diff --git a/THREAT_MODEL.md b/THREAT_MODEL.md
new file mode 100644
index 000000000..793324e8b
--- /dev/null
+++ b/THREAT_MODEL.md
@@ -0,0 +1,198 @@
+# Percolator Threat Model
+
+**Status:** Experimental — NOT AUDITED, NOT PRODUCTION READY  
+**Last updated:** 2026-04-16  
+**Spec version:** v12.17.0  
+
+---
+
+## Overview
+
+Percolator is a permissionless perpetual futures risk engine on Solana. This document describes the trust model, known deferred findings, deployment checklist, and security audit status for external auditors and security researchers.
+
+---
+
+## Trust Assumptions
+
+### Admin
+
+The admin key `7JVQvr...` is an EOA key that currently controls program upgrades and market configuration. It MUST be migrated to a Squads multisig before the mainnet production launch.
+
+The admin key is authorized to:
+
+- Update market config knobs (`UpdateConfig`, `SetMaintenanceFee`)
+- Resolve markets (`ResolveMarket`)
+- Set and rotate oracle authority (`SetOracleAuthority`, `SetOraclePriceCap`)
+- Burn admin permanently by setting to all-zeros (`UpdateAdmin` to `[0;32]`)
+- Force-close abandoned accounts after resolution (`AdminForceCloseAccount`)
+- Pause or unpause the market (via admin crank path)
+- Withdraw insurance post-resolution (`WithdrawInsuranceLimited`)
+- Set insurance withdrawal policy (`SetInsuranceWithdrawPolicy`)
+- Rotate the admin key itself (`UpdateAdmin`)
+
+The admin key cannot: redirect user fund withdrawals to arbitrary accounts, close the slab while funds remain, mutate config after resolution, or perform admin operations after the key is burned to all-zeros. Each of these constraints is covered by dedicated test cases. See the Admin Key Threat Model section of the percolator-prog README for the full enumeration.
+
+The planned mitigation is Squads multisig governance, which requires M-of-N approval for any admin operation. This eliminates the single-EOA risk surface.
+
+### Oracle
+
+Pyth price feeds are used for non-Hyperp markets. A staleness filter (`max_crank_staleness_slots`) bounds the maximum age of any accepted price. A confidence filter (`conf_filter_bps`) rejects prices with excessive uncertainty.
+
+For Hyperp markets, prices are authority-pushed via `PushOraclePrice`. The oracle authority is set by admin via `SetOracleAuthority`. A per-slot circuit breaker (`SetOraclePriceCap`) caps price movement, preventing instant mark manipulation.
+
+Oracle manipulation resistance is reinforced at the engine level: unrealized profit accrued from a price spike sits in the per-account reserve `R_i` and does not enter the matured haircut denominator until the warmup window passes. An attacker who spikes a price cannot immediately withdraw the accrued gain.
+
+### Keeper
+
+The keeper crank (`KeeperCrank`) is permissionless — any party may submit it. However, a funded keeper is required for liveness guarantees:
+
+- **Funding drift:** If the keeper is offline for more than `MAX_FUNDING_DT` (approximately 7 hours at typical slot rates), the funding accumulation caps out. Positions are not overfunded, but the cap means accumulated funding is undercharged during the outage window.
+- **Market freeze:** If the keeper is offline indefinitely, no liquidations execute, no funding settles, and no accounts are garbage-collected. Markets remain frozen but all funds remain in the vault — no funds are at risk due to keeper absence, only liveness.
+
+The keeper must remain funded in SOL for transaction fees. Keeper balance below minimum fee threshold is a monitoring alert. The deployment uses Railway with auto-restart on failure.
+
+### Matcher
+
+The matcher program is registered by each LP via `InitLP`. During `TradeCpi`, Percolator performs a CPI to the LP's chosen matcher program. The binding is enforced as follows:
+
+- **LP identity binding:** The matcher program address and context account must equal exactly what the LP registered at `InitLP` time. Substitution is rejected.
+- **LP PDA signature:** The LP PDA is derived on-chain from seeds `["lp", slab_pubkey, lp_idx_le]`. The user cannot supply a counterfeit PDA — it is derived, not accepted as input.
+- **Return data echo-validation:** The matcher's response must echo the oracle price, nonce (`req_id`), LP account ID, and size constraints back to Percolator. Any mismatch causes hard rejection. Reserved fields must be zero.
+- **Execution size discipline:** Percolator always uses the matcher's `exec_size`, never the user's requested size. The matcher may partially fill or reject.
+
+The matcher is treated as adversarial input. All ABI fields from the matcher response are validated before any state changes occur. The nonce (`req_id`) is a monotonic `u64` derived from the slab nonce. It increments on every accepted trade and is unchanged on reject, preventing replay.
+
+---
+
+## Known Deferred Findings
+
+The following findings were identified during internal audit and deferred. Each has a documented rationale and is tracked for resolution before full mainnet production.
+
+### F-6 — LP Identity Inconsistency ✅ RESOLVED (2026-04-17)
+
+**Description:** `TradeNoCpi` uses a generation-table lookup for LP identity, while `TradeCpi` uses an FNV hash of the LP public key. These are two different identity-binding mechanisms for the same logical concept.
+**Impact:** No current fund or position risk. The `NoOpMatcher` used by `TradeNoCpi` ignores the identity field. There is no path where this inconsistency allows an LP to impersonate another LP or bypass controls.
+**Resolution:** Pre-audit hygiene review confirmed the two schemes intentionally serve different contexts — `gen_table` (mat_counter) is strictly monotonic and used for internal lifecycle tracking, while FNV hash is used for matcher CPI binding and does not need to distinguish instance generations. Documented inline at `percolator-prog/src/percolator.rs` near the FNV compute site. Unifying would either break matcher binaries or lose slot-reuse detection. Not a bug; documented as intentional.
+
+### C-7 — NFT Account ID Guard Inoperative on v12.17
+
+**Description:** The NFT program's `account_id` guard does not execute correctly under the v12.17 slab layout.
+**Impact:** The guard failure allows a PDA to be created with an incorrect account ID, costing approximately 0.002 SOL in PDA rent. No position ownership bypass, no fund risk.
+**Status:** ACCEPTED RISK. Bounded blast radius (~0.002 SOL rent per occurrence), no fund or position access. Fix requires a layout-aware `account_id` read, scheduled for the next NFT program upgrade. Monitoring: count of NFT PDA creations per slab (alert if rate exceeds expected new-user rate).
+
+### C-12 — NFT Transfer Hook Stale Oracle
+
+**Description:** The NFT transfer hook does not re-fetch the oracle price at transfer time. It uses the price cached at last crank.
+**Impact:** In markets where the keeper cranks infrequently, the health check performed at NFT transfer time may use a stale price. This could allow a transfer when a freshly-priced health check would block it.
+**Status:** ACCEPTED RISK. Bounded by keeper crank frequency (`max_crank_staleness_slots`, configured to 200 slots). The fix requires an `ExtraAccountMeta` change (adding the oracle feed to the transfer hook accounts), which is a non-trivial Token-2022 interface change affecting wallet UX. Scheduled for next NFT program upgrade. Monitoring: staleness gap between last crank slot and NFT transfer signature slot.
+
+### SP-2 — `_reserved[8..16]` Market Start Slot Dead Write ✅ FIXED (2026-04-17)
+
+**Description:** The `_reserved[8..16]` field in the market header received a write of `market_start_slot` that was never subsequently read by any instruction.
+**Impact upgraded during review:** This was NOT purely cosmetic. The bytes `_reserved[8..16]` are also used by `mat_counter` (PERC-623 per-instance LP identity counter). The `market_start_slot` write at InitMarket was corrupting `mat_counter` with the slot value, causing the first `InitUser/InitLP` to receive `mat_counter = (creation_slot + 1)` instead of `1`.
+**Resolution:** Removed `write_market_start_slot` function and its call site at InitMarket. The slot value was never consumed as "market_start_slot" anywhere, so removal is safe. `mat_counter` now correctly starts at 0 and increments from 1 on the first account creation.
+
+### SP-5 — Admin Can Disable HWM Without Timelock
+
+**Description:** The admin can disable the high-watermark (HWM) policy for insurance withdrawals without any timelock or delay. An operator could in principle reduce withdrawal protections immediately.
+**Impact:** Governance policy risk, not a code vulnerability. The HWM mechanism provides withdrawal rate-limiting above and beyond the cooldown.
+**Status:** GOVERNANCE POLICY ITEM. There is no on-chain setter for `hwm_floor_bps` in current code; the field is set once at `CreateLpVault` and cannot be changed. The finding becomes relevant only if a setter is added in the future. Mitigation: with Squads multisig as admin, any future HWM-disable setter would require M-of-N signers. Tracked for post-launch governance hardening. No code change needed in current release.
+
+---
+
+## Deployment Checklist
+
+The following items must be verified before removing the mainnet beta designation.
+
+### Admin Key Migration
+
+Verify the admin key has been transferred to the Squads multisig:
+
+```bash
+solana program show ESa89R5Es3rJ5mnwGybVRG1GrNt9etP11Z5V2QWD4edv
+```
+
+The upgrade authority in the output must be the Squads multisig address, not an EOA key.
+
+### Keeper Liveness
+
+- Verify the keeper is funded and cranking on Railway
+- Check last successful crank slot is within `max_crank_staleness_slots`
+- Verify keeper SOL balance is above minimum fee threshold
+- Confirm Railway auto-restart is active
+
+### Oracle Authority
+
+- Verify oracle authority is set to the keeper bot's public key via `SetOracleAuthority`
+- For Pyth markets: verify Pyth feed IDs are correct and staleness bounds are appropriate
+- Verify `SetOraclePriceCap` is configured for Hyperp markets
+
+### Insurance Fund
+
+- Verify the insurance fund has been seeded with initial capital via `TopUpInsurance`
+- Verify insurance floor is set appropriately via `SetRiskThreshold`
+- Verify insurance withdrawal policy is configured with conservative cooldown
+
+### Stake Pool
+
+- Verify the stake pool `percolator_program` field matches the mainnet program ID `ESa89R5Es3rJ5mnwGybVRG1GrNt9etP11Z5V2QWD4edv`
+- Verify `admin_transferred` is 1 (pool PDA is admin, not EOA)
+- Verify `TransferAdmin` was called before accepting deposits
+
+---
+
+## Program IDs
+
+| Program | Mainnet Address |
+|---------|-----------------|
+| Core (percolator-prog) | `ESa89R5Es3rJ5mnwGybVRG1GrNt9etP11Z5V2QWD4edv` |
+| Matcher (percolator-match) | `DHP6DtwXP1yJsz8YzfoeigRFPB979gzmumkmCxDLSkUX` |
+| Stake (percolator-stake) | `DC5fovFQD5SZYsetwvEqd4Wi4PFY1Yfnc669VMe6oa7F` |
+| NFT (percolator-nft) | `FqhKJT9gtScjrmfUuRMjeg7cXNpif1fqsy5Jh65tJmTS` |
+
+Devnet program IDs differ. See the SDK README for the full address table across networks.
+
+---
+
+## Security Audit Status
+
+### Internal Audit
+
+- **Total findings:** 25
+- **Fixed:** 12
+- **Deferred:** 5 (documented above)
+- **Already fixed at time of audit:** 1
+- **Remaining open:** 7 (tracked in internal issue tracker, none are critical-severity)
+
+### Test Coverage
+
+- **Total tests:** 1,265
+- **Failures:** 0
+- **Composition:**
+  - percolator-prog: 707 tests (462 integration on LiteSVM, 28 unit, 8 alignment, 1 CU benchmark)
+  - percolator (core): remainder of 1,265
+  - percolator-nft: 65 tests
+  - percolator-match: 34 tests
+  - percolator-stake: 270 tests
+
+### Formal Verification
+
+- **Kani proof harnesses:** 471 total
+  - percolator (core): ~386 harnesses
+  - percolator-prog: 113 harnesses
+  - percolator-stake: 85 harnesses (35 harnesses in kani-proofs/)
+  - percolator-nft: additional harnesses
+- **All proofs passing:** yes
+
+### Upstream Contributions
+
+6 pull requests have been submitted to the upstream `aeyakovenko` repositories as part of the audit process.
+
+---
+
+## References
+
+- [Risk Engine Spec v12.17.0](spec.md) — normative specification
+- [percolator-prog README](../percolator-prog/README.md) — Admin Key Threat Model, instruction reference
+- [percolator-stake docs/AUDIT.md](../percolator-stake/docs/AUDIT.md) — 4-round stake program audit report
+- [percolator-stake docs/KANI-DEEP-ANALYSIS.md](../percolator-stake/docs/KANI-DEEP-ANALYSIS.md) — Proof-by-proof Kani analysis