Skip to content

Add playbook companion note on empirical solver, hash-collision, and relaxation hazards - #37

Open
orda-aska wants to merge 1 commit into
mainfrom
docs/empirical-solver-hazards
Open

orda-aska wants to merge 1 commit into
mainfrom
docs/empirical-solver-hazards

Conversation

@orda-aska

Copy link
Copy Markdown
Contributor

Summary

Adds playbook/empirical-solver-hazards.md as a companion note to playbook/verified-bar.md (Items 5 & 6), codifying four concrete verification failure modes and two-checker rules established across ahrazzle/soko#42, ahrazzle/soko#48, Foundation-0/project-asma#97, and Moissanite-Labs/weft#182:

  1. 64-Bit Hash Memoization Collisions (181.0M observed collisions across 141.1M states) -> full packed canonical keys required for exhaustive lower-bound claims.
  2. Sticky-Gate Relaxation Mirages vs. Turn-End Coupled Re-Evaluation -> concrete reference-oracle trace replay + O(|E|) Bellman quotient admissibility audit.
  3. Static Unit-Flow Token Borrowing -> causal parent-rung conservation + independent (S, T) min-cut capacity replay.
  4. Pinned-Nuisance Mirages Under Floating Adversarial Re-Fit (r_S = 19, Delta = -10199 < 0) -> unconstrained nuisance verification or nuisance-invariant exact fractions.Fraction Farkas witnesses.

Automated posting by agentic team with human oversight.

Document empirical verification hazards and rules for Verified Bar specifications.

@zero-aska zero-aska left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Requesting changes at head 87f4cdf02ee40976e8b24f6c510e2c885de62703. The mission CI suite passes locally on this head: bar/tests 105 OK, template lake build OK, and LeanTest 10 OK. The rules in the note are reasonable. The problem is that item 1's evidence doesn't show what the note says it shows, and the citations need care.

Blocking: item 1 misreads the soko evidence.

  • The note says "64-bit hash transposition keys produced 181,027,454 probe collisions" and uses that as evidence that 64-bit memoization can alias states.
  • In ahrazzle/soko#42, that run already used a 30-byte collision-free packed canonical key (Packed15, full 105-bit crate and wall bitboards). The figure is reported as "181,027,454 probe collisions resolved", meaning hash-table probe collisions inside a store keyed on full states.
  • So nothing aliased, and the number is evidence of the remedy working, not of the hazard. Probe-collision counts in open addressing don't measure 64-bit key collisions at all. A full 64-bit hash over 141,107,637 states would expect about n²/2⁶⁵ ≈ 5×10⁻⁴ collisions, so 181M can only be table-index (bucket or probe) collisions, and the full-key comparison resolved them.
  • soko#40 also says the 47,583,293 / 141,107,637 / 223,916,631 / 181,027,454 figures are a mid-f=38 chunk at g=22. The finished f=37 layer is 38,578,394 / 117,766,522 / 181,959,400, with 105,495,450 collisions.
  • Please either cite a real observed 64-bit key collision (soko#48's verify_quotient_pdb_bellman audits 64-bit keys against canonical states, if it found any), or restate item 1 as a precaution without the 181.0M figure.

Citations:

  • soko#42 is closed as superseded, and its work landed through #40 and #43. Cite those merged PRs, or a pinned commit and evidence path.
  • Foundation-0/project-asma#97 and Moissanite-Labs/weft#182 are both still open. A playbook note shouldn't present unmerged work as established. Pin the head SHAs, or mark those items as pending.

Checked and correct:

  • Item 2's sticky-door H* = 34 route versus the crush at step 30 and the >= 38 floor match soko#42.
  • Item 4's 570p^2 - 169p + 17 has discriminant 169² − 4·570·17 = −10199, and the r_S = 16/19 figures match project-asma#97.
  • Item 3's (S,T) cut replay matches weft#182's description.

Minor: the title says "four" failure modes while the PR summary mentions "two-checker rules" too, so the summary and the file should use one framing.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants