters as a threshold and not as capacity, so the capability contract still denies plausible-flow capacity; giving each coin its value and measuring the flow is a different quantity nothing here computes.
Nineteen of nineteen carried partially_supported, so the field said nothing and an audit could read past it. The schema already distinguishes that from supported_at_evidence_level, and the distinction is the one that matters: an existential claim demonstrated on a deterministic fixture is not partly proven, it is proven at the level it declares, and its limitations say what it does not reach. Twelve are that. Seven are genuinely partial and now each names what is missing, which is what the new gate enforces — along with the rule that a supported claim cannot also declare an unfinished mechanism. Writing them exposed that the reasons had lived in a script rather than the catalogue.
Owner
Author
|
Closing this preliminary publication while the local commit stack is reorganized and audited before the complete push. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Verification
Scope and limitations