MathCase work map · updated November 13, 2026

What has been checked, what has been measured, and what remains open.

This page is the human-readable map of the work. It keeps Lean proofs, exact finite computations, internal run summaries, and unfinished experiments separate so each result is read at the right scale.

Public source Internal summary Staged or not run
Saved public case

A failed proof attempt is part of the result.

Plain statement: if a natural prime p divides n squared, then p divides n.

p prime, p ∣ n² ⟹ p ∣ n

The first saved model candidate was rejected by Lean because it used the divisibility assumption where the theorem required the primality assumption. The corrected candidate uses both assumptions and its saved Lean and axiom checks passed.

Known theoremChecked for this formal statement · no novelty claim · the webpage does not run Lean

The useful engineering result is the inspectable sequence: statement, rejected candidate, checker response, repaired proof, and saved check. A pass supports only the exact formal statement.

Where the files are: the saved case reader is reader.html; its display code is app.js; the prime case's evidence folder is data/local/verification/proofdesk-ui-dispatch-r1/ in the MathCase repository.

Walk through the saved case
Public Lean source artifact

Collatz: bounded descent across 7,825 residue classes.

Plain statement: for every qualifying n ≥ 2 in 7,825 of the 8,192 residue classes, one of the first 21 ordinary Collatz steps reaches a smaller value.

7,825 of 8,192 classes · first descent within 1–21 steps

This is not a percentage of the Collatz conjecture solved. It does not prove that every trajectory reaches 1, and it does not close any previously uncovered whole class modulo 128. The source is a saved local compile, not a fresh public-checkout replay or a standard M4 receipt.

Saved receipt location: proofs/collatz/remaining-residues/bounded-descent-receipt.json in the MathCase repository. The receipt is named for reproducibility; this page does not send readers to a raw data view.

Internal pinned replay summary · Nov 13, 2026

229 individual Collatz shard modules replayed.

A sequential offline replay compiled the existing M15, M16, M18, and M20 shard modules in a fresh checkout. All 229 modules exited successfully. A separate audit issued 4,262 theorem-axiom queries: each report listed only propext and Quot.sound, with no sorryAx.

229 / 229individual modules compiled · one CPU · 2 GiB memory cap · network disabled
TierChecked first-descent residue classes
M1531,473 / 32,768
M1663,422 / 65,536
M18254,649 / 262,144
M20 shard total1,021,248 / 1,048,576

These counts summarize finite shard statements. They are not percentages of the Collatz conjecture solved. Individual shard passes do not, by themselves, establish a combined theorem.

This remains an internal replay summary. The pinned Docker image is not anonymously pullable, so a stranger cannot reproduce this replay from a clean public checkout yet.

Internal composition result · public evidence package pending

M20: the shard results now compose into one checked candidate.

Plain statement: the exact 2,652-family conjunction assembled from the M20 shards was compiled through explicit intermediate modules on the pinned Lean and Mathlib snapshot.

133 intermediate modules + final 2,652-family theorem · exit 0 · 2 GiB cap · [propext, Quot.sound]

The bounded local run completed without an out-of-memory termination. It used the already replayed shard outputs and stayed within the 2 GiB container limit. The final theorem covers the listed affine first-descent families; it does not prove global Collatz convergence, fill every residue class, or establish novelty.

Public status: the composition source and run receipt are recorded in a private MathCase-internal commit, c82c9685fe8294b47db8ba90c9a65d6e189553de. Visitors cannot retrieve that private receipt from this site. The public evidence package is still pending, so this is reported as an internal result rather than a public reproducibility claim. No M4 promotion is implied.

Earlier attempts remain visible

The record preserves three earlier failures: two elaboration failures and a deterministic isDefEq heartbeat timeout at 200,000. That timeout was not an out-of-memory failure. The later chunked composition is an additional result; it does not overwrite or relabel those receipts.

Internal finite-sample summary

Collatz: a 4,096-case structure sample.

The frozen sample used 4,096 unique odd residue classes left after an earlier finite screen. In the discovery half, 2,002 of 2,048 representatives reached a smaller value within 64 odd-to-odd steps. In the holdout half, 2,007 of 2,048 did so. One of three registered hypotheses passed; the sample did not yield a general rule.

97.754% / 97.998%discovery / holdout bounded-descent rates on these sampled representatives only

These finite rates are not a convergence theorem or a population-wide probability claim. The initial-valuation hypothesis failed because the earlier screen forced that statistic; the holdout confirms only the registered finite rate.

Internal bounded CPU campaign summary

Erdős–Straus: 831,860 exact finite witnesses.

The campaign checked a factor-conditioned subset of the n ≡ 1 (mod 12) residual cases through d = 10,000,000. It recorded 831,860 candidate rows and 831,860 exact Egyptian-fraction witnesses. All 427 base-search misses in the audited range resolved inside the fixed 65,536-offset extension cap.

831,860 / 831,860candidate rows with exact finite witnesses · bounded to the recorded filter and denominator range

The full Erdős–Straus conjecture remains open. This campaign established no new parameterized identity, makes no novelty claim, and says nothing about filtered-out cases or denominators beyond the cap.

What the offset and miss analysis found

The largest first-denominator offset in the audited runs was 35,498, leaving 30,038 below the fixed cap. Per-range maxima were noisy; their correlation with denominator range was weak in v9 alone (r = 0.067), so the sample gives no bound on future ranges. The 427 base-search misses shared n ≡ 1 (mod 24), which is imposed by the candidate filter; their factor patterns still varied. The analysis found no single new rule covering the misses.

The detailed manifests, witness streams, extension replays, and audit remain in the internal research checkout. This page summarizes the result but is not a public reproduction bundle.

System engineering snapshot

The research system around the proofs.

Knowledge graph

A local searchable index contains 4,294 result and attempt records, including checked statements, dependencies, axiom reports, source and receipt hashes, and preserved failures. The graph is owner-only and is not a public dataset.

Human-reviewed statement batch

Fourteen elementary statements have recorded M0 statement reviews. Nine have individual Lean passes; later batch entries and aggregate reviews are tracked separately. These are standard facts, not open-problem discoveries, and none is promoted to M4 by this page.

Transfer comparison

A library-versus-baseline protocol and held-out targets are staged. The comparison has not run, so MathCase does not yet claim its knowledge graph improves later proof work.

Research controller

The local evidence layer supports stable case identifiers, separate statuses, preserved failures, and pause/resume. Model-driven proposing and adversarial checking are not running in this local-only work period.

The index and underlying replay receipts are not bundled on this public site. The MathCase repository link in the header points to the public source tree; internal summaries are labeled here so readers can distinguish them.

Next work

What would make the next claims stronger?

Some proof dossiers and source materials remain owner-only and are not linked here. Only results with a suitable public evidence package can move into the public source ledger.

← HomeBack to top ↑