Skip to content
BicentennialGitHub

Roadmap

From developer kit to sealed board

Two tracks run side by side. Developer releases ship the tiers and tools you run on your own system; they never certify and never wait for the board. The board track validates the measurement ladder in the pilot and freezes the next pre-registration, and gates only the v1.0 board.

A release reads "released" only when its tag in the public repository dates from no later than the commit this site is synced to. The full text, with every acceptance condition, is in ROADMAP.md.

  1. v0.1 Kit

    releaseddeveloper release
    • The public repository, the generator core with the G1 and G2 genres, the scorer, the director and the Hermes adapter.
    • Ring 0 developer packs for the Remember, Survive and Act stages: T0 probe packs and T1 CI smoke.

    Done when: T0 and T1 run on Hermes native memory and two external systems, T1 is deterministic on stable fields, and the wiped-store positive control fails it.

  2. v0.2 Developer loop

    releaseddeveloper release
    • T1 for every registered system on its own lanes.
    • T2-lite: a live smoke on T1's lives, where only the probe answers come from the live pinned model.
    • A per-item flip list against the developer's own baseline, with a rerun command and a repeat count.
    • T1's loop time measured per arm, with its breakdown.

    Done when: T2-lite runs Hermes native live on the pinned model twice (A and A′) and publishes each card with its flip list, q₀, loop time, requests and spend.

  3. v0.3 Developer tools: close the loop

    nextdeveloper release
    • Repro bundles and a repro command that checks every pin and hash before it runs.
    • The developer report: the diff against your previous version, with the noise band between two identical runs and the smallest detectable effect.
    • T2 live smoke at its pre-registered size, and Ring 1: a fresh-seed run of the public templates at each release and at least monthly.
    • Bring your own system from outside the benchmark checkout, and T1's loop-time cut.

    Done when: The benchmark team takes a public failure of its own naive-RAG baseline from a T2 run, reproduces it from its bundle, fixes it, and confirms the fix on a scheduled Ring 1 run.

  4. v0.4 Developer tools: search and sensitivity

    planneddeveloper release
    • Rungs per cell, so a developer sees movement before a verdict changes.
    • The fault finder: a scripted search over levers and fault points, then shrinking a failure to a minimal public life.
    • Live scoring of G2 actions by meaning, and the per-item trace page.

    Done when: The fault finder finds and shrinks a public failure within its budget, and every cell's rungs are frozen before any real system runs on them.

  5. Pilot

    plannedboard track
    • The Image profile in Ring 0 once its spike admits it, the first Learn chains, the calibration set, and the runs of record for the pilot bars.

    Done when: The pilot bars pass, the agent-validation gates pass, the Scoring Veto Panel of agents has reviewed the candidate, and PREREG-v1 is frozen with its hash posted.

  6. v1.0 Board

    plannedboard track
    • The sealed season-1 board for G1 and G2, with each profile's required Act cells, baselines and operating metrics published before submissions.
    • The seasonal robustness study, at least three external competitors, Ring 2, and the site's leaderboard.

    Done when: The season-1 board is published with confidence intervals, cell verdicts, certification on Act per declared profile with its stability, and the operating metrics.

  7. v1.x

    plannedlater
    • A sealed slice for model labs, community templates, audio and video, more harness adapters, and vendor-submitted configurations.

    Done when: As each lands.

  8. v2 World and research tier

    plannedlater
    • World profiles, each admitted only when it exposes a memory question the earlier ones cannot: E1, a robot vacuum in a 2D home; then E2, a desk arm in simulation.
    • A research tier with long-horizon lives.

    Done when: A profile joins the board when its pilot shows it discriminates.

  9. Later

    plannedlater
    • A mobile manipulator and a household helper; longer simulated lives. The full android remains a design target.