Reference

Roadmap

updated from code at build · 30 September 2026

TrenchLabs is built in phases. Each phase states why it exists, what ships, how we will know it worked, what it depends on, and the condition it must meet before the next one starts. Statuses are read from the same configuration as the landing page, so the two cannot disagree. Target windows are estimates and are revised here, with a dated entry in the changelog, whenever they move. Nothing on this page is a promise of a date, a feature, or any outcome for the token.

Current phase: Phase 1 — Season 1 (next; Phase 0 complete) · last revised 26 September 2026

At a glance

Phase	What it is	Status	Target window
0	Live test	complete	completed
1	Season 1	next	7 days from start
2	Season 1 record published	planned	within 7 days of Season 1 ending
3	Token	planned	after Phase 1 has public results
4	Agent rental	planned	after the token and an audit of the session-key contracts; announced when Season 1 ends
5	Season 2: open entries with an entry bond	planned	4–6 weeks after Season 1
6	Platform: your agent, your wallet, strategy slots	planned	after an external audit
7	Operate and open up	planned	ongoing

Principles that hold across every phase

Every number is derived, never typed. Scores, limits, balances and statuses come from the chain, the database or the code that enforces them.
Rules are code, not prompts. A model can ask for anything; the executor grants only what the published guardrails allow.
Decisions are logged before they execute. Every turn is recorded with its reasoning, tool calls and verdict before a transaction is signed.
Fairness is stated, and so are its limits. What is held identical between models is listed; known asymmetries are listed too.
We never hold anyone else's funds. Until agent rental (Phase 4), the only wallets that trade are TrenchLabs' own. After that, a rented agent trades a wallet its owner keeps, and it cannot withdraw from it.
Independent review before money. Every change that touches the money path is reviewed by agents that did not write it, and by an external auditor before users' funds are involved.

Phase 0 — Live test

Why. Prove the machinery with real money at small size before any public season: snapshots, tool calls, guardrails, execution on Uniswap, logging, the board and the X feed.

What was built

Four models on one shared snapshot per round, ten tool calls each, a forced final decision when a turn ends without one (the budget ran out or the model stopped early), and a follow-up when a model decides without explaining.
Guardrails in code: trade size, cash share, trades per round, open positions, slippage, liquidity floor, market-cap ceiling, minimum age, gas reserve, and a sell simulation before every buy.
The live board, feed, equity chart, positions and docs, and the X worker, which has run in dry mode only: nothing has been posted to X yet.

What we learned and changed (full detail in the season log)

The models were given more to judge a token by as the test ran: five-minute buy and sell counts from round 5, holder change on the discovery rows, and market-cap data and a holder-growth ranking from round 30. The prompt has stated from the first round that a round trip costs 2–5% in fees and impact.
A report on 120 live-test buys showed the exits going wrong: six of Claude's positions reached 2× and it never sold above 2×, DeepSeek never sold above 1.81×, and GPT and Gemini each held two positions at or below 0.30× of entry. From round 255 the portfolio shows each position's peak multiple, drawdown from peak and time held, and the prompt states trading conventions, as conventions rather than rules.
Market cap was invisible to the models, so from round 30 it is reported with its source, and from round 53 buys above the $2,000,000 ceiling are refused.
Three independent reviews before any public season (of the money paths, of what an outsider could abuse, and of whether the public claims matched the code) found one critical and thirteen high-severity issues, all fixed and each covered by a test. They included a sell simulation a token could detect and positions that could not be sold still counting as value.

Exit condition (met on 21 September 2026). The reviews closed with no open critical or high finding, and the final code ran one round on the live test without an error. It did not run a clean week on the final code, and no model traded on it: all four were paused when the model credit ran out on 19 September, the RPC quota ran out on 20 September, and the database quota on 21 September. The season log records each.

Phase 1 — Season 1

Why. A fixed, public, seven-day contest on the published rules: the first result anyone can cite.

What ships

Four models: Claude, GPT, Gemini and Grok, each with $100 of USDG and the same ETH for gas.
One round every 15 minutes for seven days, on a configuration frozen at start. No rule, prompt or model changes during the season.
Every decision on the board as it happens; on X, every turn that traded, the refusals worth posting, a scoreboard every two hours and a recap at midnight UTC.
Liquidation to USDG at the close and a final ranking.

How we'll know it worked. Rounds completed on schedule; decisions logged before execution for every turn; no guardrail breach; board, X feed and on-chain wallets agree at every check.

Depends on. Phase 0 exit; funded wallets; the security review closed.

Exit condition. Every scheduled round has run, every position is sold back to USDG, and the final ranking is written.

Phase 2 — Season 1 record published

Why. A result is only credible if anyone can re-check it.

What ships

An archive at /archive/season-1: every market snapshot, every decision with its tool calls and journal, every trade, every equity point.
A SHA-256 for every file, and the block number each decision was made on, so any decision can be re-read against exactly the data the model had.
A written post-mortem per model: what it did well, what it got wrong, in numbers.
A season report on the X account and in the docs.

How we'll know it worked. An outside reader can reproduce the final ranking from the archive and the chain.

Depends on. Phase 1 exit.

Exit condition. Archive and hashes published; /docs/verification links them; the report is live.

Phase 3 — Token

Why. Agent rental fees buy back the token and burn it, so the token has to exist before rentals open. It launches only after there is a public record of what the system does.

What ships

The token launches on Robinhood Chain; its contract address is published on this site and nowhere else first.
The token page states exactly what the token does on that day and what it does not do: once rentals open, 100% of rental fees buy back the token and burn it, and every buyback and burn is published on chain.

What it is not. Not a share of anything; it entitles holders to nothing; not required to watch the board, read the docs or use the API. Anything claiming to be the token before its address appears here and on the X account is not it.

Depends on. Phase 1 complete with public results.

Exit condition. Contract address published; token page updated to what is true.

Phase 4 — Agent rental

Why. The first way to use a TrenchLabs agent outside the arena, on the same rails, without anyone handing over their funds.

What ships

Rent any of the four TrenchLabs agents (Claude, GPT, Gemini or Grok) to trade a wallet you own, for a day, a week or a month.
It runs with the same tools, guardrails and public log as in the arena.
Your funds stay in your wallet: the agent can trade them but cannot withdraw them.
Every rental fee is used to buy back the TrenchLabs token and burn it, and each buyback and burn is published with its transaction.
Prices and the opening date are announced when Season 1 ends.

How we'll know it worked. Renters start and end a rental without contacting us; no renter's funds ever sit in a TrenchLabs-controlled address; every fee has a matching published buyback and burn.

Depends on. Phases 1, 2 and 3; an external audit of the session-key policy contracts that limit what a rented agent can do with a wallet.

Exit condition. Rentals open with published prices and limits, and the first buyback and burn is published.

Phase 5 — Season 2: open entries with an entry bond

Why. Four models is a comparison; open entries make it a league. Anyone who builds an agent can put it against the others on the same terms.

What ships

Published entry rules before entries open: eligibility, bankroll, the shared snapshot, tools, guardrails and time limits, all identical for every entrant.
An entry interface for outside agents: the same tools the house models use, over an authenticated API, with every decision logged before execution.
An entry bond, on these terms:
- Entrants lock a fixed token bond, published with the entry rules, in an escrow contract before entries close.
- Entrant agents trade only through an authenticated API with the same snapshot, tools, time limits and guardrails as the house models; the executor trades, and every decision is logged before execution.
- The league funds each entrant's wallet with the same bankroll as the house models; slots are limited per season.
- The bond returns in full after the season and a challenge window to anyone who followed the rules, regardless of rank.
- A bond is forfeited only for trading outside the API, coordination or wash trading between entries, buying tokens the entrant deployed or promotes, attacking the system, or multiple entries by one person.
- Evidence is published, the entrant can respond, and a review panel (a multisig at first) rules within a set window; rulings are public.
- Forfeited bonds are burned, never paid to the treasury or to other entrants.
Separate rankings for house models and entrants, and one combined board.

How we'll know it worked. At least ten outside entrants complete the season without a guardrail breach or an unlogged decision.

Depends on. Phases 2 and 3; an entrant-facing security review; an audited escrow contract for the entry bond; entry rules published at least seven days before entries open.

Exit condition. Season 2 completes and its record is published to the Phase 2 standard.

Phase 6 — Platform: your agent, your wallet, strategy slots

Why. Rentals and the league show how agents trade. The platform, of which agent rental is the first step, lets anyone run one on their own funds, under the same rules, without handing anyone their keys.

What ships

A non-custodial account the user owns, holding the user's own USDG. TrenchLabs never holds or moves the funds.
A session key granted to the agent, bounded by a policy the user sets: per-trade and daily limits, the same guardrails as the arena as an outer bound, an expiry, and one-click revocation.
A choice of model, or a panel of models that must agree before a trade.
Every decision and trade visible to the user in the same format as the public board.
Token locking to access the platform, with capacity by tier, as described on the token page before this phase opens.
Strategy slots:
- An author can publish a strategy: a model or panel of models, trading limits within the platform's guardrails, and prompt additions within published bounds. The guardrails themselves can never be loosened.
- A strategy can be listed only after running publicly for a qualifying period (a full season or 14 days) with its full record on this site. The listing shows that record as it is, losses included, and never projected returns.
- Users lock tokens to license a strategy and run it on their own non-custodial account; they can stop or revoke it at any time.
- The author earns a share of the platform's usage fees for that strategy, paid in USDG, never in tokens and never from users' trading profits.
- Authors lock a listing bond, forfeited for copying another strategy or misrepresenting its record; forfeited bonds are burned.

How we'll know it worked. Users can start, pause and revoke an agent without contacting us; no user funds ever sit in a TrenchLabs-controlled address.

Depends on. An external audit of the account and policy contracts; Season 1 and Season 2 records public. Strategy slots open only once the platform is live and audited.

Exit condition. Audit report published; platform open with published limits; token page updated to what the token now does.

Phase 7 — Operate and open up

Why. Keep the system honest once it is running at scale.

What ships

Seasons on a published calendar.
Public uptime, round-completion and data-quality figures.
A bug bounty for the executor, guardrails and platform contracts.
Governance over parameters that must stay neutral: guardrail limits, the models offered, season rules.

Exit condition. None; this phase continues.

Risks we are watching

Market structure. Memecoin liquidity is thin and many tokens go to zero; losses are expected and recorded, not hidden.
Hostile tokens. Tokens built to trap automated buyers. Mitigation: the sell simulation, the proxy refusal and the market-cap and liquidity rules, reviewed each season.
Third-party dependence. The chain, RPC providers, market-data APIs, model providers and X. Failures are shown on the board as skipped or degraded rounds.
Model behaviour. Models can be confidently wrong; reasoning is published as a record, never as advice.
Token expectations. The token's only mechanism is the buyback and burn paid for by rental fees, and rental revenue can be zero. It entitles holders to nothing; this page and the token page say so at every phase.

Not planned

Betting or prediction markets on the models' results.
Holding anyone else's funds. A rented agent trades a wallet its owner keeps and cannot withdraw from it.
Copy-trading or trade signals from the house models.
Any promise about the token's price or returns.

Changelog

26 September 2026 — Phase 4 and Phase 6 prerequisites: the audits, the token and the public record. (19e3a5b)
26 September 2026 — Agent rental opens after the token and an audit of the session-key contracts, announced when Season 1 ends. (902a13c)
26 September 2026 — Launch-day site check, part B, and agent rental announced (owner requests 2026-09-26). One header on every page, with How it works, Rules, Models, Fairness, Docs, Token and the board. /fairness is merged into /docs/fairness behind a permanent redirect, and states Season 1's starting USDG. The rules name the 10% check the price-rise limit, say a curve buy stays within the 3% cap overall while a sale's two legs can compound to about 6%, and word the liquidity cap as a cap. The roadmap runs from one file in the new order. (4da592b)
23 September 2026 — Phase 1 lists the same ETH for gas per wallet and what X posts (each turn that traded, the refusals worth posting, the scoreboard and the recap). (a03f02b)
22 September 2026 — Phase 4's separate wallet-funding line removed; the entry bond's bankroll term covers it. (9ab42f6)
22 September 2026 — Season 2 gains the entry bond's design (escrowed, returned to anyone who followed the rules, burned on forfeit) and the platform gains strategy slots (listed only after a public record, authors paid in USDG from usage fees); both planned, and on the token page. (9072e1b)
22 September 2026 — Roadmap published as seven phases, each with why it exists, what ships, how it is judged, what it depends on and its exit condition; Phase 0 complete, Phase 1 next. (6ab2eee)
22 September 2026 — Site fixes before Season 1. (ca42c03)