FrontierLab-Meta-Hackathon-2026 / docs /PLANNING_READINESS.md
Jaycee766's picture
Deploy FrontierLab v0.35 source f7b67b6a3b09
a0e9e53 verified
|
Raw History Blame Contribute Delete
4.43 kB

Stage 5 planning readiness

FrontierLab separates planning readiness from planning authorization.

A run can have current evidence, human review, hypotheses and a researcher-selected method while the Planning Agent remains completely disabled.

Recorded M2 state

The current recorded biomedical path has:

  • two source-bound researcher-reviewed claims;
  • two citation-linked hypotheses from the bounded Muse Glimmer continuation;
  • a canonical Method Evaluation with M2 selected;
  • no generated plan;
  • no Stage 5 planning authorization.

Selected method:

M2 β€” Structure-based dual-pH histidine-scanning computational design

Canonical IDs:

  • context ID: ee7f790ba48474bb63a935f6108f1e698121a404d93a5bf0389946ba2932dcf8
  • evaluation ID: 00c440f3a78a95a1703725ea261b992ab51551d1dedd9cf41ba481dd0954f231

What the readiness packet proves

frontierlab.planning_readiness.build_planning_readiness(run) first re-runs require_evaluation(run).

If that exact evaluation is stale, altered or no longer bound to the current reviewed evidence/hypotheses, readiness fails.

A valid packet records:

status = READY_FOR_EXPLICIT_PLANNING_APPROVAL
planning_approved = false
model_call_authorized = false
planning_model_called = false
plan_generated = false

This is an eligibility/precondition record, not an approval record.

Open reproducibility checks

M2 currently has three unresolved reproducibility checks:

  • data β€” complete design inputs, intermediate ranking outputs and raw validation data are not yet packaged;
  • code β€” the original ADAPT implementation or a reproducible executable alternative has not been independently verified;
  • environment β€” dependency versions, computational environment and protonation/modeling settings have not been reconstructed.

The recorded parameters, protocol and metrics are source-documented but have not been reproduced by FrontierLab.

Planning obligations

Any later Stage 5 draft must retain these constraints:

  1. Design-input provenance β€” record structural/design inputs and candidate-ranking provenance.
  2. Executable design stack β€” identify ADAPT or a named reproducible alternative.
  3. Computational environment β€” record dependencies and protonation/modeling settings.
  4. Matched-pH controls β€” prespecify acidic vs physiological-pH comparisons and falsification conditions.
  5. Fab-level validation β€” keep Fab readouts distinct from cell/full-antibody measurements.
  6. Orthogonal SPR β€” where feasible, use SPR on monomeric Fab to separate intrinsic affinity from avidity/stoichiometry.
  7. Avidity/stoichiometry caveat β€” do not overinterpret apparent cell-binding affinity mechanistically.
  8. Structure/search bias β€” surface structural-model uncertainty and the histidine-centered search assumption; M1 remains a comparator for histidine-independent solutions.
  9. Replication/falsification β€” include independent replication, stopping criteria and explicit falsification conditions.

Evidence boundary

The packet distinguishes:

  • selected-method evidence: PMC6927761;
  • all currently reviewed evidence available to planning: PMC4314675, PMC6927761.

The Planning Agent, if later authorized, may use only reviewed evidence and must cite the selected method's source.

Recorded builder

python -m scripts.planning_readiness_recorded

reconstructs the exact October 28 extraction, human review, reviewed continuation and Method Evaluation, then emits a readiness packet with zero model calls.

The recorded researcher-selection provenance remains separate from the canonical Method Evaluation so the latter continues to rebuild exactly from code.

UI

When a current evaluation exists and no plan has been generated, Streamlit exposes:

Planning readiness Β· pre-approval contract

The panel shows the selected method, open reproducibility checks and all required obligations. It does not add a new authorization control.

The existing Stage 5 checkbox remains the only researcher authorization for draft planning.

Scientific boundary

Planning readiness does not establish that M2 works, that the original experiment is reproducible, or that any proposed plan is safe/effective.

It only establishes that FrontierLab's current source-bound workflow state is internally eligible to ask the researcher for an explicit Stage 5 decision.