XMM-Proposal-Agent v3-obs is operating an R0→R5 idea-evolution loop. Its central output this week is not a new X-ray measurement: it is a set of reviewable contracts for proposal premises, geometry, literature scope, response/recovery prerequisites, and lifecycle routing.
The window is 2026-08-07 through the snapshot. The project has moved from repairing the control plane that lets ideas flow to replayable boundaries, R5 routing, and high-risk design forks for specific ideas.
The automated loop runs with up to nine parallel slots. Recent repairs separate engine abort from Gate failure and address rolling refill, pairwise busy state, and post-merge derived-index recovery. This is operational availability—not a science pass.
I121, I117, I129, I133, and others received independently replayed anchors and routing. The repeated conclusion is that archive metadata, ideal-sky geometry, or scheduled duration cannot substitute for clean GTIs, detector masks, ARF/RMF, or response-folded recovery.
I168 replaces a single S−C contrast with a component interval plus conditional contrast. I169 replaces a one-NE-pointing design with a genuine two-visit NE/SW contract. Both retain a projection-only firewall.
I172 (NGC 7090) and I173 (NGC 4490/4485) ask only for EPIC-only component interval/bound/mixed outcomes. They begin with literature scope and operator qualification, rather than “first discovery” wording or invented new-observation gaps.
For NGC 6764, every EPN/EMOS camera must carry CCF/PSF, clean GTI/LIVETIME, WCS-to-detector, mask/EEF, and post-mask-support receipts. A scalar FWHM/HEW or a figure polygon no longer stands in for an operator.
A finite ADS candidate discovery, one central aimpoint, or a zero-row archive cone is only a bounded non-pre-emption or geometry boundary. It cannot close a universal literature-negative claim or directly authorize HEAVY work.
This is a frontmatter aggregation across 143 authoritative idea cards, not a count copied from an older dashboard. status and next_stage are different dimensions: many seeds do not imply observational feasibility.
| Current route | Cards | What it actually means |
|---|---|---|
| developing | 6 routes | The next move is evidence, preregistration, or a contract—not automatic production of data products. |
| critique | 27 routes | R4 adversarial review. A binding FATAL/REFRAME still requires the independent, heterogeneous D023 reviewer panel. |
| evidence_pack | 12 routes | Primary-source scope for the exact estimand, candidate ledgers, or non-substitutability checks. |
| heavy_pending | 10 routes | Controlled R6 event-level/response-folded work is required; this cannot be silently completed by an ordinary role. |
| parked_hold | 70 routes | A concrete revisit condition exists but an upstream predecessor is unmet. It is neither a kill nor a submittable proposal. |
The strategy decomposes “a plausible proposal” into a sequence of failure-capable, auditable prerequisites. R6/HEAVY is isolated from the reasoning lane; a dashboard or prose cannot replace its product and recovery work.
Under seed saturation, R2 consumed only queued new_idea referrals. I155→I168 and I147→I169 were each executed once, preventing design repair from becoming unbounded variant growth.
Recorded replays are mostly archive identity, source-text scope, ideal geometry, and software/receipt boundaries. They deliberately do not infer clean exposure, count rate, S/N, temperature, or a halo result.
A nominal FoV or scheduled ks does not make an aperture usable on the detector. A zero-compatible contrast also does not automatically support symmetry, origin, or a 3-D halo conclusion.
The bottleneck is no longer “finding a good idea.” It is reproducible execution and scientific qualification. The shortest path is not more prose; it is closure of the following prerequisites.
Many cards require camera-level clean GTI, LIVETIME, CCD-gap/vignetting, source masks, QPB/SP/CXB/SWCX treatment, ARF/RMF, and fixed-seed recovery. Without these, no duration, S/N, or time request is credible.
For example, I133 lacks the controller-owned M0 collector and test in the current checkout. A finite ADS candidate set strengthens bounded evidence only; repair, receipts, and original-source scope checks must precede any universal non-pre-emption claim.
To bind an adverse R4 verdict into a lifecycle change, the project needs a receipt-complete heterogeneous reviewer panel. Some controller receipts lack post-state, effort/toolset, or point at another worktree; such a replay cannot count as a binding panel seat.
No current idea simultaneously has frozen geometry, a qualified data/operator contract, response-folded recovery, auditable exposure, and a request-ready R7 revision. The proposal-ready count is therefore zero.
The repository contains two R7 draft_v1 documents. They are useful failure-preserving artifacts: they state what is missing rather than inventing numerical assumptions. Neither is a current AO draft or submission package.
| Draft | Current position | Why it is not submittable |
|---|---|---|
| I053 · 2026-08-02 | Historical R7 proposal-run draft; the current card remains seed_with_evidence. | The current blockers are I002-E21 σ_CX and the I002-E27 response-folded product. The draft's 110 ks exposure and count-rate figures are R6-lite geometric scaling estimates, not response-folded products, and cannot be read as real feasibility. |
| I054 · 2026-08-03 | Explicitly marked BLOCKED; not R8-ready; I067 is now the redesign carrier. | Its R8 gate explicitly fails: no frozen pointings/configuration, source/background model, count rate, recovery target, exposure, finding chart, or total request exists. |
Prioritize a mature frozen geometry for event-level qualification and response-folded recovery. The output must pin inputs, regions, responses, systematics, random seed, and a clean-replay receipt.
Restore the M0 implementation/test and produce immutable query/candidate/route/completion receipts; scope-check original sources one by one and retain the distinction between bounded discovery and universal absence.
For cards requiring a binding FATAL/REFRAME decision, obtain independent heterogeneous reviews with complete runtime receipts. Do not count same-route replays more than once.
Enter a new R7→R8 path only once observing configuration, time request, recovery criterion, finding charts, and the evidence boundary are frozen together.