diff --git a/.gitignore b/.gitignore index 3482d5e..3ea1f31 100644 --- a/.gitignore +++ b/.gitignore @@ -68,3 +68,16 @@ audit-scratch/ # Playwright MCP session output (page snapshots taken while verifying a view). .playwright-mcp/ + +# Working notes, reviews, audits and handovers stay on the workstation; only +# the English operational docs listed below are published. +/docs/* +!/docs/devnet-rollouts.md +!/docs/MONGO_BACKUP_RUNBOOK.md +!/docs/NGINX_HEADERS_RUNBOOK.md + +# Session handoffs, the working plan and concept material: local only. +/plan.md +/HANDOFF*.md +/deliverables/ +/.workflows/ diff --git a/.workflows/scale-150-runbook.js b/.workflows/scale-150-runbook.js deleted file mode 100644 index 8524dc2..0000000 --- a/.workflows/scale-150-runbook.js +++ /dev/null @@ -1,244 +0,0 @@ -export const meta = { - name: 'scale-150-runbook', - description: 'Investigate the live devnet fleet read-only and produce a verified runbook for scaling 80 -> ~150 masternodes across 8 new VPS', - phases: [ - { title: 'Investigate', detail: 'six parallel read-only investigators over the live fleet' }, - { title: 'Verify', detail: 'adversarial check of each dimension\u2019s claims' }, - { title: 'Synthesize', detail: 'merge into one executable runbook' }, - ], -} - -const CONTEXT = ` -CONTEXT — DeFCoN devnet "defcon-q60" fleet, READ-ONLY INVESTIGATION. - -Access (all already configured, use Bash): -- \`ssh devnet\` -> the explorer/seed VPS, you are ROOT. Runs: defcond-devnet.service (seed, BDB binary - /usr/local/bin/defcond, datadir /home/defcon/.defcon) and defcond-devnet2.service (node2, nobdb binary - /usr/local/bin/defcond-nobdb, datadir /home/defcon/.defcon2). Node RPC: - \`sudo -n -u defcon /usr/local/bin/defcon-cli -datadir=/home/defcon/.defcon \` - Explorer API on 127.0.0.1:4100 (curl). -- \`ssh devnet-jump\` -> the jump host, ROOT. Fleet inventory: /root/fleet-nodes.txt (8 IPs, one per line). - Fleet key: /root/.ssh/defcon_nodes. Reach a fleet host FROM the jump: - ssh devnet-jump "ssh -n -i /root/.ssh/defcon_nodes -o StrictHostKeyChecking=no root@ ''" -- Fleet hosts: each runs 11 systemd instances defcon-devnet-mn@1..11 (10 masternodes + instance 11 = a - staker, because a masternode cannot stake). Binaries in /opt/defcon-devnet/bin/, per-instance datadir - /opt/defcon-devnet/mn/ with defcon.conf. -- The project's hard-won ops lessons live in d:\\www\\devnet .deftrack\\CLAUDE.md — READ IT (Read tool) for - the "Operational notes earned the hard way" and "Where things run" sections. They record real failures. - -CURRENT STATE: 8 fleet hosts x 11 instances = 88 services, plus seed + node2 = 90 daemons. 80 masternodes -enabled. Chain tip ~5553. Fleet binary md5 d1be1264468ab1ee0119110f7b467837 (--without-bdb build). - -GOAL BEING PLANNED: the operator will supply 8 NEW VPS with ssh access, to raise the masternode count from -80 to roughly 150 (so ~70 new masternodes, ~9 per new host). Your job is to establish, from EVIDENCE on the -live systems, exactly what a new host must look like and what the operation requires. - -RULES: -- READ-ONLY. Never modify a file, never restart a service, never send a transaction, never register anything. - Only inspect: cat/grep/systemctl cat/systemctl is-active/ls/RPC read calls/curl GETs. -- Ground every claim in evidence you actually observed. Quote the command and the real output. -- If you could not verify something, say so in "gaps" — do NOT guess. A guessed port range or conf line - would be executed later against 8 production hosts. -` - -const FINDINGS_SCHEMA = { - type: 'object', - properties: { - dimension: { type: 'string' }, - facts: { - type: 'array', - items: { - type: 'object', - properties: { - claim: { type: 'string', description: 'One concrete, actionable fact' }, - evidence: { type: 'string', description: 'The command run and the actual output that proves it' }, - }, - required: ['claim', 'evidence'], - }, - }, - gaps: { type: 'array', items: { type: 'string' }, description: 'What you could NOT verify and why' }, - risks: { type: 'array', items: { type: 'string' }, description: 'What could go wrong in execution' }, - }, - required: ['dimension', 'facts', 'gaps', 'risks'], -} - -const VERDICT_SCHEMA = { - type: 'object', - properties: { - dimension: { type: 'string' }, - confirmed: { type: 'array', items: { type: 'string' } }, - refuted: { - type: 'array', - items: { - type: 'object', - properties: { claim: { type: 'string' }, why: { type: 'string' }, correction: { type: 'string' } }, - required: ['claim', 'why'], - }, - }, - missed: { type: 'array', items: { type: 'string' }, description: 'Important things the investigator did not check' }, - }, - required: ['dimension', 'confirmed', 'refuted', 'missed'], -} - -const DIMENSIONS = [ - { - key: 'host-layout', - prompt: `Establish exactly what a fleet HOST looks like, so an identical new one can be built. -Inspect a real fleet host (take the first IP from the jump inventory): -- the full systemd unit: \`systemctl cat defcon-devnet-mn@\` — every line: ExecStart, User, LimitNOFILE, - Restart, ExecStartPost, WorkingDirectory, environment. -- the on-disk layout under /opt/defcon-devnet: bin/ contents (ls -la, md5sum of defcond and defcon-cli), - the mn directories, ownership and permissions. -- which OS/release the host runs (cat /etc/os-release), and whether libdb is present (this matters: the - fleet uses a --without-bdb build). -- \`ldd /opt/defcon-devnet/bin/defcond | grep -c "not found"\` to confirm the binary's runtime deps resolve. -- how many instances are enabled/active: \`systemctl is-active defcon-devnet-mn@{1..11}\`. -- whether there is any provisioning script or README left on the host that built this layout. -- ufw/firewall state (\`ufw status\` or iptables -S | head) — CLAUDE.md warns two of eight hosts filter. -Report the exact reproducible layout a new host needs.`, - }, - { - key: 'node-config', - prompt: `Establish the exact per-instance node CONFIG, so a new instance's defcon.conf can be written correctly. -On a real fleet host, read several instance configs (at least mn1, mn5, mn10 and the staker mn11) and diff -them mentally: -- \`cat /opt/defcon-devnet/mn1/defcon.conf\` etc. Report EVERY key, and which keys differ per instance - (port, externalip, masternodeblsprivkey, rpcport, ...) versus which are identical everywhere. -- the exact PORT ASSIGNMENT scheme: which P2P port and which RPC port each instance N uses. CLAUDE.md - mentions ports 19799-19808. Verify against the real configs and against listening sockets - (\`ss -lntp | grep defcond\` on the host). -- how staking differs on instance 11 (the staker) versus the masternodes 1-10 — which keys are present or - absent (masternodeblsprivkey, staking, disablewallet), and whether ExecStartPost enables staking. -- confirm dslactivationheight and maxconnections are present and their values. -- REDACT every secret: never output a masternodeblsprivkey, rpcpassword or any key material. Report only - that the key EXISTS and its config key name. -Report a precise template for a new instance's config with the per-instance variables named.`, - }, - { - key: 'funding', - prompt: `Establish exactly how ~70 new masternode COLLATERALS can be funded from the seed wallet. -On the seed (ssh devnet, RPC as shown in the context): -- \`getbalances\`, and the shape of the UTXO set: \`listunspent 26\` — how many outputs, what denominations - (bucket them), how many are >= 1000000, what the typical amount is. Use python3 to aggregate; do NOT - print the whole list. -- the devnet collateral amount for a REGULAR masternode (1,000,000 DFCN per CLAUDE.md) — confirm from the - chain if you can (e.g. an existing masternode's collateral output value via \`protx list registered 1\` - then \`gettxout\` on one collateralHash/collateralIndex). -- READ CLAUDE.md carefully for the funding traps and report them precisely: the coinstake-maturity trap - ("never let the wallet pick its own inputs on the seed node"), why sendmany can return a txid that never - reaches the mempool, the COINBASE_MATURITY depth rule, the stakeValueRange caveat, and the - WAIT_FOR_ISLOCK_TIMEOUT 10-minute mining delay that applies to every transaction once masternodes exist. -- work out the arithmetic: to create ~70 outputs of exactly 1,000,000 DFCN, how many transactions are - needed given the real UTXO denominations, and what the per-transaction output limit implies. -- check whether the seed wallet is currently staking (\`getstakinginfo\`) since that is what creates the - immature coinstake outputs that poison automatic coin selection. -Report a concrete, safe funding procedure with explicit input selection.`, - }, - { - key: 'registration', - prompt: `Establish exactly how the existing 80 masternodes were REGISTERED, so ~70 more can be registered the same way. -- On the seed, inspect an existing registration on-chain: \`protx list registered 1\` (limit output with - python3 — print ONE entry pretty-printed, plus the total count). Identify the fields: proTxHash, - collateralHash/collateralIndex, service (ip:port), ownerAddress, votingAddress, payoutAddress, - pubKeyOperator, operatorReward, and the state block. -- Determine whether all 80 share one payout address (CLAUDE.md says every masternode here shares one payout - address) — verify by aggregating payoutAddress across all entries. -- Determine which protx RPC variant is appropriate: run \`defcon-cli help protx\` and report the available - subcommands and the exact argument list of \`protx register_fund\` and \`protx register\` (use - \`defcon-cli help "protx register_fund"\`). State precisely which arguments are needed and in what order. -- Establish how the BLS operator keypair is produced: \`defcon-cli help bls\` / \`bls generate\`. Do NOT - generate or print any key — just report the command and its output field names. -- Look for any leftover registration script or notes on the seed or jump host (find /root /home -maxdepth 3 - -iname "*regist*" -o -iname "*mn*setup*" 2>/dev/null | head) that shows how the original 80 were done. -- Note the ISLock 10-minute mining delay implication for registering ~70 masternodes. -Report a concrete registration procedure, including what must be unique per masternode.`, - }, - { - key: 'ops-integration', - prompt: `Establish what must be UPDATED OUTSIDE the new hosts when the fleet grows from 8 to 16 hosts. -- The jump host inventory: \`cat /root/fleet-nodes.txt\` (count lines, show the format — IPs are not secret - within this investigation but note that the explorer is public and must use host LABELS, never IPs). -- Read d:\\www\\devnet .deftrack\\ops\\fleet-deploy.sh (Read tool) and report exactly what it expects: - which env vars, the FLEET_INSTANCES default, what it does per host, and what would happen if a new host - were added to the inventory but had a different instance count. -- The explorer's operator attribution: how host IP maps to an operator label. Query the explorer admin/API - read endpoints on the seed (curl 127.0.0.1:4100/api/v1/... — find the operators endpoint) and read - d:\\www\\devnet .deftrack\\server\\src\\domain\\operatorIndex.ts to establish how a NEW host would be - attributed and what has to be declared so 70 new masternodes are not all "unattributed". -- Check the explorer's masternode-count-dependent config: read d:\\www\\devnet .deftrack\\server\\src\\config\\llmq.ts - and report anything that assumes a fixed masternode count or profile size that a jump to 150 would break. -Report the exact list of off-host updates required.`, - }, - { - key: 'risk', - prompt: `Establish the FAILURE MODES of adding ~70 masternodes at once to this live chain, and how to avoid them. -- Read d:\\www\\devnet .deftrack\\CLAUDE.md in full, especially "Operational notes earned the hard way" and - "Measurement caveats that are easy to get wrong". Extract every lesson that bears on adding many - masternodes at once — in particular the PoSe ban-wave history, the "do not measure in the first rounds - after a revive or a restart" rule (46 masternodes revived at height 2404 caused a punishment cascade), - the dkgBadVotesThreshold history, and the masternode maxconnections requirement. -- On the seed, establish the CURRENT quorum profile parameters actually in force: \`defcon-cli help quorum\`, - then read the live profiles the chain uses. Specifically get the LLMQ types and sizes in play - (llmq_50_60, llmq_60_75, llmq_400_60, llmq_defcon) via \`quorum list\` and report how many quorums of each - type currently exist, and how a jump from 80 to 150 candidates changes selection. -- Establish the DSL implication: the service commitment carries a bitfield sized to the masternode list - (currently size=80 per the explorer /api/v1/dsl/epochs). Reason precisely about what happens to the DSL - measurement while ~70 nodes join and have not yet announced liveness. -- Establish the practical constraint: how long ~70 registrations take given the 10-minute ISLock mining - delay per transaction batch, and whether registrations can be batched. -Report a prioritized risk list with a concrete mitigation for each.`, - }, -] - -phase('Investigate') -log(`Investigating ${DIMENSIONS.length} dimensions of the live fleet, read-only`) - -const investigated = await pipeline( - DIMENSIONS, - (d) => - agent(`${CONTEXT}\n\nYOUR DIMENSION: ${d.key}\n\n${d.prompt}`, { - label: `investigate:${d.key}`, - phase: 'Investigate', - schema: FINDINGS_SCHEMA, - }), - (findings, d) => { - if (!findings) return null - return agent( - `${CONTEXT}\n\nYou are ADVERSARIALLY VERIFYING another investigator's report on dimension "${d.key}".\n\n` + - `Their report:\n${JSON.stringify(findings, null, 2)}\n\n` + - `Independently re-check the load-bearing claims by running the commands YOURSELF (read-only). ` + - `A claim that would be executed against 8 production hosts must not rest on one reading. ` + - `Refute anything wrong, imprecise, or unverifiable, and give the correction. ` + - `Then list what the investigator MISSED that matters for building 8 new hosts and registering ~70 masternodes. ` + - `Be skeptical: default to refuting a claim you cannot reproduce.`, - { label: `verify:${d.key}`, phase: 'Verify', schema: VERDICT_SCHEMA } - ).then((verdict) => ({ dimension: d.key, findings, verdict })) - } -) - -const results = investigated.filter(Boolean) -log(`${results.length}/${DIMENSIONS.length} dimensions investigated and verified`) - -phase('Synthesize') -const runbook = await agent( - `${CONTEXT}\n\nSix investigators mapped the live fleet and an adversarial verifier re-checked each. ` + - `Here is everything, including refutations and corrections:\n\n${JSON.stringify(results, null, 2)}\n\n` + - `Write the RUNBOOK for scaling this devnet from 80 to ~150 masternodes across 8 new VPS.\n\n` + - `Requirements:\n` + - `- Where the verifier refuted or corrected a claim, the CORRECTION wins. Never carry a refuted claim forward.\n` + - `- Ordered phases with explicit commands, each grounded in what was actually observed. Mark anything still ` + - `unverified as [UNVERIFIED - confirm before running] rather than presenting it as fact.\n` + - `- Cover: (1) prepare each new host, (2) fund ~70 collaterals of exactly 1,000,000 DFCN safely, ` + - `(3) register ~70 masternodes, (4) update the jump inventory and operator attribution, ` + - `(5) verify the result, (6) what to do if it goes wrong (rollback/triage).\n` + - `- A dedicated section on TIMING and MEASUREMENT: the ISLock 10-minute rule, how long the whole operation ` + - `realistically takes, and precisely what happens to the in-flight DSL shadow measurement and the DKG/PoSe ` + - `readings while the new nodes join — including which readings must be excluded as artefacts.\n` + - `- A pre-flight CHECKLIST of things to confirm before touching anything.\n` + - `- Call out every irreversible step.\n` + - `- Secrets: never include key material; say where a key is generated and that it stays on the host.\n\n` + - `Be precise and executable. This will be run against 8 production hosts on a live chain.`, - { label: 'runbook', phase: 'Synthesize' } -) - -return { dimensionsCovered: results.map((r) => r.dimension), runbook } diff --git a/CLAUDE.md b/CLAUDE.md index 82bd43e..bcaeb73 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -4,7 +4,7 @@ Working notes for this repository. Everything here was verified against source, not assumed — the "Verified facts" section carries file:line references so it can be re-checked when the node is upgraded. -**Open work lives in [`plan.md`](plan.md)**: the owed experiment runs and what +**Open work lives in `plan.md`** (a local file, not in this repository): the owed experiment runs and what blocks each, what in the current binary is still unproven and why, the tooling debts every measurement exposed, and the v23/M-02 items parked by decision. This file records what is known; `plan.md` records what is not yet done. Read it @@ -1154,7 +1154,7 @@ The explorer database is backed up nightly since 2026-09-14. `ops/mongo-backup.s `/usr/local/sbin/deftrack-mongo-backup`, with the `deftrack-mongo-backup.{service,timer}` units. It runs mongodump as the read-only user, checks the archive before pruning, and writes a `.sha256` and a manifest beside each archive. The procedure, restore check and rollback are in -[`docs/MONGO_BACKUP_RUNBOOK_HU.md`](docs/MONGO_BACKUP_RUNBOOK_HU.md). The archives hold non-public host +[`docs/MONGO_BACKUP_RUNBOOK.md`](docs/MONGO_BACKUP_RUNBOOK.md). The archives hold non-public host addresses, so every copy that leaves the VPS must stay private. **Measure an RPC client change against its mechanism, not only against the error count.** #181 closes pooled diff --git a/HANDOFF-v23-audit-2026-09-04.md b/HANDOFF-v23-audit-2026-09-04.md deleted file mode 100644 index 566263f..0000000 --- a/HANDOFF-v23-audit-2026-09-04.md +++ /dev/null @@ -1,257 +0,0 @@ -# HANDOFF — DeFCoN Core v23 pre-release audit - -**Written** 2026-09-04. **Repo** `defcon-project/defcon`, branch `v22.1.x`, -tip at handoff **`27f6e7b10a`** (#192). Working tree `~/DEFCON` in WSL Ubuntu. - -Written in English to match the repo, commit messages and PR text. - -> **Correction, 2026-09-07.** The body below is left as written on 2026-09-04; -> one of its headline risks has since been measured out of existence. -> -> **The mainnet premine cannot be frozen, because it is already spent.** -> Verified on 2026-09-05 against a copy of mainnet at height 130,100, with a -> control: `gettxout` on the coinbases of blocks 1, 2, 450, 899 and 900 all -> answer spent, and `scantxoutset` on the premine script returns -> `0.00000000` while a control script returns a real balance. The #184 -> finding stands as a bug -- a strict rule that compared against 96 bytes -> would have refused every genuine 97-byte BLS spend -- but the sentence -> "activating the gate would have made those outputs unspendable and split -> the chain on the first spend" describes a spend that can no longer happen. -> What remains open, and is the actual reason M-02 stays out of v23 (owner -> decision, 2026-09-05), is whether any *other* unspent BLS-locked output -> exists on mainnet: `dumptxoutset` plus a parser proven on a known output, -> not `scantxoutset`, which cannot filter by type. See `plan.md` §5. - ---- - -## 1. The task - -Audit the node before the v23 release, which will carry the first coordinated -mainnet consensus activation this chain has had. Mainnet runs **v22.1.4** -(every peer reports `/DeFCoN:22.1.4/`, protocol 70239); the release will be cut -from the `v22.1.x` tip. The audit had to answer two questions: - -1. Is there anything in the code that makes v23 unsafe to release? -2. Does v23 validate the entire existing mainnet history exactly as v22.1.4 - does — i.e. did any change since that tag alter validity without an - activation height to separate the two? - -Both are now answered. The bug-hunting phase is closed. - ---- - -## 2. Completed - -### 2.1 Five blockers found and fixed (all merged) - -| PR | What it was | -|---|---| -| **#184** | **M-02 would have frozen every BLS output, and the premine is one.** A script carries a signature with its hash type appended, so a BLS signature reaches the encoding check at **97** bytes; the strict rule compared against 96. With the gate active every genuine BLS spend failed as `SCRIPT_ERR_SIG_DER`. `consensus.premineAddress` is a 48-byte key under `OP_CHECKSIG` — a `BLSPUBKEY` output — on all four networks, and mainnet paid its first 900 blocks into it. Activating the gate would have made those outputs unspendable and split the chain on the first spend. | -| **#185** | **The evodb verifier was innocent; two tests were stale.** `evodb_diff_round_trip` planted corruption under `dmn_D3` while `DB_LIST_DIFF` has been `dmn_D5` since the DSL migration — it corrupted nothing, the verifier rightly said clean, and the test then reported it for missing damage that did not exist. `verify_db_legacy` spent a collateral one block after registering it, which `minStaticCollateral` (8064) correctly refuses. Suite is back in the CI gate; it runs in 8 s and was never the "long mining loop" its exclusion claimed. | -| **#186** | **Q60 formation was hard-wired devnet-only.** `IsQuorumTypeEnabledInternal` returned false for `LLMQ_DEFCON` on every non-devnet network *before* reading the activation height, so a mainnet height would have formed no quorum and forked nobody — and at the height the resolver would have flipped to a profile with no quorums, stopping ChainLocks silently. Formation is now keyed on the height alone. Added `CheckLLMQConfiguration`, which refuses to start on a role naming an unregistered profile, on either half of a switchover without the other, or on an InstantSend switchover below the ChainLock one. | -| **#188** | **The proof-of-stake nonce rule is no longer optional.** #120 rightly made the header proof-of-work requirement follow height, but above `lastPowBlock` that removed the check entirely — so a header with a non-zero nonce entered the index, which reads a nonce the same way, stored it as proof-of-work, and `LoadBlockIndexGuts` then failed the whole index load on the next restart. **One unsolicited header would stop a node from starting.** The gate that closes it was unset on mainnet, so the release would have shipped that open. Gate removed; the rule now follows `lastPowBlock` everywhere. | -| **#192** | **A refused ChainLock must leave no trace.** `ProcessNewChainLock` wrote `bestChainLock` *before* checking that the named block sits at the claimed height, so a refused signature stayed as the node's best lock — suppressing every later lock at that height and relaxing the superblock check in `ConnectBlock`, which reads that height. Also, a CLSIG we could not verify because we lack the quorum was punished like a forgery and remembered as seen. Both inherited from Dash and still present in their v23 rewrite; no upstream patch exists. | - -Also merged the same day by another session: **#189, #190, #191** (Sentinel -Layer). All classified; none touches mainnet — DSL activation and enforcement -heights remain unreachable there. - -### 2.2 Classification of every consensus-path change since v22.1.4 - -Done by net diff per validity-deciding file, not commit by commit. Result: -**exactly one ungated change had a real consequence** (#120, fixed by #188), and -**one is ungated but confined to paths that were already broken** (#119, settled -by the reindex below). Everything else is gated, unreachable on mainnet, or not -consensus. - -Facts worth keeping, because the safety often rests on something non-obvious: - -- **#124 retiring the Evo masternode type cannot invalidate history**, because - Evo was never registrable on mainnet: it required `BASIC_BLS_VERSION`, which - requires V19, and `V19Height` is unreachable on mainnet and testnet. -- **`simplifiedmns.h` could have broken every block and does not.** The entry is - hashed into `cbTx.merkleRootMNList`, which consensus checks. Its changed - fields are written only inside `if (obj.nVersion == BASIC_BLS_VERSION)` — - again unreachable without V19 — so mainnet serialization is byte-identical. -- **`GetMNPayee`'s multi-slot branch went from never-executed to always-executed - on mainnet.** The outcome is unchanged only because no registrable type has a - payment weight above 1. **Compute has 5.** This is not dead code; it is code - waiting for a gate. -- `llmq_50_60` / `llmq_60_75` bad-votes thresholds changed ungated (3 → 40/48) - but both profiles are testnet/devnet-only and the sole reader is the DKG - session. No mainnet-enabled profile changed shape. -- `serialize.h`: no block or transaction field gained a bound. `LIMITED_VECTOR` - is applied only to spork signatures, bloom filters, one P2P vector and the - dormant Compute descriptor. - -### 2.3 Both mainnet validation tests passed - -Run against `b70a7e292e` with a throwaway build in `~/DEFCON-v23test`. - -- **Migration** (what every operator will do): v23 started on a genuine v22.1.4 - datadir with no reindex, loaded in 5 s, migrated `dmn_S3`/`dmn_D3` → - `dmn_S5`/`dmn_D5` with the marker moving `b_b4` → `b_b6`, and reached the same - tip. -- **Full reindex** (`-reindex -assumevalid=0 -connect=0 -txindex=1`): rebuilt - 130 100 blocks from raw block files in **190 s**, verifying every signature. - **Zero validation problems.** Same tip hash. The masternode list derived from - blocks alone came out **220 total / 138 enabled**, the enabled count matching - the live explorer that day. - -Two independent derivations agree. This settles #119, #120's stricter half, -#188's rule across all ~129k proof-of-stake blocks, and the CbTx merkle roots. - ---- - -## 3. Decisions taken - -- **Q60 goes to mainnet in v23** (user, 2026-09-04). This makes v23 a mandatory, - coordinated upgrade. -- **K-03 (the `GetBlockProof` chainwork weighting) stays out of v23**, recorded - as accepted risk. Skipping leaves v23 *strictly better* than what mainnet runs - today; doing it would require re-deriving `nMinimumChainWork` in the same - commit, and getting that wrong freezes every restarted node in permanent IBD. - Compensating controls: ChainLocks, the checkpoint at 103536, the fork-recovery - anchor from 113117. -- **The nonce rule was ungated rather than scheduled**, because leaving it - optional is what was dangerous and no block on any network violates it. -- **Registration and activation height for Q60 must land in one commit.** #186's - startup check now makes separating them impossible by accident. -- **DSL enforcement is testable but not advisable anywhere.** The commitment - still carries only a `missed` bitfield, so "no evidence" heals; the unmetered - `POSECHALLENGE` is still open. - ---- - -## 4. What remains - -### 4.1 Blocked on the user — the Q60 mainnet activation height - -This is the only real consensus event in v23. - -- Pick **H as a multiple of 24**. -- **Non-upgraded nodes fall off at H − 110**, not at H: quorum formation opens - at H − 120 and commitments of a type they do not know are rejected as - `bad-qc-commitment-type`. They strand on an invalid tip and need a reindex - after upgrading. **H − 120 is the upgrade deadline to communicate.** -- Q60 needs **44 valid masternodes of 60** to form, with no fallback. Mainnet had - 138 enabled / 220 total, so there is roughly 3× headroom — but 79 were - PoSe-banned. Watch mainnet quorum health for a few weeks before fixing H. -- Then one commit: `AddLLMQ(LLMQ_DEFCON)` on mainnet (and testnet), - `llmqTypeChainLocksV2`, `nChainLocksV2ActivationHeight`, - `llmqTypeDIP0024InstantSendV2`, `nInstantSendV2ActivationHeight` - (≥ the ChainLock height), and `WAIT_FOR_ISLOCK_TIMEOUT` 600 → 120. - -### 4.2 Release hygiene — can be done now, the tree is free - -- **`_CLIENT_VERSION_IS_RELEASE` → false until tagging.** This bit us in - practice: a datadir could not be attributed to a build because every - development build reports `v22.1.5`. Only the evodb keys revealed the truth. -- **`chainTxData`** is still Dash's, so a fully synced mainnet node reports - `verificationprogress: 0.0043`. Recompute with `getchaintxstats`. -- `nMinimumChainWork`, `defaultAssumeValid`, checkpoints — refresh **at tagging** - from the then-current chain, not before. -- Release notes: `doc/release-notes.md` is still Dash's v22.1.2. Nothing exists - for v22.1.3/4/5 or v23. -- **Issue #15** — the published Linux tarball for release `6ea2f52` still - contains zero-filled `defcond`, `defcon-wallet` and `defcon-qt`. No process - change followed. Add an artifact smoke test (`file` + `ldd` + `-version` on a - target host) and record the md5 in the release notes. -- Note for release notes: #183 moved the **mainnet Tor listen port 9996 → 8189**. - -### 4.3 Must be repeated on the release candidate - -Both validation tests bind only the commit they ran against. Re-run them on the -tagged binary. Adding the Q60 activation at a height above the tip should not -change either result, since the gate is never reached. - ---- - -## 5. Exact next steps - -```bash -# 1. Rebuild the test binary at the release candidate (throwaway tree, -# never touches ~/DEFCON) -cd ~/DEFCON-v23test && git fetch --depth 1 origin v22.1.x && git reset --hard FETCH_HEAD -make -C src -j10 defcond defcon-cli - -# 2. Fresh copies of the pristine datadir — migration is one-way, so never -# run against d:/x/Defcon itself -rm -rf ~/mn-test && mkdir -p ~/mn-test/migrate ~/mn-test/reindex/blocks -cp -r /mnt/d/x/Defcon/{blocks,chainstate,evodb,indexes,llmq,sporks.dat,mncache.dat,governance.dat} ~/mn-test/migrate/ -cp /mnt/d/x/Defcon/blocks/blk*.dat /mnt/d/x/Defcon/blocks/rev*.dat ~/mn-test/reindex/blocks/ - -# 3. Migration test — must reach the same tip and move dmn_S3/D3 -> dmn_S5/D5 -~/DEFCON-v23test/src/defcond -datadir=$HOME/mn-test/migrate \ - -connect=0 -listen=0 -disablewallet -txindex=1 -daemon=0 -printtoconsole=1 - -# 4. Reindex test — -assumevalid=0 matters, without it script checks below the -# fork anchor are skipped, which is exactly the BLS path M-02 changed -~/DEFCON-v23test/src/defcond -datadir=$HOME/mn-test/reindex \ - -reindex -assumevalid=0 -connect=0 -listen=0 -disablewallet -txindex=1 \ - -daemon=0 -printtoconsole=1 -``` - -Watch for `ERROR: ConnectBlock`, `txn-spends-premature-collateral`, -`bad-pos-nonce`, a proof-of-work failure below height 999, or any CbTx merkle -mismatch. Success is reaching the datadir's own tip hash. - -Ready-made scripts from this session are in the session scratchpad -(`migrate.sh`, `reindex.sh`, `cigate.sh`); the shape is worth keeping. - ---- - -## 6. Artefacts and where things are - -| What | Where | Note | -|---|---|---| -| Pristine **v22.1.4** mainnet datadir | `d:\x\Defcon\` | height 130100, tip `d48a0fc4…1223d1`, evodb `dmn_S3`/`dmn_D3`, marker `b_b4`, clean shutdown. **Never run a test against it directly** — migration is one-way. Contains a small `teszt` wallet; copies exclude it. | -| The **v22.1.4 binary** that produced it | user's machine | More valuable than the datadir: with it a fresh master copy can be regenerated at any height. Keep it. | -| Throwaway v23 build tree | `~/DEFCON-v23test` | Independent shallow clone, `--without-bdb --with-gui=no --disable-tests --disable-bench`. Never touches `~/DEFCON`. | -| Stash and branch backups | `~/stash-backup-20260904/` | Three stash patches plus the commit list of the deleted `release/v22.1.5`. All verified to hold nothing unique; deletable once nobody misses them. | - ---- - -## 7. Open risks - -- **The largest risk is not in v23 — it is live on mainnet now.** The ChainLock - quorum is `llmq_400_60` at **size 400 / minSize 4 / threshold 3**, and the - quorum is capped by the masternode count, so with 138 enabled masternodes - **any three of them can produce a ChainLock**. Nothing merged fixes this; the - Q60 activation does. It also set the severity of #192. -- **Two gates must stay unset**, or code that is currently inert goes live: - `nComputeNodeActivationHeight` (makes `GetMNPayee`'s multi-slot branch - meaningful) and `nDSLActivationHeight` / `nDSLEnforcementHeight` (opens the - type-10 path, whose empty-vin exemption in `tx_check.cpp` is genuinely - ungated). -- **The audit covered changes since v22.1.4.** Code that predates the tag was - not re-audited, and **16 unit suites remain excluded from CI**, including - `transaction_tests`, `validation_chainstate_tests`, `miner_tests` and - `denialofservice_tests`. An audit finding nothing more does not prove there is - nothing more. -- CI builds **linux64 only** and runs **no functional tests**, while releases - ship win64 and macOS. -- `spork@{1}` in the deleted stashes carried 12 files the reverse-apply check - called unlanded; that check is unreliable after months of drift, and the patch - is saved. Worth a proper look if anything seems missing. - ---- - -## 8. Working notes for whoever continues - -- **Verify the verifier.** Three times in this project a tool gave a clean wrong - answer. Twice today the thing that looked broken was innocent: the evodb - verifier (#185) and the ChainLock race branch the user asked about. Hand-check - one instance and run a negative control before believing any measurement. -- **Every fix here shipped with a negative control** — the change reverted, the - new test required to fail, and only it. Keep that. -- A restriction rule needs **both halves proven**: that it rejects what it must - reject, and that it accepts what it must accept. M-02 had only the first, and - that is exactly how it nearly froze the premine. -- The WSL sibling trees `~/DEFCON-seed`, `~/DEFCON-fleet`, `~/DEFCON-qt` are - **git worktrees sharing `~/DEFCON/.git`** — checking out in them writes to the - shared git directory. They now sit at the current tip, which means a rebuild - there would produce a binary different from what the devnet fleet runs. - **The devnet has deliberately not been updated.** -- `wsl -d Ubuntu -- bash -c '...'` expands `$VAR` and `$(...)` on the Windows - side. Write a script file and run it instead. diff --git a/HANDOFF.md b/HANDOFF.md deleted file mode 100644 index bb3cfa5..0000000 --- a/HANDOFF.md +++ /dev/null @@ -1,237 +0,0 @@ -# HANDOFF — devnet session, 2026-09-05 - -**Repo** this one (`devnet .deftrack`), branch `main`. **Node** `e15e29b136` -(#193) on the whole network. The previous handoff, on the DeFCoN Core v23 -pre-release audit, is preserved unchanged as -[`HANDOFF-v23-audit-2026-09-04.md`](HANDOFF-v23-audit-2026-09-04.md). - -Open work is tracked in [`plan.md`](plan.md); this file is the narrative of what -happened and what to do next. Where the two overlap, `plan.md` is the list to -work from. - ---- - -## 0. Since this handoff (2026-09-07) - -Read `plan.md` first; this section only says what moved after the narrative -below was written, so the numbers in it are not taken for current. - -- **The node is `c739d9f504` (#208) on all 162 daemons** since the 2026-09-07 - rollout (`docs/devnet-rollouts.md`, run `fleet-rollout-208-2026-09-07`). - The `e15e29b136` in section 2 is history. One masternode was banned by two - interleaved profiles ten blocks apart during the roll and revived at 9234; - the network is 152/152. -- **A passive observation is running and forbids every intervention until it - closes:** `absent-epoch-rate-post-208-2026-09-07`, boundary 9192 to 10608, - about 2026-09-09 evening UTC. No E4b, no roll, no restart, no revive, no - InstantSend probe before that; any intervened epoch comes out of the - sample. Intermediate readings at 9648, 9888 and 10128, none conclusive. -- **The explorer gained five merged changes on 2026-09-07**, all deployed - (#126 to #130): the plan and CLAUDE.md caught up (the nonce rule #162 is - ungated everywhere); the four remaining writing services got integration - tests over a real MongoDB, which found and fixed a tight loop in the - ChainLock ZMQ derivation; an operator can now confirm a reorg deeper than - the 200-block cap through `GET`/`POST /api/v1/admin/sync/rewind` (runbook - in CLAUDE.md); the experiment outcome carries the mean block interval and - the producer concentration, and the InstantSend probe lives in the repo as - `ops/instantsend-probe.py` (dry-run proven on the VPS, not yet run for - real); and the simulator's two unkept promises -- action audit events, - resume directive -- were removed and documented as absent. -- **Local integration tests need a throwaway, no-auth MongoDB** on 27018 in - WSL; the recipe is in CLAUDE.md under the local environment. -- **What is owed next**, in the order agreed: the InstantSend probe's first - real run once the observation closes; E4b only after measuring what share - of the target's quorum links are inbound (the chaos filter reaches only - those); the `defcon-enable-staking` installation on the VPS and the eight - fullnodes (no restart needed); and the two M-02 measurements in `plan.md` - §5 if a later release schedules it. - ---- - -## 1. The task - -The newest binary had just been rolled out. Two questions: what is actually in -it, and which experiments are worth running against it. That turned into six -Experiment runs, one operational change to block production, and the first real -installation of the fault-injection tooling. - ---- - -## 2. What is deployed, verified rather than assumed - -| | | -|---|---| -| Node commit | `e15e29b136` (#193) | -| Seed binary | `/usr/local/bin/defcond` md5 **f0a2d5ce** (BDB build) | -| Fleet + devnet2 | `/usr/local/bin/defcond-nobdb` md5 **356f2b25** | -| Uniformity | 16 hosts, 160 instances, one chain, 0 forked, 0 unreachable; plus seed and devnet2 = **162** | -| Masternodes | 152 total / 152 enabled throughout the session | - -Compared by `md5sum`, never by version string — all three report `v22.1.5`. - ---- - -## 3. Completed - -### 3.1 Five Experiment runs closed - -- **`sentinel-hardening-rollout-2026-09-04`** (7745 → 8004). 31 rounds formed, - 0 failed, median *and* worst health 1.00 on all four profiles, 0 PoSe - penalties, ChainLock coverage 1.00, block spacing on target. Ten consecutive - Sentinel epochs (323–332) committed after the change against a 93.5–96.6 % - baseline; the single absent epoch, 322, is the one whose signing window - contained the restart itself. -- **`bls-strict-size-spend-2026-09-04`** (7756 → 8004). #184 proven on-chain: a - BLS-locked output created and spent with the strict rule active, mined in - block **7760**, scriptSig one `OP_PUSHDATA1` push of **97 bytes**. Negative - control: the previous binary reindexes to 7759 and refuses 7760 - (`ConnectBlock: CheckQueue failed`), and offered the same transaction answers - `non-mandatory-script-verify-flag (Non-canonical DER signature)`. -- **`instantsend-lock-latency-2026-09-05`** (8015 → 8017). 20 transactions: - median **1.143 s**, p90 2.020 s, min 0.768 s, one outlier at 18.157 s. The - conflicting spend was refused with **`tx-txlock-conflict`**, not an ordinary - mempool conflict — the null result written into the hypothesis did not occur. -- **`chaos-wrapper-pilot-2026-09-05`** (8022 → 8022). Installed on one fleet - host; every refusal path fired with its exact message; a marker job retired - 16 s after expiry and a 1 ms netem fault 5 s after expiry, both with nobody - calling `clear`; `eth0` restored to `fq_codel` on its own. -- **`stake-redistribution-2026-09-05`** — **still open**, see §5. - -### 3.2 An owed verification that had been skipped - -The rollout run's `expected` demanded a full reindex on the new binary reaching -the network's tip hash, and the run was closed this morning without it. Done -afterwards on binary `356f2b25`: **8028 blocks** rebuilt from raw block files -with `-assumevalid=0` in **45 s**, **zero validation errors**, tip hash equal to -the live chain's (`a1118849…469a1e`). That single result closes #193 (every -historical DSL commitment re-derived its attesting quorum — a mismatch is -`bad-dsl-quorum-hash` and would have stopped the reindex), #164 (the -connect-time stake modifier is deterministic, which RPC cannot show) and #188. -The evidence was appended to the closed run's notes. - -### 3.3 Unit coverage confirmed on the deployed commit - -- `llmq_chainlocks_tests` — no errors. Includes `llmq_configuration_coherence` - and `formation_follows_the_height_not_the_network` (#186). -- `pos_stake_rules_tests`, `pos_multiwallet_tests` — no errors. Includes - `a_watch_only_output_is_never_offered_as_a_kernel` (#167). - -### 3.4 Findings worth more than the runs that produced them - -- **The premine key is identical on all four networks** (`chainparams.cpp` 224, - 416, 587, 924). The devnet premine is locked to the main network's own - premine key, so the originally planned "spend a premine output" test would - have required a production secret. It was rebuilt on a freshly generated BLS - key instead. -- **The mainnet premine is already spent.** Verified with a control against a - copy of mainnet at 130100: `gettxout` on the coinbases of blocks 1, 2, 450, - 899, 900 all answer spent; `scantxoutset` on the premine script returns - `0.00000000` while a control script returns 2,990,000. The v23 audit's - headline risk therefore rested on an assumption that no longer holds — and - M-02's real benefit is a different one (see `plan.md` §5). -- **One kernel script won 44 % of 250 blocks**, top five 57 %, and it is the - seed's wallet, holding 58.6 % of network stake weight — while the experiment - outcome reported `distinctStakers: 42`. - ---- - -## 4. Decisions taken - -- **The seed's staking is off, durably** (user chose this over letting it come - back at E1a's restart). `staking=0` in the seed conf plus a systemd drop-in - clearing `ExecStartPost` for `defcond-devnet` only. Verified by restarting: - service active, `ExecStartPost` empty, 175 peers, tip advancing. -- **M-02 / v23 is postponed** by the user; everything found about it is in - `plan.md` §5 so it is not lost. -- **The chaos pilot host was chosen by the inventory's own rule**, not by - preference: `fleet-nodes.txt` exists precisely "so a routine rollout does not - reach the hosts that also carry production services". Among those 11, three - carry no fleet staker — which matters while block production rests on the - staker hosts — and one of those three runs no containers. -- **The chaos pilot applied no real fault.** A marker job and a 1 ms delay were - enough to prove the recovery timer; stopping a masternode belongs to the - outage run, not to a tooling pilot. - ---- - -## 5. What remains - -Work from [`plan.md`](plan.md). The immediate items: - -1. **Close `stake-redistribution-2026-09-05`** once the LWMA-3 retarget settles - (N = 36, recomputed every block; expect two to three hours from 07:41 UTC). - The number the run exists for is the top-1 producer share: **44 % before**, - expected 10–15 % after. Difficulty had already fallen 519 M → 428 M and six - blocks after the switch came from five distinct producers, none of them the - seed. -2. **E1a**, the DSL enforcement gate. Then **E1b** (the punishing branch, first - time on any chain) and **E2** (the mass-outage guard edge pair). -3. **E4b** and the real InstantSend security test — the latter needs a partition - fault the wrapper does not implement yet. - ---- - -## 6. Exact next steps - -### 6.1 Closing the staking run - -```bash -# top-1 producer share over the last 250 blocks; run on the seed -# (the same sampling used before the switch, so the two are comparable) -# then close at the current tip and write the notes. -``` -Compare against the recorded baseline in that run's `notes`: 41 distinct kernel -scripts, top-1 110/250 = 44.0 %, top-5 57.2 %. - -### 6.2 E1a, the enforcement gate - -- Pick **H = tip + 100 rounded up to a multiple of 24** (epoch boundaries are - exactly the multiples of 24; epoch 323 began at 7752). -- Add `dslenforcementheight=H` to the **`[devnet]` section** of every conf: 160 - fleet instances, the seed, devnet2. Then restart all of them. -- **This is consensus.** `IsBanned()` reads `nDSLBanHeight` (`dmnstate.h:454`) - and `fRewardSuspended` changes payee selection, so a node started without the - argument forks — the same mechanism as `dslactivationheight`, which has - already stranded a node once on this project. -- The node refuses an enforcement height below the activation height, so the - configuration cannot be half-applied in that direction. -- The seed's restart no longer re-enables its staking; that interaction was - removed in §4. - -### 6.3 Reverting anything from this session - -```bash -# seed staking back on -rm /etc/systemd/system/defcond-devnet.service.d/no-staking.conf -sed -i 's/^staking=0$/staking=1/' /home/defcon/.defcon/defcon.conf -systemctl daemon-reload && systemctl restart defcond-devnet - -# chaos wrapper off the pilot host -bash /root/chaos/uninstall.sh && userdel chaosops -``` - ---- - -## 7. Working notes for whoever continues - -- **Remote configuration writes and service restarts are blocked in this - session's permission mode.** The seed staking change had to be handed to the - user as a script to run from their own PowerShell. Read-only SSH, the - Experiments admin API and throwaway-datadir work all pass. Expect the same for - E1a's fleet-wide conf edit and restart — plan it as a script the user runs, or - get a Bash permission rule added first. -- **Every measurement in this session that looked like a failure was checked - before being reported, and two of them were the tool's fault, not the - network's**: the InstantSend probe raced the ChainLock that prunes the record - it polls, and the first negative-control reindex died on a startup flag - (`Cannot set -bind together with -listen=0`) rather than on validation. Keep - doing that; `plan.md` §3 lists the tooling debts it exposed. -- **Experiment field limits are 2000 characters** for `hypothesis`, `expected` - and `notes`, and 1000 for `intervention.description`. Write to the limit - before posting; the API refuses rather than truncating. -- The admin API key lives in `/opt/devnet-deftrack/app/.env` on the explorer - host as `ADMIN_API_KEY`; read it there and post from the host so it never - crosses the wire. -- Scratch scripts from this session are in the session scratchpad - (`isprobe.py`, `reindex-current.sh`, `chaos-install.sh`, - `chaos-exercise.sh`, `seed-staking-off.sh`); the shapes are worth keeping. diff --git a/README.md b/README.md index b8bc782..83c6172 100644 --- a/README.md +++ b/README.md @@ -1,9 +1,53 @@ -devnet.deftrack is a DeFCoN devnet blockchain explorer that monitors DKG/LLMQ quorum formation and PoSe behaviour during masternode testing. +# devnet.deftrack -Its purpose is to measure current network behaviour and identify issues, allowing future quorum configuration changes to be compared against a real baseline. +An explorer for the DeFCoN **devnet** (`defcon-q60`), live at +[devnet.deftrack.xyz](https://devnet.deftrack.xyz). Its primary job is to record every DKG round — +including the ones that did not happen — and to attribute masternode failures, so that quorum +configuration changes can be compared against a measured baseline. It observes the network; it never +touches consensus. -DEVNET — test network. Coins have no value. +**DEVNET — a test network. Its coins have no value.** -License +## What it records + +- DKG / LLMQ rounds for every quorum profile, reconstructed from the expected schedule, so a round that + failed to form is a row, not a gap +- PoSe penalties, bans and revivals, block by block +- ChainLock coverage and InstantSend behaviour +- Sentinel Layer (service PoSe) epochs and commitments +- the Experiments record: every intervention on the network, with its hypothesis declared before the + outcome is known + +## Layout + +| Path | What | +|---|---| +| `shared/` | types shared by server and client | +| `server/` | Express + Mongoose indexer and API (`/api/v1/...`) | +| `client/` | Lit front-end, built with Vite | +| `ops/` | deployment, backup, fleet and measurement tooling | +| `docker/` | the regtest lab image | +| `docs/` | the rollout record and operational runbooks | + +## Development + +```bash +npm install # workspaces: shared, server, client +npm run typecheck # all three workspaces +npm run build # shared -> server -> client +npm run dev # server :4100 + client :5190 (Vite proxies /api) +npm test # server unit tests (vitest) +``` + +The server needs a MongoDB instance and a DeFCoN node's RPC; see `.env.example`. `CLAUDE.md` holds the +verified facts about the node that the explorer relies on, with source references. + +## Documentation + +- [`docs/devnet-rollouts.md`](docs/devnet-rollouts.md) — what has been deployed to the devnet, and when +- [`docs/MONGO_BACKUP_RUNBOOK.md`](docs/MONGO_BACKUP_RUNBOOK.md) — the nightly database backup +- [`docs/NGINX_HEADERS_RUNBOOK.md`](docs/NGINX_HEADERS_RUNBOOK.md) — security headers on the served site + +## License MIT diff --git a/client/e2e/run-status.spec.ts b/client/e2e/run-status.spec.ts index b1b6aa7..b98eb9b 100644 --- a/client/e2e/run-status.spec.ts +++ b/client/e2e/run-status.spec.ts @@ -498,7 +498,7 @@ test.describe('run status', () => { /** * Everything below was found by the final independent review (2026-09-13), * and each case failed on main before it was fixed. The review's own probes - * are in docs/review-2026-09-13; these are the gate's versions, and the order + * stayed with the review; these are the gate's versions, and the order * of the answers in them is stated with held responses rather than hoped for * with delays. */ diff --git a/deliverables/defcompute_koncepcio/assets/defcon-logo.png b/deliverables/defcompute_koncepcio/assets/defcon-logo.png deleted file mode 100644 index 3f19b42..0000000 Binary files a/deliverables/defcompute_koncepcio/assets/defcon-logo.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/index.html b/deliverables/defcompute_koncepcio/index.html deleted file mode 100644 index d346c77..0000000 --- a/deliverables/defcompute_koncepcio/index.html +++ /dev/null @@ -1,888 +0,0 @@ - - - - - - DeFCompute – decentralizált számítási hálózat - - - - - -
-
-
-
- DeFCon -
DeFCompute
-
-
-
Koncepció és megvalósítási vázlat
-

Decentralizált számítási hálózat

-

Hogyan kapcsolható össze a DeFCon masternode-hálózat, a független oracle-réteg és a világ számítási kapacitása egy ellenőrizhető piactérré?

-
- -
-
- - -
-
DeFCompute / ÁttekintésKoncepciódokumentum
-
-
Az ötlet egy oldalon
-

A kihasználatlan gépekből ellenőrizhető számítási piac

-

A szolgáltató számítási kapacitást ad el, a vevő munkát küld, a DeFCon-hálózathoz kötött szereplők pedig bizonyíthatóan ellenőrzik, hogy a munka valóban és helyesen elkészült.

- -
-
1

Szolgáltató

GPU-t, CPU-t, memóriát és tárhelyet kínál.

-
4M

Oracle node

Közvetít, mér, tárol és bizonyítékot replikál.

-
5/7

Validátorok

Hét 1M node közül öt jóváhagyása szükséges.

-
✓

Elszámolás

Csak elfogadott eredmény után történik fizetés.

-
- -
-
-
A lényeg
-

Aki kiszolgál, nem dönthet egyedül

-

A 4 milliós node-ok futtatják az oracle-szolgáltatásokat, de az általuk kezelt munkát nem hagyhatják jóvá saját hatáskörben. A végső döntés egy külön, 1 milliós masternode-okból választott bizottságé.

-
-
-
Minőségi kapu
-

Nem kerülhet fel akármilyen gép

-

Oracle-ok által indított hálózati, hardveres és próbafeladatos mérés előzi meg a listázást. A mérés később véletlenszerűen megismétlődik.

-
-
- -
- Fontos tervezési alapelv: a 4 milliós fedezet nagyobb szolgáltatási kötelezettséget és felelősséget jelent, nem automatikusan négyszeres szavazati vagy kifizetési súlyt. -
-
-
02
-
- - -
-
DeFCompute / RendszerképSzétválasztott felelősség
-
-
Szereplők és kapcsolatok
-

Így áll össze a teljes rendszer

-

A munka végrehajtása, a bizonyíték őrzése, az ellenőrzés és a fizetés külön szereplőkhöz kerül. Ettől lesz a rendszer hibatűrő és nehezebben manipulálható.

- -
- - - - - - - - - EGY MUNKA ÚTJA - - - V - Vevőfeladat + letét - - - G - GPU / CPU gépszolgáltató agent - - - PIACTÉR ÉS ÜTEMEZŐ - Megfelelő gép kiválasztása - ár · teljesítmény · régió · minősítés - - - - - - 4M COMPUTE / ORACLE MASTERNODE-RÉTEG - - közvetítés - mérés - bizonyíték - replikáció - - - - - 1M MASTERNODE VALIDÁTOROK - - - 5 igen a 7-ből → elfogadva - - - ELSZÁMOLÁS - Escrow fizet - - - - -
-
-
03
-
- - -
-
DeFCompute / Felhasználói folyamatA vevőtől a kifizetésig
-
-
Példa hétköznapi nyelven
-

Mi történik egy számítási feladattal?

-

A vevő nem kap hozzáférést a szolgáltató gépéhez. Egy pontosan meghatározott, elszigetelt feladat fut le, majd csak az ellenőrzött eredmény után történik elszámolás.

- -
-
-

A gép ajánlatot tesz

Az agent közzéteszi a ténylegesen kiadható erőforrást, árat és minőségi bizonyítványt.

-

A vevő kiválasztja

A piactér ár, régió, GPU, tárhely és minőségi osztály alapján párosít.

-

A díj letétbe kerül

A pénz nem közvetlenül a szolgáltatóhoz megy; az escrow a döntésig zárolja.

-

Az agent letölti a feladatot

Konténerképet és bemenetet kap, ellenőrzött hash-ekkel.

-

Elszigetelt futtatás

A feladat Dockerben, később erősebb microVM-szigetelésben fut.

-
-
-

Eredmény és bizonyíték készül

Kimeneti hash, napló, futási mérés és proof-of-resource csomag jön létre.

-

Több oracle eltárolja

Az evidencia nem egyetlen gépen marad; tartalomazonosító alapján replikálódik.

-

Hét validátor ellenőriz

A kijelölt 1M masternode-ok egymástól függetlenül szavaznak.

-

Öt aláírás dönt

Legalább öt egyező, érvényes aláírás nélkül nincs elfogadás.

-

Fizetés vagy visszatérítés

Elfogadáskor fizetés, hibánál szabály szerinti újrapróba vagy visszatérítés.

-
-
- -
- Biztonsági határ: a vevő soha nem kap SSH/root hozzáférést, a számítási agent pedig soha nem kaphatja meg a Core wallet, tulajdonosi vagy masternode BLS-kulcsait. -
- -
-
Hash
változtathatatlan azonosító
-
3–5×
bizonyíték-replika
-
5/7
elfogadási küszöb
-
-
-
04
-
- - -
-
DeFCompute / 4M szolgáltatási rétegNagyobb fedezet, több feladat
-
-
Compute / Oracle Masternode
-

Mitől ér többet a 4 milliós node?

-

Nem attól, hogy ugyanazt a blokkláncot négyszer tölti le. Attól válik hálózati szempontból értékessé, hogy kötelező, mérhető és számonkérhető többletszolgáltatásokat nyújt.

- -
-
↔

Oracle API és relay

Összeköti a szolgáltató agenteket, a piacteret és a validátorokat. A forgalom hitelesített és visszakövethető.

-
◆

Bizonyítéktár

Megőrzi és több társhoz továbbítja a futási bizonyítékokat, naplóhash-eket és eredményazonosítókat.

-
⌁

Minőségmérés

Aktív hálózati teszteket és véletlen próbafeladatokat indít a szolgáltatói gépeken.

-
⚙

Aggregálás és beküldés

Összegyűjti az öt validátoraláírást, majd bárki által ellenőrizhető csomagként továbbítja az elszámoláshoz.

-
- -
-
-
-
4 000 000
-
DFCN fedezet a tervezett új típushoz
-

A nagy fedezet gazdasági felelősséget ad, a műszaki feladat pedig valódi hasznot teremt a hálózatnak.

-
-
-

Ajánlott Core-tervezési szabály

-
    -
  • külön Compute masternode-típus;
  • -
  • külön devnet aktiválási magasság;
  • -
  • Platform/Evo mezők nélkül;
  • -
  • szavazati súly: 1;
  • -
  • fizetési súly: külön szabály, nem automatikus 4×;
  • -
  • régi kliens aktiválás előtt kötelezően frissítendő.
  • -
-
-
- -
- Miért fontos a külön típus? A meglévő Evo-típus 4M fedezete a kódban más mezőkhöz és súlyozásokhoz kötődik. Egyszerű újrahasználata akaratlanul extra governance- vagy kifizetési befolyást adhatna. -
-
-
05
-
- - -
-
DeFCompute / 1M validátorokFüggetlen 5-of-7 döntés
-
-
Ellenőrzési réteg
-

Miért hét ellenőr és miért kell öt igen?

-

Az 1 milliós masternode-ok közül feladatonként kiválasztott bizottság ellenőrzi a bizonyítékot. Így az oracle, a szolgáltató és a döntéshozó nem ugyanaz a fél.

- -
- - - RÖGZÍTETT BIZOTTSÁG EGY ADOTT MUNKÁHOZ - - ✓MN 1 - ✓MN 2 - ✓MN 3 - ✓MN 4 - ✓MN 5 - –MN 6 - –MN 7 - - - KÜSZÖB ELÉRVE5 aláírás → elfogadás - -
- -
-
-

Mit ellenőriz a validátor?

-
    -
  • a bizottsági tagság és referencia ChainLock;
  • -
  • a feladat és eredmény hash-e;
  • -
  • a bizonyíték teljessége és szabályossága;
  • -
  • a szolgáltató aláírása és időkorlátja;
  • -
  • nincs-e kettős, ellentmondó eredmény.
  • -
-
-
-

Mit kell kizárni?

-
    -
  • oracle által gyártott ál-szavazatok;
  • -
  • utólag változó bizottság;
  • -
  • szavazat csendes felülírása;
  • -
  • egy közös API-kulccsal leadott voksok;
  • -
  • production módban mock fallback.
  • -
-
-
- -
- Minden munkához rögzíteni kell a referenciamagasságot, ChainLock hash-t, hét tagot, 5-ös küszöböt és szabályverziót. A szavazatok aláírt, hozzáfűzhető események legyenek; az ellentmondó szavazat bizonyítékként megőrizhető. -
-
-
06
-
- - -
-
DeFCompute / Minőségi kapuBelépés csak mérés után
-
-
Gépfelvétel a piactérre
-

A bemondott teljesítmény helyett bizonyított képesség

-

A szolgáltatói agent önbevallása csak kiindulási adat. A listázást oracle-ok által indított, időzített és aláírt vizsgálat engedélyezi.

- -
- - - - 01FelismerésCPU · GPU · RAM - 02Oracle-kihívásnonce + időlimit - 03Terhelés10–15 perc - 04Mintafeladatvalódi eredmény - 05Bizonyítványlejáró minősítés - - - MEGFELELTlistázható + minőségi osztálykezdetben korlátozott munkakeret - NEM FELELT MEGnem bérelhetőjavítás után ismételt vizsgálat - -
- -
-
GPU

Hardver

Modell, VRAM, tartós sebesség, hőmérséklet, hibák, throttling.

-
SSD

Tárhely

Szabad hely, szekvenciális és véletlen I/O, stabilitás.

-
NET

Hálózat

Fel- és letöltés, ping, jitter, csomagvesztés több régióból.

-
- -
- Nem egyszeri vizsga: rövid újramérés 6–24 óránként, véletlen próbafeladatok és minden valódi munka után frissülő minőségi pontszám. Így a belépés után visszavett vagy túlértékesített kapacitás is észlelhető. -
-
-
07
-
- - -
-
DeFCompute / Mérési rendszerÉrthető és összehasonlítható
-
-
Ajánlott induló küszöbök
-

Mitől számít megfelelőnek egy GPU-s gép?

-

A minimumértékek nem örök szabályok. Devnetes mérésekből kell őket kalibrálni, majd géposztályonként és feladattípusonként verziózni.

- - - - - - - - - - - - - -
TerületInduló minimumMiért fontos?Hogyan mérjük?
Memória32 GB fizikai / kb. 24 GB kiadhatóA rendszernek és az agentnek is maradjon tartaléka.OS-adat + terhelés alatti mérés
GPU memória8–12 GB, munkakategóriától függőenA modellnek és a batch-nek el kell férnie.Driverlekérdezés + allokációs próba
Szabad SSD150 GB; ajánlott 300–500 GB NVMeKonténerképek, modellek, munkatér és cache.Szabad hely + kontrollált I/O teszt
Letöltés≥ 100 MbpsBemenetek és image-ek érkezése.Több oracle-régióból aktív teszt
Feltöltés≥ 50 MbpsEredmények és bizonyítékok visszaküldése.Oracle oldalon is mért adatfolyam
Csomagvesztés< 1%; prémium osztályban < 0,5%Megszakadás, hibás interaktív élmény ellen.Időzített minták több célponttal
Jitter< 30 msA késleltetés kiszámíthatósága.Rövid és hosszabb mintasorozat
GPU tartós teljesítmény≥ 80% típusreferenciaKiszűri a túlmelegedést és erős limitálást.10–15 perces változó kihívás
- -
-
-
86
-

Quality Score

Egységes, de részleteiben is megtekinthető pontszám. Nem helyettesíti a munkaspecifikus minimumot.

-
-
-
35% tartós teljesítmény
-
-
25% hálózat · 25% rendelkezésre állás · 15% sikeres munkák
-
-
- -
- Egy 32 GB RAM + RTX 3090 konfiguráció számításra erős lehet, de 80 GB SSD általános piactérhez kevés. Kis munkakategóriában engedhető, nagy modellekhez legalább 250–500 GB NVMe célszerű. -
-
-
08
-
- - -
-
DeFCompute / Szolgáltatói osztályokMinőség kizárás helyett
-
-
Otthoni vagy adatközponti?
-

Ne a címet, hanem a szolgáltatás minőségét jutalmazzuk

-

Az adatközpont részben felismerhető IP/ASN, útvonal és stabilitás alapján, de nem bizonyítható tökéletesen. Egy otthoni optikai RTX 3090 jobb lehet egy túlterhelt olcsó szervernél.

- -
-
- Community -

Közösségi gép

-

Otthoni vagy nem igazolt környezet, amely műszakilag átment a mérésen.

-
  • rövid batch munkák
  • alacsonyabb értékhatár
  • gyakoribb újramérés
  • reputáció építése
-
-
- Verified -

Ellenőrzött

-

Tartós teszteken megfelelt, stabil és jó munkatörténettel rendelkező gép.

-
  • közepes és hosszabb munkák
  • magasabb értékhatár
  • jobb rangsorolás
  • nyilvános részeredmények
-
-
- Datacenter Verified -

Adatközpont igazolt

-

Hosting ASN, stabil IP, hosszú próbaüzem és magas rendelkezésre állás.

-
  • nagy értékű munkák
  • interaktív feladatok
  • szigorúbb SLA
  • prioritásos párosítás
-
-
- -
-
-

Datacenter Verified feltételek

-
  • hosting szolgáltatóhoz tartozó ASN;
  • stabil, közvetlenül mért hálózati útvonal;
  • legalább 72 órás próbaüzem;
  • később ≥ 99,5% gördülő rendelkezésre állás;
  • prémium hálózati küszöbök.
-
-
-
Decentralizációs előny: az otthoni GPU-k földrajzilag és szolgáltatónként szétszórják a kapacitást.
-
Kockázat: VPN vagy proxy miatt a fizikai hely nem bizonyítható teljesen; a rendszer ezért a ténylegesen mért szolgáltatást tekintse elsődlegesnek.
-
-
- -
-

Feladatosztályok automatikus párosítással

- - - - - - - -
OsztályMire való?Legfontosabb követelmény
Standard BatchRender, AI batch, kódolás, hosszabb számítástartós teljesítmény és megbízhatóság
Interactivegyors válaszú fejlesztési vagy vizualizációs munkaalacsony ping, jitter és csomagvesztés
High Performancenagy modell, nagy adat, időérzékeny GPU-feladatgyors GPU, NVMe, nagy sávszélesség
-
-
09
-
- - -
-
DeFCompute / Csalás- és hibavédelemBizalom helyett bizonyíték
-
-
A mérés kijátszásának nehezítése
-

Honnan tudjuk, hogy valóban az ígért gép dolgozik?

-

A szolgáltató által küldött adat önmagában nem bizonyíték. A kihívást az oracle adja, a mérést mindkét oldal rögzíti, az eredményt pedig több független szereplő ellenőrzi.

- -
-
#

Véletlen, aláírt kihívás

Egyszer használható nonce, változó bemenet és szűk időablak akadályozza az előre elkészített válaszokat.

-
⇄

Kétoldali hálózati mérés

Nem az agent állítása számít: az oracle saját oldalán is méri az átvitelt, késleltetést és megszakítást.

-
3×

Több független oracle

Eltérő régióból, szolgáltatótól és lehetőleg ASN-ből érkező mérések csökkentik az összejátszás esélyét.

-
↻

Véletlen újratesztelés

A minősítés lejár; valódi munkák közben és között is érkezhet rövid próba.

-
- -
-
-

Automatikus riasztási jelek

-
    -
  • a teljesítmény hirtelen eltér a referenciaértéktől;
  • -
  • azonos hardverazonosító több helyen egyszerre;
  • -
  • túl sok kapacitás kerül párhuzamosan értékesítésre;
  • -
  • eredményhash vagy időbélyeg ellentmondása;
  • -
  • oracle-ok eltérő evidencia-rootot látnak;
  • -
  • gyakori megszakadás vagy feladatátadás.
  • -
-
-
-

Fokozatos következmények

-
-
Figyelmeztetés
pontszámcsökkenés, újramérés
-
Korlátozás
kisebb értékű munkák
-
Felfüggesztés
nem kap új munkát
-
Bizonyított csalás
protokoll szerinti szankció
-
-
-
- -
- A távoli mérés korlátja: hagyományos szoftverrel nem bizonyítható tökéletesen a gép fizikai helye vagy tulajdonosa. A TEE/hardveres attestáció később erősíthet ezen, de önmagában az sem bizonyít adatközponti elhelyezést. -
-
-
10
-
- - -
-
DeFCompute / InfrastruktúraAjánlott szerepkörök
-
-
Hol mi fusson?
-

Különítsük el a hálózati bizalmat és a számítási kockázatot

-

A nagy fedezetű oracle/Core node stabil Linux szerveren fusson. A Windows, macOS vagy Linux számítási gépek távoli worker-ként kapcsolódjanak hozzá.

- -
-
- 4M Oracle / Core -

Stabil dedikált Linux szerver

-
  • 8–16 modern CPU-mag;
  • 64–128 GB ECC RAM;
  • 2× NVMe, lehetőleg tükrözve;
  • 1–4 TB használható tárhely;
  • garantált 1 Gbps kapcsolat;
  • monitoring, mentés, több régió.
-

GPU általában nem szükséges az oracle, relay, indexer és bizonyítéktár feladataihoz.

-
-
- Compute Worker -

Windows · macOS · Linux

-
  • GPU/CPU kapacitás értékesítése;
  • kifelé induló, hitelesített kapcsolat;
  • elkülönített munkatér;
  • erőforráskorlátok;
  • Docker, később microVM;
  • Core-kulcsok nélkül.
-

A Linux/Docker út a legegyszerűbb kezdeti cél; a Windows GPU és macOS Metal támogatás külön fejlesztési feladat.

-
-
- -
- A legfontosabb izoláció: önkényes idegen kódot ne futtassunk ugyanazon a hoston, ahol a 4M fedezethez tartozó Core folyamat és érzékeny kulcsok vannak. A worker külön gép vagy legalább erősen izolált környezet legyen. -
- -
-

Hálózati élmény feladattípusonként

-
-
<100 ms
batch feladat

A nagy adatátvitel fontosabb, a válaszidő kevésbé érzékeny.

-
<40 ms
interaktív munka

Alacsony jitter és csomagvesztés is szükséges.

-
1/1 Gbps
kiváló kapcsolat

Szimmetrikus optika előny, de nem minden munkához kötelező.

-
-
-
11
-
- - -
-
DeFCompute / HibatűrésTöbb oracle, nem egy központ
-
-
Decentralizált oracle-szolgáltatás
-

Minél több oracle van, annál jobb — bizonyos feltételekkel

-

A több példány növeli a rendelkezésre állást és megőrzi a bizonyítékokat, de csak akkor, ha nincs mind ugyanannál a szolgáltatónál, és ugyanazt az eseménysorrendet tudják ellenőrizni.

- -
- - - WORKERBizonyítékcontent hash - - AOracle EUszolgáltató 1 - BOracle USszolgáltató 2 - COracle APszolgáltató 3 - - - VALIDÁTOR3 forrás összevetéseegyező proof/evidence root?ELLENŐRZÖTT - - ELTÉRŐ ROOT → konfliktus, nincs automatikus fizetés - -
- -
-
-

Replikált események

-

Minden állapotváltozás aláírt, tartalomazonosított esemény: jobHash, proofHash, evidenceRoot, outputHash, időbélyeg és aláíró.

-
-
-

Elhelyezési szabály

-

Egy munkához választott replikák lehetőleg külön országban, szolgáltatónál és ASN-ben fussanak.

-
-
- -
- Egy közös SQLite adatbázis jó fejlesztői MVP, de nem decentralizált igazságforrás. Productionben eseményalapú replikáció kell; a helyi adatbázis csak gyorsítótár vagy lokális nézet legyen. -
-
-
12
-
- - -
-
DeFCompute / MegvalósításDevnettől az élesítésig
-
-
Javasolt lépéssorrend
-

Előbb a bizonyítható működés, utána a konszenzusváltozás

-

A legkisebb kockázatú út az alkalmazásszintű 5/7 validáció működőképessé tétele a meglévő 1M node-okkal, majd a 4M Compute típus külön aktiválása.

- -
-
1

Mock szavazatok cseréje valódi aláírásokra

Rögzített bizottság, aláírt és append-only voksok, production mock fallback nélkül.

-
2

1M validator sidecar devnetes próba

Legalább hét valós operátorral: offline tagok, késleltetés, hibás és ellentmondó szavazat tesztje.

-
3

Több oracle és evidencia-replikáció

Több régió, aláírt eseményfolyam, eltérő root felismerése és reconciliation.

-
4

Minőségi kapu és szolgáltatói osztályok

Aktív benchmark, próbafeladat, pontszám, újramérés, korlátozás és felfüggesztés.

-
5

Elszámolási gateway

Öt egyedi, ellenőrizhető validátoraláírást fogad el; bármely oracle beküldheti.

-
6

Új 4M Compute masternode-típus

Önálló típus, devnet activation height, külön jogosultságok és kötelezettségek.

-
7

Terhelési és támadási tesztek

Oracle-kiesés, 2/7 és 3/7 hibás validátor, hálózati partíció, kettős eredmény, túlértékesítés.

-
8

Mainnet csak mérhető sikerkritériumok után

Dokumentált upgrade ablak, régi kliensek kizárása, monitoring és vészleállítási szabályok.

-
- -
- Miért nem rögtön natív új LLMQ? A Regular-only tagszűrés és a más láncon ellenőrizhető aláírás további Core-konszenzusmunkát igényel. Az alkalmazásszintű EIP-712/secp256k1 5/7 modell gyorsabban tesztelhető; natív LLMQ későbbi fejlesztési irány lehet. -
-
-
13
-
- - -
-
DeFCompute / ÜzemeltetésMitől mondjuk, hogy működik?
-
-
Mérhető sikerkritériumok
-

A hálózat egészségét is folyamatosan mérni kell

-

Nem elég, hogy egy demó egyszer lefut. A rendszer akkor érett, ha terhelés, kiesés és rosszindulatú viselkedés mellett is kiszámítható.

- -
-
p95
feladat-hozzárendelési idő

Mennyi idő alatt kap megfelelő gépet a vevő?

-
≥99%
sikeres elszámolás

Nem vitatott, szabályosan lezárt munkák aránya.

-
5/7
validátor rendelkezésre állás

Mennyi idő alatt áll össze az elfogadási küszöb?

-
3+
független oracle-régió

A bizonyíték túléli-e egy szolgáltató kiesését?

-
<1%
hamis pozitív mérés

Jó gépet tévesen ne tiltsunk ki.

-
0
kulcsszivárgási út

Workerből ne legyen elérhető masternode/wallet kulcs.

-
- -
-
-
-

Technikai telemetria

-
  • API/RPC hívások száma, p50/p95 és hibaarány;
  • oracle eseményeltérés és replikációs késés;
  • validátorszavazat késése és kiesési aránya;
  • munkánként CPU/GPU/VRAM/SSD kihasználtság;
  • hálózati ping, jitter, loss és átviteli görbe;
  • event-loop lag, memória és adatbázis p95.
-
-
-

Piaci és minőségi mutatók

-
  • aktív és sikeresen minősített gépek száma;
  • Community / Verified / Datacenter arány;
  • megszakadt, újraütemezett és vitatott munkák;
  • tényleges teljesítmény az ígérthez képest;
  • szolgáltatói koncentráció régió és ASN szerint;
  • vevői költség és szolgáltatói kihasználtság.
-
-
- -
- A publikus explorerben ne csak „online/offline” látszódjon. Jelenjen meg a minőségi osztály, utolsó mérés, rolling uptime, teljesített munkák, hálózati kategória és a bizonyítvány lejárata — érzékeny nyers IP vagy hardverazonosító nélkül. -
- -
- A pontszám algoritmusa legyen dokumentált és verziózott. A szolgáltató lássa, melyik mérés rontotta a besorolását, és kérhessen újramérést. -
-
-
14
-
- - -
-
DeFCompute / Döntési pontokAmit specifikálni kell
-
-
Nyitott kérdések
-

Mit kell eldönteni a fejlesztés előtt?

-

A fő architektúra összeállt, de néhány gazdasági és protokollszabályt még a kódolás előtt pontos, tesztelhető mondatokba kell foglalni.

- -
-
Tagság

Ki lehet validátor?

Minden 1M node kötelező sidecarral, vagy csak külön regisztrált, szolgáltatást hirdető operátorok?

-
Jutalom

Ki miből részesedik?

Oracle-díj, validátordíj, worker díja, DAO-rész és sikertelen feladat költsége.

-
Adat

Meddig őrizzük az evidenciát?

Méretkorlát, replikaszám, titkosítás, törlés és vita esetén hosszabb megőrzés.

-
Szankció

Mi jár hibáért és csalásért?

Külön kell választani a véletlen kiesést, a rossz teljesítményt és a bizonyítható manipulációt.

-
Bizottság

Hogyan választunk 7 tagot?

Determinista, munkához és ChainLockhoz kötött kiválasztás, újrajátszás elleni védelemmel.

-
Vita

Mi történik 4/3 eredménynél?

Új validátorkör, újrafuttatás, kézi DAO-felülvizsgálat vagy automatikus visszatérítés?

-
- -
-

Ajánlott alapállás az első devnethez

- - - - - - - - - -
TémaInduló döntésMiért?
Validátorkülön regisztrált 1M sidecar operátoroka nem telepített sidecar ne okozzon állandó 5/7 liveness hibát
Bizottság7 fix tag / job, 5 aláírásegyszerűen auditálható és reprodukálható
Oracle-replikaminimum 3, célérték 5egy-két kiesés és szolgáltatói hiba túlélhető
Új szolgáltató72 órás próba + alacsony munkalimitreputáció nélkül korlátozza a kockázatot
Fizetéscsak érvényes 5/7 csomagraaz oracle egyedül ne tudjon pénzt felszabadítani
- -
- A DAO feladata lehet a szabályverziók, díjak és vészleállítási paraméterek jóváhagyása — nem szükséges, hogy a DAO központi oracle-szervereket vagy signer kulcsokat üzemeltessen. -
-
-
15
-
- - -
-
-
-
DeFCon
DeFCompute
-
-
Összegzés
-

A bizalom helyét átveszi a mérés, a bizonyíték és az elosztott döntés.

-

A 4M node-ok szolgáltatják az oracle- és bizonyítékréteget. Az 1M masternode-ok 5/7 bizottsága ellenőriz. A worker csak sikeres minősítés után értékesíthet. A vevő pedig csak elfogadott munka után fizet.

-
-
-
4M
szolgáltatási felelősség
-
5/7
független elfogadás
-
3–5
oracle-replika
-
100%
mérés előtti listázási kapu
-
- -
-
- - - - diff --git a/deliverables/defcompute_koncepcio/previews/page-01.png b/deliverables/defcompute_koncepcio/previews/page-01.png deleted file mode 100644 index f37e1ef..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-01.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-03.png b/deliverables/defcompute_koncepcio/previews/page-03.png deleted file mode 100644 index 6523441..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-03.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-07-check.png b/deliverables/defcompute_koncepcio/previews/page-07-check.png deleted file mode 100644 index d37ca1f..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-07-check.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-07.png b/deliverables/defcompute_koncepcio/previews/page-07.png deleted file mode 100644 index 05b7685..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-07.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-08.png b/deliverables/defcompute_koncepcio/previews/page-08.png deleted file mode 100644 index 3a6c119..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-08.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-09-check.png b/deliverables/defcompute_koncepcio/previews/page-09-check.png deleted file mode 100644 index ef5cd37..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-09-check.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-09-final.png b/deliverables/defcompute_koncepcio/previews/page-09-final.png deleted file mode 100644 index 799a934..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-09-final.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-09.png b/deliverables/defcompute_koncepcio/previews/page-09.png deleted file mode 100644 index aa87180..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-09.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-13.png b/deliverables/defcompute_koncepcio/previews/page-13.png deleted file mode 100644 index 50595bd..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-13.png and /dev/null differ diff --git a/deliverables/defcompute_koncepcio/previews/page-16.png b/deliverables/defcompute_koncepcio/previews/page-16.png deleted file mode 100644 index 350daa7..0000000 Binary files a/deliverables/defcompute_koncepcio/previews/page-16.png and /dev/null differ diff --git a/docs/AUDIT_2026-09-05_HU.md b/docs/AUDIT_2026-09-05_HU.md deleted file mode 100644 index eea8a06..0000000 --- a/docs/AUDIT_2026-09-05_HU.md +++ /dev/null @@ -1,208 +0,0 @@ -# Audit – devnet.deftrack explorer és szimulátor (2026-09-05) - -Hatókör: a teljes repó a `feat/outcome-concentration` ágon (`3c7b498`), hat -területre bontva: gyűjtő/observation pipeline, HTTP-réteg és biztonság, -szimulátor-mag (`server/src/simulator`), szimulátor szolgáltatási réteg -(services/routes/models/CLI), Lit-kliens, valamint ops/chaos/lab/CI. Minden -fájl teljes terjedelmében olvasva; a kiemelt (kritikus/magas) találatokat a -forrásban másodszor is visszaellenőriztem, ezeket **[ellenőrizve]** jelöli. -A node-oldali feltevéseket az `~/DEFCON` `v22.1.x` (`e15e29b136`) fa ellen -olvastuk vissza. - -Kiindulási állapot: `npm run typecheck` zöld mindhárom workspace-en; `npm test` -69 fájl / 626 teszt zöld; `npm audit` 3 mérsékelt (qs az Express → body-parser -láncban, **most már** van hozzá `npm audit fix`). A VPS-en a `NODE_ENV=production` -a unitban és a futó folyamatban is be van állítva, az `ADMIN_IDENTITY_HEADER` -nincs (csak olvasó SSH-val ellenőrizve). - -Ami jól áll, és amit nem érdemes szétszedni: a tiszta domain-réteg (dkgSchedule, -dkgWindows, mnListDiff, runState, liveRunLock) alaposan tesztelt és helyes; az -admin auth-terv (kulcs XOR session, CSRF, Cloudflare Access JWT RS256 + iss/aud, -timing-safe összehasonlítás, hash-elt session-id) rendben van; a szimulátor -publikus DTO-i allowlist-alapúak és nem szivárogtatnak; az audit-stream -append-only és ujjlenyomattal visszajátszott; a chaos-csomag fájljogai, -symlink-ellenőrzései és job-írásai gondosak; secret nincs a munkafában. - ---- - -## 1. Kereszt-területi top-lista (ezekkel kezdeni) - -| # | Súly | Hol | Mi | -|---|---|---|---| -| 1 | **KRITIKUS** [ellenőrizve] | `server/src/services/sync.service.ts:333-347` | `rollbackIfReorged` a sikertelen `getblockhash`-t (null) hash-eltérésként kezeli; a visszagörgető ciklusban a null sosem töri meg a keresést, a kurzor -1-ig fut, és **minden** Block/Transaction/ServiceEpoch/QuorumCommitment törlődik, minden QuorumRound `pending`-re áll. Kiváltó: node-restart, `-28 warmup`, axios timeout (nincs a transient-regexben, `rpc.service.ts:122`), vagy **node-reindex** (a tip fölötti `getblockhash` hibát ad). A `signingActiveQuorumCount`-nál régebbi formálódott körök utána sosem figyelhetők meg újra, tehát a sikertelen-kör történet – a projekt célja – visszaállíthatatlanul elveszik. | -| 2 | **KRITIKUS** [ellenőrizve] | `ops/chaos/defcon-chaos:323-325`, `server/src/simulator/netemLease.ts:234-238` | A `prio` root az **alapértelmezett priomappal** épül, a netem az `1:3` sávra kerül. A Linux a TOS-bulk osztályú forgalmat – köztük az OpenSSH interaktív `af21`-ét – szűrő nélkül is a 3. sávba teszi, ezért 100 % loss mellett a host SSH-ja meghal (a plan.md §3a-ban élesben megtörtént), és egy közös hoston a mainnet forgalom is a netemen megy át. A „production sosem kap faultot” tulajdonságot a wrapper nem tartja. | -| 3 | **MAGAS** [ellenőrizve] | `defcon-chaos:325` | A szűrő egress `dport`-ra illeszt, ami a *távoli* portot jelenti: nem a célzott daemon linkjeit, hanem a host összes daemonjának az adott portszámú távoli peerek felé menő kapcsolatait vágja el. Egy megosztott hoston a másik 9 masternode-ot is rontja. Izolációhoz cgroup/unit-alapú osztályozó + `sport` kell. | -| 4 | **MAGAS** [ellenőrizve] | `netemRunner.ts:42-48`, `netemLease.ts:283`, `netemWrapperMain.ts:63-64` | A wrapper **nyugtázza, amit nem alkalmazott**: lejárt lease-re `planApply` üres akciólistát ad, `applyLocked` mégis jobId-t ad vissza, a ciklus `ack()`-ol és `dispatched`-ként számolja. Ugyanez a kimenet a nem validált partíciós peer-lista és az `1e-7%` formátumú loss esetén. A dispatcher már az enqueue-nál `applied`-re állítja az akciót (`simulationDispatcher.service.ts:131`). A futam „fault_active”-ot hisz, miközben semmi nem történt. | -| 5 | **MAGAS** [ellenőrizve] | `simulationDispatcher.service.ts:181-195`, `liveExecutorPlan.ts:327`, `dockerLiveExecutor.ts:163-165` | Az ütemezett `fault-clear` sorok bekerülnek a dispatcher-táblába, de `scheduledLabActionsForPlan` kihagyja őket, így a lookup `UnsupportedLiveFaultError`-t dob: **a netem-fault sosem törlődik a tervezett végén**, csak kézi `recover`-re vagy a wrapper TTL-jére. | -| 6 | **MAGAS** [ellenőrizve] | `quorumRound.service.ts:272-278, :179`, `dkgSchedule.ts:117-121`, `collectorPolicy.ts:24` | Egy `masternode count` hiba `null`-t ad, `effectiveSize` null → `classifyRound` az `impossible` ágat kihagyja → **végleges, hamis `failed`** ítélet; a `failed` státuszt a frissítési szabály sosem nézi újra. A CLAUDE.md „soha ne írjunk 0-t hibából” szabályának ítélet-változata. | -| 7 | **MAGAS** | `simulationControl.service.ts:646-656, :713-730, :741-747` | Három élő-lock hiba: (a) két egyidejű `start` esetén a vesztes CAS-ág `releaseForRun`-nal a **nyertes** lockját engedi el; (b) `abort()` a `begin_recovery`-t a végrehajtó-ellenőrzés *előtt* perzisztálja, így egy aktiválás előtt abortált live futam örökre `recovery`-ben ragad és a `no-active-experiment` preflight minden későbbi live futamot blokkol (a `create` alapból `network:'devnet'`-et fogad live módban is); (c) a lock sikertelen recovery után is elenged, miközben `faultMayBeActive: true`. | -| 8 | **MAGAS** | hat publikus route + `shared/src/index.ts:53,183,231,247,326` + öt kliensoldal | Masternode host-címek (`service`, `hostIp`, `serviceBefore/After`, `byHost[].hostIp`, `punishmentsInBlock.hosts`, fairness `hosts[].host`) a publikus DTO-kban és a publikus oldalon. A repó saját szabálya („a host-címek nem publikusak”) ellen. | -| 9 | **MAGAS** | git-történet, `fb6bb7c` → `a2f86c6` | Egy valós, nem dokumentációs tartományú IP-cím került egy tesztfixtúrába, később kicserélve, de a **publikus GitHub-történetben** benne maradt. Nem a VPS címe. Döntés kell: history-rewrite + force-push, vagy „kiszivárgottnak tekintjük és tűzfalazzuk”. (A cím szándékosan nincs ebben a fájlban.) | -| 10 | **MAGAS** | `ops/chaos/install.sh`, `plan.md` §4 | Nincs host-jogosultsági kapu: a wrapper és a 15 s-os recovery timer **három** hoston aktív, kettő véletlen maradvány, egyikükön fleet-staker és futó production `defcon-node` konténer. A `fleet-nodes.txt` csak leltár, a chaos-csomag nem olvassa. | -| 11 | **MAGAS** | `mnListDiff.service.ts:107-148` | Batch közbeni RPC-hiba után a memóriabeli penalty-baseline előrébb jár, mint a perzisztált kurzor; az újrajátszás minden büntetett node-ra hamis `penalty_up`-ot ír – pont az, amit a `seed()` megelőzni hivatott. | -| 12 | **MAGAS** | `dkgWindows.ts:1340-1353`, `measurementWindows.ts:107-110`, `simulationMeasurement.ts:338`, `scenarioRegistry.ts` (`MAX_OUTAGE_BLOCKS = 6`) | A DKG-anchor kapu és a mérési ablak együtt **sosem** tartalmazza a célzott kört: az anchor a ciklus + `dkgPhaseBlocks`-ra tesz, a megfigyelés faultStart + 2-től indul, a kört viszont `expectedHeight` = ciklus-kezdet szerint választja. Minden horgonyzott futam `not-evaluable`. Az egyetlen „matched” fixtúra 11 blokkos faulttal dolgozik, amit a registry (max 6) nem enged. | - ---- - -## 2. Területenkénti részletek - -### 2.1 Gyűjtő / observation pipeline - -Magas (az 1., 6., 11. fenti tételen túl): - -- **`effectiveSize`/`impossible` a *mostani* MN-számmal**, nem a kör alapblokkjának listájával (`quorumRound.service.ts:179`). Ban-hullám alatt (152→21) egy valóban sikertelen kör `impossible` lesz és sosem kerül újra elő; revive után az igazán lehetetlenek `failed`-re fordulnak. `protx list valid ` vagy az indexelt snapshot kell. -- **`minedHeight` sosem töltődik vissza** `detailsComplete` után (`:358-367` vs `:325`); a reorg-reset `minedHeight > cursor`-ra szűr (`reorg.ts:25`), így pont azokat a köröket hagyja formálódottnak, amelyek elhagyott blokkban voltak. -- **Két eseményíró, félig ütköző kulcsokkal** (`masternodePoller.service.ts:144-175` és `mnListDiff.service.ts:115`): a `banned/revived/registered` kulcs azonos, a poller (60 s, poll-magassággal) rendszerint előbb ír, a blokk-pontos listdiff sor no-op lesz; reorgnál csak a `source:'listdiff'` törlődik, a poll-dátumú ban megmarad. A `penalty_up` és `service_changed` viszont eltérő kulccsal **kétszer** íródik → minden lezárt kísérlet `penaltyIncreases` száma akár 2×. - -Közepes: - -- `QuorumCommitment.quorumHeight` a **bányászati** magasság, nem a kvórumé (`sync.service.ts:466`; konszenzus: `qcTx.height = pindexPrev+1`, `commitment.cpp:201`). Minden sorban `quorumHeight === minedHeight`. -- Holt ZMQ-ciklus `connected: true`-t jelent (`zmq.service.ts:82-85`), nincs újracsatlakozás; a ChainLock-időzítés csendben 5 perces felbontásra esik. -- `/chainlocks` latencia-percentilisek keverik a poll-forrású (akár 300 s) és a ZMQ-s értékeket (`chain.v1.routes.ts:308-313`), a `Block.ts:62-68` saját tiltása ellenére. -- `seedStatus.report()` RPC-hibából `peers: 0`, `verifiedMasternodes: 0`, `stakeScripts: []`-t ír (`seedStatus.service.ts:272-334`). -- Üres `protx list` (reindex, warmup) minden masternode-ra `removed` eseményt gyárt (`masternodePoller.service.ts:266-296`), IBD/mnsync-kapu nélkül. -- Batch közbeni reorg észlelhetetlen: nincs `previousblockhash` láncellenőrzés a ciklusban (`sync.service.ts:292`). -- `recomputeConsecutiveFailures` 0-ról indul egy önkényes ablakkezdetnél, és a nem rögzített (unseeable/gated) magasságokat átugorva egymás utáninak számol (`quorumRound.service.ts:418-437`). - -Alacsony: `valueSat` ~1,1e17 sat fölött nem safe integer (a `JSON.parse` már kerekít, `sync.service.ts:50-56`); hibás ZMQ-frame `lastSeq = -1`-re állít; `computeOutcome.chainLockCoverage` nevezője PoS-blokk, számlálója minden lockolt blokk; nincs `{hasChainLock, height}` index; `admin.v1.routes.ts:77` korlátlan `members.0 $exists` lekérdezés. - -Terv-szintű megjegyzés: két igazságforrás van a körökre – a `QuorumRound` egy 15 s-os RPC-ablakból, a `QuorumCommitment` blokk-pontosan és reorg-biztosan a láncból – és semmi nem egyezteti őket. A commitment-index már ismeri az egész lánc minden formálódott körét (és a null-commitment = sikertelen kör), reorg-biztosan. - -### 2.2 HTTP-réteg és biztonság - -- **Fejléc-identitásos bejelentkezés csak `NODE_ENV`-től függ** (`middleware/adminSession.ts:43`), és a „megbízható proxy” ellenőrzés nginx mögött üres (`adminSession.v1.routes.ts:55`, mindig `127.0.0.1`). Élesben zárva (ellenőrizve), de egy explicit `ADMIN_ALLOW_HEADER_IDENTITY` kapcsoló és nginx-oldali fejléc-törlés kell, hogy ne egy környezeti változó hiánya legyen a határ. -- `/admin/operators` és `/admin/experiments` csak `requireAdminApiKey`-t használ (`admin.v1.routes.ts:18`), nem `requireAdminAuth`-ot: elveszik a cookie/Origin-elutasítás, és a böngészős session nem tudja kezelni őket. -- **Nincs Express error-middleware**: hibás JSON, >256 kB body, `next(error)` HTML-oldalt ad, fejlesztői módban stack-tel; a `{ success, data }` boríték törik. -- `qs` sérülékenység a publikus API-n (`npm audit fix` + `app.set('query parser','simple')`). -- `offset` korlátlan minden lapozott route-on; `/blocks` és `/txs` `total`-ja `estimatedDocumentCount()`; `/experiments/:runKey` futó futamnál kérésenként újraszámol a teljes ablakon 10 s cache-sel; `sharesBlock` lap-lokális. -- **A CLAUDE.md „withCachePolicy profilok + in-flight dedup” állítása nem igaz**: a middleware csak fejléceket állít, dedup csak az RPC-rétegben van. -- `Cache-Control: public` 4xx-re is; `/metrics` és `/health` publikus (RSS, event-loop lag, RPC-metódusonkénti latencia); ingest `host` regex enged IPv4-et, `stakeScripts` korlát nélkül nő. -- Leállítás: `server.close()` nincs megvárva, dispatcher és observation sweep nem áll le, nincs timeout; nincs `unhandledRejection` handler; csak `MONGODB_URI` kötelező a configban; `db.ts:50-52` nyers driver-üzenetet logol (URI-töredék). - -### 2.3 Szimulátor-mag - -A 4., 5., 12. tételen túl: - -- `chain-synced` preflight a live tipet a draft `capturedAtHeight`-jével **egyenlőségre** ellenőrzi (`preflight.ts:213-216`, `simulationEvidence.service.ts:251-255`): validate és arm egy blokkon belül kell hogy lezáruljon (devneten 150 s). A 5 perces snapshot-kor már korlátoz; ez egy második, ellentmondó korlát – és a PREFLIGHT_HU így írja elő, tehát a terv hibás. -- `quorum-member-outage` `phase:'dkg'` a **régi** (utolsó formálódott) kvórumot célozza, a DKG viszont újat választ; ~40 % átfedésnél az `estimateImpact`/`expectedAvailability` szisztematikusan téves. A „next quorum nem elérhető” handoff-nyitott kérdés pont ez. -- `isBenignFailure` bármely `RTNETLINK answers`-t „már tiszta”-ként kezel egy `clear`-en, az `Operation not permitted`-et is: az élő qdisc kikerül a nyilvántartásból (`netemWrapperHost.ts:50-55`). -- Q60-küszöbök beégetve (44/41 a `dryRunExecutor.ts:333-347`-ben, literál típusok a `scenarioTypes.ts:79-80`-ban, 0,8/-0,1 a `compareAvailability`-ben): a laborban (3/2/2) minden riport `degraded` vagy `not-evaluable`. -- A futam-szintű lease minden akció-TTL-t felülír; `restart-flapping` 0. ciklusa 900 s-ig marad lent, ha a dispatcher meghal. -- Nincs automatikus `observing → recovery` átmenet a tervezett végén; a reconcile csak `abortRequested: true`-val tüzel, így egy kézzel nem recoverelt futam `aborted`-ként záródik, és a mérési vég-anchor az operátor reakcióidejétől függ. -- `preflight.ts:330` `(snapshotMatches ?? true)` fail-open; minden draft az aktuális kvórum összes tagjának regisztrációját követeli, `mn-stop`-ra is – részleges devnet-regiszterrel semmi nem draftolható. -- Wrapper boot-verseny (`netemWrapperMain.ts:433-455`: timer élesítés a `recoverInflight`/`bootCleanup` előtt); olvashatatlan state-fájl `{jobs: []}`-re cserélése; `NETEM_WRAPPER_INTERVAL_MS` NaN → 1 ms-os pörgés. -- Blast-radius: `host-outage` a `seed`-et és `staker`-t is viszi; `network-degradation` engedi a `seed` szerepet (az explorer bizonyítékforrása); flapping 10 stakert enged `maxStakers: 5` mellett; `node-isolation` peer-listája korlátlan a wrapper 32-es plafonja ellen. -- `interfaceRef: 'devnet-p2p'` dekoratív: a wrapper `eth0`-t és a teljes interfészt rontja, az ARCHITECTURE_HU „csak a P2P port” ígérete ellenére. -- `clear-recover` élőben nem futtatható (nincs lease-elt akció → `faultLeaseExpiresAtForStart` dob). Halott scenario. -- Egy régi/ismeretlen `parameters` dokumentum a teljes publikus `/simulations` listát 500-ra viszi (`simulationPublicDto.ts:351-356`). - -### 2.4 Szimulátor szolgáltatási réteg - -A 7. tételen túl: - -- **Finalize-jelölt kiéhezés**: a `find(...).sort(updatedAtMs).limit(50)` után szűr a már riportolt futamokra; 50 riportolt futam után az újak sosem kerülnek sorra (`simulationMongo.repository.ts:104-124`), és index nélküli collection scan 20 s-enként. -- **Dupla dispatch**: 30 s lease, `renewLease` implementálva, de sosem hívva; a `settle` eredménye eldobva; a parancsfájlnév nem `actionId`-alapú, tehát nincs dedup. Egy lassú docker-probe után a második worker is enqueue-ol, és a régi `stop` a frissebb `start` után landolhat. -- `failed` akciók sosem próbálódnak újra; a `maxAttempts` csak crash-reclaimre számít – a kód kommentje mást ígér. -- Recovery/abort események a request `acceptedAtMs`-ével vannak bélyegezve; egy rendszer-átmenet (`system_resume_recovery`) után ugyanazon idempotency-kulccsal ismételt kérés örökre `STALE_EVENT`. -- Sima `Error`-ok az időzítésből 500 „internal error”-ként jönnek vissza (határidő-lecsúszás 409 kellene, üzenettel). -- `PUT /targets/:targetId` bármely szerepnek engedi egy **engedélyezett** target `hostRef/unitRef/network` átírását, az `enabled` marad. -- Idempotency-kulcs nem globális (ugyanaz a kulcs `validate`/`arm`/`start`-ra elfogadva), a hash indexeletlen. -- `pre('validate')` invariánsok négy modellen sosem futnak (minden írás `updateOne`/`bulkWrite`); a modell-tesztek holt kódot tesztelnek. -- A dispatcher és az observation sweep az explorer-folyamatban is fut (`index.ts`), ahol live futam sosem aktiválódhat; leállításkor egyik sem áll le. -- Duplikáció: `isDuplicateKey` ×3, append-only hook-blokk ×4, három privát `Logger`, `abort()`/`recover()` azonos 12 sora, `chainTip` closure ×3. - -### 2.5 Kliens (Lit 3) - -- **Az áttekintő minden profilt egy sorozatba kever** (`dd-page-overview.ts:373` `healthTimeline` `llmqName` nélkül; a szerver csak paraméterrel szűr, `quorumRounds.v1.routes.ts:205`): a „formation rate (7d)”, medián/legrosszabb health és a streak öt interleaved ütemezésen fut – pont amit a CLAUDE.md tilt. A `minSize` az első (tie-break nélkül rendezett) körből jön (`:402`), így hol 3, hol 350 („152/350 minimum – kvórum nem lehetséges”), sosem a Q60 73 %-a. -- **Sikertelen vs. büntető kör vizuálisan fordítva**: a `failed` piros pill „0 büntetett” felirattal, a formálódott és 12-t büntető zöld pill sima számmal (`dd-page-rounds.ts:269-291`). Az incidens zöld, a nem-esemény piros. Három explicit állapot kell: `formed · clean`, `formed · punished N` (borostyán, N félkövér), `did not form · nobody punished` (semleges, szöveggel). -- Formation rate health nélkül a shell telemetriájában (`dd-shell.ts:307`) és a `/rounds`, `/pose` oldalon; `maxPossibleBan` a ChainLock-profilé, de nincs kimondva; profil-paraméterek csak a kísérlet-részletben. -- Host-IP-k a publikus oldalon (`dd-page-masternodes.ts:199,144`, `pose:202,245`, `fairness:192,249`, `blocks:269`). -- `.err`/`.note` osztály 12 oldalon definiálatlan; „Seen” oszlop `detectedAt`-ot mutat, amit minden poll felülír (mindig „Ns ago”); fetch-versenyek sequence-guard és AbortController nélkül minden oldalon; dupla fetch mountkor; admin-dashboard 30 s-ig üres bejelentkezés után (`_loading` kapu); scenario `