Repository navigation
Gap analysis after Phase 1 — and the gaps this repo can close, closed - #1
Merged
Merged
Conversation
…he task ladder docs/GAPS.md records the self-audit: uncalibrated tolerance band, missing within-band damage scenario, centralized dispute resolver, grindable round seed, the skipped two-GPU replay criterion, verifier economics at zero fraud, and repo hygiene (LICENSE, CI, SECURITY.md). TASKS.md gains the items that existed only as footnotes: within-band budget scenario, production verifier daemon, multi-epoch re-join, Genesis bond values, dispute-limitation disclosure (Phase 2); security audit, VRF hardening, bonded dispute committee, mainnet key regeneration (Phase 3). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KbsW5FgMKkc9kXBhWoumjW
|
…ts and CI The within_band attack spends exactly the published tolerance band: each malicious worker biases against the honest mean at 0.9x band, so replay audits pass by construction and only aggregation pushes back. The --band-sweep experiment measures loss damage across band widths with explicit replay evidence per band. economy.py gains the zero-fraud projection (audit fees are pure treasury burn when there is nobody to slash, ~9.2% of rewards at p=0.1) and genesis_parameters() publishing the Genesis operating point under the (1-p)/p discipline. 28-test suite over aggregation, attacks, replay verification, economy and sparse transport runs on CPU torch with a synthetic corpus; CI runs it on every PR. SECURITY.md opens a private-disclosure channel, CONTRIBUTING.md documents the workflow, PRD records the Phase 1 as-built substitution, and TASKS/GAPS reflect what closed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KbsW5FgMKkc9kXBhWoumjW
30 rounds, 5/16 coalition at 0.9x the band, replay audits blind by construction (distance = 0.9x band, zero fraud verdicts), coalition accepted 100% throughout. Damage vs the 2.173 honest reference: +0.012 at band 0.02, +0.019 at the 0.05 operating point (~0.9%), +0.032 at 0.1, +0.063 at 0.2 — roughly linear in band width. Curve and replay evidence promoted to docs/assets; README, WHITEPAPER honest-limitations, TASKS and GAPS now quote the measured number instead of admitting it qualitatively. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KbsW5FgMKkc9kXBhWoumjW
kh0ra
changed the title
kh0ra
marked this pull request as ready for review
July 20, 2026 15:06
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
2 participants
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.
What
Started as a self-audit after Phase 1 closed on devnet; now also implements every gap that lives in this repo.
The audit (
docs/GAPS.md, updateddocs/TASKS.md): uncalibrated tolerance band, missing within-band damage scenario, main_authority-gated dispute resolution through Genesis, grindable round seed with no hardening task, the two-GPU replay criterion folded into 1.9 and never run, zero-fraud verifier income sustainability, repo hygiene. Footnote-only items from CODEMAP/1.6/1.9 became real checklist entries (Phase 2: verifier daemon, multi-epoch re-join, two-GPU replay, Genesis bond values, dispute disclosure; Phase 3: independent Anchor audit, VRF hardening, bonded dispute committee, mainnet key regeneration).The implementations:
--band-sweep): a 5/16 coalition biasing against the honest mean at 0.9x the published band passes every replay audit by construction (measured distance = 0.9x band, zero fraud verdicts) and is 100% accepted by the aggregator. Measured damage over 30 rounds vs the 2.173 honest reference: +0.012 at band 0.02, +0.019 at the 0.05 operating point (~0.9%), +0.032 at 0.1, +0.063 at 0.2 — roughly linear. Curve promoted todocs/assets/band_sweep.png; README, whitepaper honest-limitations, TASKS and GAPS now quote the measured number instead of admitting it qualitatively.audit_burn_projection()(zero-fraud equilibrium: audit fees are pure treasury burn, ~9.2% of rewards at p=0.1, preset-independent) andgenesis_parameters()(1B preset, p=0.1, band 0.05, bond = 9 rounds of reward), both emitted into results.json..github/workflows/ci.ymlruns it on every PR.Still open by design: the LICENSE decision (owner call), and everything living in leviathan-net or requiring hardware.
Why
A checklist item that is not on the checklist is a lost item — and a limitation admitted qualitatively is cheaper to admit than one that is measured. ARCHITECTURE.md claimed "sim/ tracks how much loss damage fits inside a given band"; now it actually does.
🤖 Generated with Claude Code
https://claude.ai/code/session_01KbsW5FgMKkc9kXBhWoumjW