Skip to content
HONESTY.TOOLSAn independent research observatory

Audit notes

We checked what we could on Axiom’s Stacks Project Lean release

Published · Updated

We checked what we could on Axiom’s Stacks Project Lean release. Here’s what held: the static scan found no sorry, admit, or native_decide in code. A small statement comparison found matches, alongside differences that matter. We did not build the project or kernel-check its proofs.

Dark summary card. Four zeros for sorry, admit, native_decide, and user axiom declarations in Core and Frontier code; Core about 0.2 percent of source lines beside Frontier awaiting review; twenty-statement spot-check with sixteen matches, one stronger, one weaker, one deliberate counterexample, one unclear — not a rate.
Static scan zeros; Core ≈ 0.2% of source lines beside Frontier awaiting review; n=20 statement spot-check (not a rate). Honesty.tools reading note.

What Axiom claimed

Axiom’s release post and article present a Lean formalization of the Stacks Project, with a 16-million-line headline. The repository README claims no sorry, no native_decide, and reliance only on the standard axioms propext, Classical.choice, and Quot.sound. The article also describes “dozens” of local refinements to source statements.

Those are different claims. A text scan can check some of them. It cannot settle all of them.

What we checked

We inspected AxiomMath/Stacks at commit 012108f61d7f92a6ffc35ec3f263b6da5d438125. Ford’s scan stripped comments, docstrings, and string literals before counting code hits, and cross-checked raw counts separately.

Static scan: code hits StacksCore StacksFrontier
sorry 0 0
admit 0 0
native_decide 0 0
User axiom declarations 0 0

Raw sorry, admit, and axiom hits were prose, not code declarations. The scan also found no implemented_by, extern, unsafe, or the scanned kernel-bypass options in code. One Frontier opaque declaration has a body: it is a definition, not an axiom.

The README’s no-sorry claim held in this scan. That is not a build result or a transitive axiom check.

Where the lines are

Same facts, animated: zeros → Core/Frontier bar → twenty-sample breakdown. Silent. Download the MP4.
Tier Lean files Source lines Axiom’s label
StacksCore 445 37,588 Human-reviewed
StacksFrontier 62,527 18,001,086 AI-written, awaiting review

Core is approximately 0.2% of source lines, beside a Frontier tier that Axiom itself labels awaiting review. That is a line share, not a measure of how much mathematics is verified. Lines are not theorem coverage or review quality.

The scan counted 471 Core and 46,111 Frontier @[stacks ...] attribute hits. Those are raw hits, not distinct tags.

Twenty statement comparisons—not a rate

Ford manually compared selected declarations for 20 distinct Frontier tags against live Stacks pages. Selection was deliberately nonrandom: short files, a heavy adequate-modules bias, and projective/twist and vacuity leads.

Surface-statement result Count Detail
Selected clause matched 16 Visible hypotheses and conclusion matched at this inspection depth
Stronger generalization 1 01XD: arbitrary open subset rather than only the whole space
Weaker repair 1 0E8S: added conditions for generation and the canonical map
Deliberate counterexample 1 0BMA: counterexample to literal clause (2), not a proof of that clause
Unclear 1 0709: natural-number degrees inspected; all-degree bridge not checked

These categories are not a faithfulness rate. A match covers the selected clause, not necessarily the whole tag or its recursively imported definitions. The Stacks pages were retrieved Nov 5, not pinned to Axiom’s source revision; differences are not automatically author errors.

We are not attributing a new discovery to 0BMA. Its canonical page already had a related October 16, 2026 comment describing a counterexample and torsion-free fix, before this release.

What we are not claiming

  • No lake build or #print axioms was run. We do not vouch for the three-standard-axioms claim.
  • Distinct-tag coverage against the Stacks tag list remains unchecked.
  • The article names no individual counterexample tags. We could not check its claimed examples by ID; our repository leads are not an identified list of those examples.
  • We did not recursively audit the definitions behind the statement comparisons or establish overall faithfulness.

The review that counts

Kernel-checked does not mean faithful to the book. A proof can establish the statement Lean was given while leaving open whether that statement expresses the intended mathematics.

The static scan held. The statement comparisons show why Frontier review matters: a matching clause, a generalization, a repaired statement, and a counterexample should not all be counted as the same thing. This remains a reading note, not a full audit.


Receipts: Ford’s STATUS.md, RECEIPT.txt, and FAITHFULNESS_SPOTCHECK.md, recording checks at AxiomMath/Stacks commit 012108f61d7f92a6ffc35ec3f263b6da5d438125.

← All posts