ADR-0006: 100 ms confirmations

Product wants "100 ms settlement." Physics: nothing lands on Solana inside one ~400 ms slot ⟨VERIFY⟩, and withdrawable finality is bounded below by L1 rooting…

Status: accepted (Written). Dated 2026-11-19.

Status: accepted, 2026-11-19, by the owner ("accept ADR-0006 and ADR-0063"), with the answers to its three Open questions at the end of this record. Written 2026-08-15 · Prompt: §6 C2, App. E2 · Builds on: solieum-receipt (Step 7), ADR-0003/0004. Accepted as a design: nothing in it is built.

Context

Product wants "100 ms settlement." Physics: nothing lands on Solana inside one ~400 ms slot ⟨VERIFY⟩, and withdrawable finality is bounded below by L1 rooting (~13 s ⟨VERIFY⟩) plus the proof system. So the 100 ms rung must be engineered as user-irreversibility (S1) — an enforceable, insured confirmation — while S2 (anchored) and S3 (withdrawable) are compressed separately and never conflated with it.

Decision

  1. Micro-block sealing. The sequencer seals L2 micro-blocks on a fixed 50 ms cadence. A receipt refers to a sealed block hash and position, not a mempool ack.
  2. Receipts stay the base layer (already built): signed {order_hash, market/lane, block, position, l2_slot}; double-assignment or a published batch contradicting a held receipt is provable equivocation → slash (solieum-receipt's two proofs).
  3. Coverage on top. A receipt MAY carry covered_up_to: Lamports, backed by a dedicated on-chain coverage bond (separate from the proposer bond; hard collateral per ADR-0003).
    • Exposure accounting: the sequencer tracks rolling exposure = Σ covered value on receipts whose blocks are not yet S2-anchored at confirmed ⟨policy: confirmed vs finalized — see Open⟩. Admission is fail-closed: when headroom is exhausted, further receipts are issued uncovered and say so (covered_up_to: 0); never silently covered.
    • Sizing inequality (the whole scheme): coverage_bond ≥ peak_TPS × avg_covered_value_per_tx × S2_latency_p99 × safety_factor. Faster anchoring (smaller S2 window) linearly reduces the capital the promise needs — the pipeline work below is what makes 100 ms coverage affordable.
    • Claims: holder submits receipt + equivocation/exclusion proof to the L1 coverage program → automatic payout up to covered_up_to, no discretion, pause cannot block a proven claim (exits-page rule extends here).
  4. S2 pipeline target: post one batch frame per L1 slot (continuous streaming), multi-path submission, priority-fee escalator with a published congestion playbook. Targets: S2 p50 ≤ 1.5 s, p99 published, never advertised as fixed.
  5. Language rule (site + docs): the shipped phrase is "100 ms accountable confirmation" or "insured up to X" — the bare phrase "100 ms settlement" is banned exactly as "200 ms finality" was (§4.2-class vocabulary rule).
  6. Geography is disclosed: 100 ms is a same-region ingress promise; cross-continent RTT alone can exceed it. Publish per-region p50/p99 for ack→sealed; no global single number.

Consequences

  • New on-chain program: coverage vault + claim instruction (needs ed25519 receipt verification on-chain — precompile cost ⟨VERIFY⟩ gates the claim-tx design).
  • solieum-receipt grows coverage fields + exposure-accounting + claim-verdict functions (pure, std-only, testable like everything else).
  • solieum-batch's sustainable_tps/cost model couples in: coverage headroom is a function of anchor cadence the batcher actually sustains.
  • Verifier (Step 8) gains the receipt-vs-batch cross-check and an auto-claim generator — watchtowers get paid twice: challenges and coverage claims.

Reversal triggers

Coverage bond underwater vs exposure at any TVL review → cut coverage caps automatically (fail closed). Multi-prover live and S3 in minutes → revisit whether coverage is still worth its capital at all.

Open (decide before implementation)

  • Exposure release at L1 confirmed vs finalized: confirmed frees ~10× capital, accepts reorg-window risk on coverage accounting ⟨quantify with L1 reorg data⟩.
  • Per-lane vs global coverage pools (markets appendix wants per-market lanes).
  • Who may top up the bond mid-epoch, and the timelock on withdrawals from it.

Answers to Open (accepted 2026-11-19)

Drafted at the owner's request so this record could be accepted as a design. Nothing below is built: there is no coverage bond, no claim program and no covered_up_to field anywhere in the code. The owner accepted these answers, and the record with them, on 2026-11-19.

Where things stand, measured

This record was written against targets. Three facts have since been measured, and they change the arithmetic of the first answer:

This record assumedMeasuredSource
Block cadence50 ms micro-blocks100 ms blocks on gamma and the coin chainnotes/speed-2026-10-26-100ms-validator.md
Receipt ("processed")—14–16 ms p50same, and notes/settlement-speed-2026-11-13.md
Ordering → root on Solana (the S2 window coverage must span)p50 ≤ 1.5 s target192 s median, 278 s worst under a 200-transaction burst; roots drain at about one per secondnotes/settlement-speed-2026-11-13.md §1

The sequencer bond (ADR-0012) is built and slashable on devnet; it is separate from the coverage bond this record proposes, and stays so.

A1. Exposure is released at L1 finalized, not confirmed

Decision: a covered receipt counts against the coverage bond until the batch that carries it is anchored on Solana at finalized commitment.

Why: the Open item said confirmed frees about 10× the capital. That was true against a 1.5 s anchoring target, where Solana's confirmed-to-finalized gap (about 13 s ⟨VERIFY⟩) would dominate the window. Against the measured window it does not: 13 s on top of a 192 s median is about 7% more capital, not 10×. At that price, the reorg risk confirmed carries in the coverage accounting is not worth taking, and finalized makes the accounting depend on nothing that can be rolled back.

Revisit when: the measured S2 p99 falls below about 30 s. Below that the gap is a large share of the window again, and the trade this item described becomes real. Revisiting needs L1 reorg data for confirmed, which this record still does not have.

A2. One coverage pool, with an exposure cap per lane

Decision: one coverage bond per chain. Each lane (a market, once markets exist; until then the chain is one lane) has its own exposure cap, a share of the pool. When a lane reaches its cap, its receipts are issued uncovered and say so (covered_up_to: 0), while other lanes keep their headroom. The fail-closed rule in Decision 3 applies per lane as well as to the pool.

Why: separate pools per lane split capital that is scarce at the start and need a funding decision for every new market; one pool with no caps lets a single busy market consume all the headroom and leave every other user uncovered. Caps give the markets appendix the isolation it asks for without a pool per market. No market or lane exists today (crates/solieum-clob is a library wired to nothing), so this costs nothing to adopt now.

Revisit when: a lane has a different backer (a market maker who wants to fund coverage for its own market only). That is the case for a separate pool, and it would be added beside the shared one, not instead of it.

A3. Anyone may top up; withdrawals wait, and stop new coverage first

Decision:

  • Top-up is permissionless and immediate, at any time, mid-epoch included, in SOL. More collateral can only make the promise safer, and ADR-0012's bond already works this way (solieum-node bond --top-up).
  • Only the bond's owner may withdraw, in two steps: request, then withdraw after a delay. The request takes effect at once on the sequencer side: from that moment it issues no new covered receipts against the bond, so exposure can only fall while the delay runs.
  • The delay is at least the sequencer bond's (unbond_delay_secs, 604,800 s on devnet, ADR-0012), and never shorter than the longest time a covered receipt can still be claimed: the measured S2 worst case plus the challenge window (48 h today, ADR-0016) plus a claim margin. With the current figures the sequencer bond's seven days covers that with room.
  • A proven claim is paid even during the delay, and a pause cannot block it, as Decision 3 already says. A withdrawal takes only what is left after every claim filed before it.
  • The bond is SOL, like every bond that protects users (ADR-0018); its upgrade authority follows ADR-0031's multisig and timelock like the other programs.

Why: a coverage promise is only as good as the money still standing behind it when a claim arrives. Instant withdrawal would let a sequencer cover receipts, misbehave, and leave with the collateral before a claim lands; stopping new coverage at the request closes that door without trapping honest capital for longer than the claims need.

What accepting these answers does and does not mean

Accepting them makes this record a complete design. It does not build anything, and it does not change what may be said publicly: Decision 5's language rule stands, and no page may say "insured" or "covered" until a coverage bond exists on a chain and a claim has been paid in a test.