burn-the-house-down
Skill4 starsChanged 21 days agoMITView on GitHub
Escape a local optimum in a long-horizon optimization campaign by running a structurally free descent beside the protected incumbent. Use when improvements shrink while machinery grows, candidates are rejected by incumbent-fitted gates, the same wall recurs, work has collapsed into knob-fencing, or the user says "local optimum", "stuck polishing", "optimization plateau", or "burn the house down".
In this repository79 other files it gives its agents.
- AGENTS.mdAGENTS.md
- accepted-proof-exact-retuneSkill
- arxiv-portable-researchSkill
- auto-autoresearch-pipelineSkill
- beachhead-ratchetSkill
- bounded-commuting-foldSkill
- calibrated-static-cost-proxySkill
- circuit-edit-integration-safetySkill
- colab-laneSkill
- computer-algebra-oraclesSkill
- constraint-construction-searchSkill
- cross-model-review-gateSkill
- deepseek-pressure-testSkill
- ecdsafail-fastevalSkill
- ecdsafail-lowq-playbookSkill
- ecdsafail-oracle-loopSkill
- ecdsafail-pareto-ratchetSkill
- ecdsafail-scratch-dietSkill
- elicit-researchSkill
- evaluator-scoring-surface-auditSkill
- exact-evaluator-replicaSkill
- exact-hold-passSkill
- exact-postpass-stackSkill
- exact-search-engineeringSkill
- fast-lean-proof-loopSkill
- fit-probe-and-spliceSkill
- gpu-affine-table-synthesisSkill
- greedyfrog-methodSkill
- hard-problem-contractSkill
- hold-demand-ledgerSkill
- hosted-carry-lendingSkill
- inert-op-screen-to-structural-ruleSkill
- invariant-chartSkill
- jev-second-perspectiveSkill
- lambda-exchange-deskSkill
- leader-rebase-stackSkill
- lean-arklib-proofsSkill
- lean-finite-certificate-splittingSkill
- lever-switches-and-break-controlsSkill
- mission-commandSkill
- nasqret-playbookSkill
- paged-walk-descentSkill
- peak-lifetime-ledgerSkill
- pinned-table-stackingSkill
- pip-disciplineSkill
- portable-source-assertionsSkill
- proximity-prize-campaignSkill
- proximity-prize-frontier-archaeologySkill
- quantum-affine-lookup-synthesisSkill
- quantum-borrowed-wire-counterexampleSkill
- quantum-dense-qromSkill
- quantum-dirty-borrowingSkill
- quantum-divstep-reconstructionSkill
- quantum-encoded-arithmeticSkill
- quantum-graded-divider-placementSkill
- quantum-guarded-permutation-auditSkill
- quantum-mbundo-lifetimesSkill
- quantum-record-miningSkill
- quantum-replay-arithmeticSkill
- quantum-residue-scan-ambiguitySkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 20
Run at least two independently contextualized blank-slate explorers and maintain three architecture fingerprints.
- Distinctive: line 23
Run the cheapest falsifier first.
- Distinctive: line 24
Before scoring a local resource cut, census every tied global owner and price every forward, inverse, cleanup, boundary, and repeated-use …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Abbruchbire (burn-the-house-down)
The figures below are board metrics of public ecdsa.fail rows (official grader, August 2026).
A local optimum rarely looks like inactivity. It can look like real but mostly sub-percent, uneven gains purchased with more machinery: wider searches, tighter certificates, longer reviews, and more elaborate explanations for why the structure itself cannot change.
That was the ECDSA.fail frontier before 2026-08-21. The leading campaigns refined one divstep lineage while their own fitted gates increasingly defined what a valid next candidate was allowed to look like.
Then a separate descent produced a different family: fixed-depth ping-pong binary division. The move added 171 qubits and removed 332,069 rounded average executed Toffoli, lowering the product from 1,477,492,400 to 1,258,525,947—14.82%. Its realized trade was 1,941.9 removed Toffoli per added qubit, against a 972.6 exact whole-jump break-even.
The method was not “try random things,” and a later strict improvement on the same family is not another burn:
- Protect the verified leader.
- Open a campaign charter with a discontinuity-sized target; by default, 5% for structural admission and 10% for the campaign.
- Run at least two independently contextualized blank-slate explorers and maintain three architecture fingerprints. A new worktree alone proves nothing.
- Turn every wall into an assumption and an executable overturn condition.
- Give explorers a score direction and a protected invariant-progress horizon, so a worse intermediate score is not an automatic veto.
- Run the cheapest falsifier first. Let proxies rank the queue, but let only the trusted evaluator certify the exact artifact.
- Before scoring a local resource cut, census every tied global owner and price every forward, inverse, cleanup, boundary, and repeated-use recomputation traversal.
- Cap incumbent grinding while structural questions remain open; reconcile each provider/account ceiling independently, and let unused machines stay idle.
- Repeat the descent from earlier clean ancestors.
- Compose the top compatible pair in both orders; discovery order need not be the best application order.
- Grind only after the architecture is frozen, competitive, and passes the campaign checker.
- Name every deployed candidate as an immutable generation. A replacement gets a new manifest and exact hashes; it supersedes rather than silently editing the prior generation.
- Split qualification from execution. Qualification may fetch, build, and prove parity, but it cannot begin expensive screening. Execution requires a fresh qualification receipt, provider-local budget authority, and a bound lifecycle guard with the original absolute deadline.
- Port the winner through a clean shipping tree and re-run the trusted correctness contract.
- Land one measured increment on the new family, refresh the live frontier after it, and capture final capital plus branch/worktree inventory for both shipping sources.
The phrase is deliberately dramatic; the operation is not. Do not delete the incumbent, discard evidence, or weaken correctness. Remove the old artifact's authority over the explorer while preserving the artifact as the leader until it is genuinely beaten.
When gains shrink while machinery grows, protect the leader, challenge the assumptions behind the wall, and re-run the descent on cleared ground.
If the run produces a 0.07% same-family win, celebrate and ship it as defense. Do not use it to certify the burn. Structural classification requires a different resource equation that clears the declared discontinuity threshold; a COMPLETE archive additionally requires the post-jump increment, later frontier refresh, and final snapshots.
What this copy is
This is the doctrine: the fifteen rules above. The full campaign package we run it with (charter templates, architecture passports, conformance checkers and the companion skills for clean-room re-descent, portfolio management and the structural audit) lives in a separate repository and is not published here. Where a rule says "campaign checker" or "qualification receipt", write the check yourself: a script that refuses to start expensive work until the named evidence exists.
The move is Teddy Pender's, shown in public by his 21 August 2026 ecdsa.fail row. The rules are our own reading of what made such a row possible and of where our own campaigns failed to do the same; they are not his words.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

