> ## Documentation Index
> Fetch the complete documentation index at: https://docs.usetissue.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Feedback & Roadmap

> What we learned building against TxLINE, and what's next.

## Feedback for the TxLINE team

What worked especially well:

* One normalized StablePrice schema makes strategy inputs simple and deterministic
* Message IDs and timestamps provide strong replay/dedupe keys
* Merkle proof endpoints and Solana validation make source evidence independently
  checkable
* Dual SSE streams support direct event-to-market latency analysis

Friction encountered (see the project's `feedback.md` for full detail):

* Activation response shape varies (plain-text vs JSON) and required handling both
* Snapshot `asOf` semantics needed clarification during corpus capture
* The odds proof endpoint was, at the time, undocumented
* A real process-level restart drill surfaced a devnet/mainnet IDL mismatch and public
  devnet RPC rate limiting under load
* Reconfirmed on a later rebuild: TxLINE's proof endpoints don't reliably serve proofs
  for historical/replayed messages, which blocks a live SIGKILL-during-anchor-submission
  drill specifically — worked around with an independent stream-drop/reconnect drill
  that doesn't depend on any message passing proof verification (see below)

## Known limitations and future improvements

* Capture more real TxLINE fixtures to move the calibration/holdout comparison out of
  "underpowered" territory
* Extend on-chain anchoring evidence once TxLINE exposes any additional oracle
  primitives beyond `validate_odds`/`validate_stat`
* Confirm with the TxLINE team whether `/api/odds/validation` and
  `/api/scores/stat-validation` are meant to serve proofs for historical/replayed
  messages, or are intentionally scoped to recent/live data only — the one remaining
  blocker on a full live anchor-submission chaos drill

## Changelog

<Accordion title="Unreleased">
  * Ed25519 signing of every ledger record
  * Real Merkle tree + inclusion-proof API (`/ledger/proof`)
  * Periodic on-chain checkpoint anchoring, extending the Pre-Match Commitment
  * Aggregate proof-failure-rate circuit breaker
  * Signed policy config snapshots
  * Human-reviewed tuning-suggestion tool (`evaluate:tuning`)
  * N-way regime ablation matrix in the Strategy Arena
  * Decision-loop and proof-verification latency histograms on `/metrics`
  * Equity curve on the dashboard
  * Structured pattern-recall tool (`find_similar_decisions`) in the analyst layer
  * Local Solana anchoring test suite against Surfpool (`test:surfpool`) — real
    transaction-level scenarios without racing public devnet's rate limits
  * Adversarial input suites for both the feed ingest pipeline and the analyst's
    LLM+MCP tool loop — found and fixed two real bugs: a malformed odds price could
    silently poison an entire market's consensus with NaN, and the analyst's tool-loop
    fallback returned a blank answer instead of an explanatory message
  * Dashboard end-to-end suite (Playwright) covering every page under every desk status
  * A second real process-level chaos drill (`drill:streamdrop`) exercising SSE
    disconnect/reconnect resilience independent of TxLINE proof-verification success
  * Real order execution on Slip (`exec/slipExec.ts`): a risk-approved decision that
    clears a second, stricter, off-by-default capital-risk gate
    (`risk/gates.ts::evaluateSlipExecution`) is turned into a real signed, confirmed
    transaction on a real settlement venue — TxLINE's own `txoracle` has no order/execution
    instructions at all, so this closes what was previously the honest limitation above.
    Found and fixed a real bug along the way: the vendored `@slip/sdk` tarball had been
    packed before a required field (`settlement_mode`) was added to the on-chain program's
    instruction shape, so every transaction it built was silently corrupted — diagnosed by
    isolating the failure to a byte-identical call that passed in Slip's own repo but
    failed in Tissue's, then diffing the committed source against the stale packed build.
    Rehearsed end to end (independently provision a two-sided market, buy, resolve from a
    protocol-valid score proof, claim, and retry reconciliation, with independent on-chain
    balance checks) against a local Surfpool instance
    (`test:slip:surfpool`)
</Accordion>

## Contributing

This repository is currently private under a competition embargo. Once publication is
authorized:

* Follow the existing pattern discipline: every tunable constant belongs in
  `policy.toml`, never hardcoded
* New heuristics must be flagged, individually toggleable, and testable in isolation
  via the regime ablation matrix
* No feature may compromise the analyst layer's read-only isolation or the
  deterministic replay guarantee
* Every PR should include real test coverage — this project treats "tests exist" as a
  hard requirement, not a nice-to-have