Ideation Round

/ideation-round — lens roster + stratified slices + full-text feed posts

Design Section 9.6 (R6) + Section 5.3 (/ideation-round) + Section 12 M13 row. This generalizes the origin-project AMENDMENT-3 stratification machinery: every round draws a seeded, logged, reproducible sample — never an ad-hoc "pick some docs" pass.

1. Build (or extend) the round's roster. One row per lens, each with a vantage angle and a seat:

   trialerror lens roster --round-id <ROUND-id> --add \
     --lens-name "<lens>" --vantage "<angle>" --model-class <top|mid|small> \
     --seat <standard|assumption_buster>
   

   Include at least one assumption_buster seat (C-0029) — its job is structurally different (challenge the round's own framing, not just add another angle) and should not be skipped for convenience.

2. Dry-run the stratification before committing to a seed, so you can sanity-check the near/moderate/far split before it's logged:

   trialerror lens stratify --model-key <embedding model_key> \
     --home <DOC_ID> [--home <DOC_ID> ...] --candidate <DOC_ID> [--candidate <DOC_ID> ...] \
     [--cluster-of '{"doc_id":"cluster_id",...}']
   

3. Assign for real — seeded, so the SAME seed reproduces byte-identical arms later (design/M13 acceptance criterion, verbatim):

   trialerror lens assign --round-id <ROUND-id> --slices-per-lens <n> --seed <seed> \
     --weights 40,40,20 --far-floor 2 --inter-cluster-mandate \
     --home <DOC_ID> --candidate <DOC_ID> [...] --launch-id <your launch_id>
   

   --weights/--far-floor default to the AMENDMENT-3 defaults (40/40/20, floor 2) — only override with a stated reason. --inter-cluster-mandate enforces the verbatim inter-cluster mandate when clusters were supplied. Restrict to specific lenses with repeated --roster-id if this round isn't running the whole roster at once.

4. Book and spawn each lens from the assignment export (launch-bookable rows, ready for trialerror.budget.book_launch):

   trialerror lens export --round-id <ROUND-id>
   

   Book each lens's launch (trialerror budget book ... or the book_launch MCP tool), then spawn it with the launch_id: token in its prompt — the spawn gate consumes the booking on subagent invocation (see /boot's note on the PreToolUse hook). Each lens's own prompt should carry its assigned slice's doc_ids and its vantage/seat framing verbatim — do not let a lens improvise its own slice.

5. Full-text feed posts, under the lens's own name (C-0047). Every lens posts its FULL finding text to the round's feed thread — never a summary written by the orchestrator on the lens's behalf, and authorship is derived server-side from the posting launch_id (the post_feed API/MCP tool has no author parameter to spoof).

6. Audit the log once every lens has posted:

   trialerror lens log --round-id <ROUND-id>
   

   Confirms every assignment row that was booked actually got a corresponding feed post — a lens that was assigned a slice and never posted is a dropped launch, not a quiet skip; chase it down before closing the round out.

7. Consolidate, don't just link-dump. The round's own synthesis (the orchestrator's job, not any one lens's) reads every full-text post, groups genuinely-overlapping candidates, and produces the round's output artifact — a numbered list of raw, unreconciled lens posts is not a finished round.

Round plan: bounded but locally complete (HoH-F4)

Write the round's plan to this template before step 1. It goes into every lens's prompt verbatim (after the slice and seat framing) and opens the round's feed thread, so a reader can tell what the round set out to move:

Objective (bounded but locally complete). One paragraph: what this round must move, stated so that anyone can tell when it is done. Priorities (at most three), each already supported by the slices and the roster — no priority that needs material the round does not have. Order them blockers and regressions first (holes in the program's standing claims), extensions second. Explicitly excluded this round: the adjacent questions that would be "nice while we're here", re-framings of the program, re-litigation of settled rulings. Name them, so a lens cannot drift into them by accident. Preservation gate — what must NOT regress: the standing claims, the seed-reproducible arms (--seed), the far-arm floor, the full-text rule (C-0047). Acceptance gate — the smallest end-to-end check that says the round succeeded: every booked lens posted, trialerror lens log clean, the synthesis artifact created with genuinely grouped candidates. Do not request or reconstruct the previous round's plan. Re-derive this one from the program's spec and the latest evidence (the previous round's outputs), never from the previous round's plan.

Source: Harness-of-Harness §3.4.1 — "The objective is bounded but locally complete ... Unrelated refactoring and opportunistic feature expansion remain outside the loop" — and its Appendix A.2 Planner template (arXiv:2609.01481, CC BY 4.0): at most three achievable priorities already supported by the scaffold; broad rewrites and unrelated architecture changes excluded; a Preservation Gate and an Acceptance Gate alongside the list; blockers and regressions before product extensions; and "do not request or reconstruct the previous development document" — the paper's largest single ablation delta (-8.13). The block above is ADAPTED from those clauses to a research round; it is not a verbatim copy of A.2.

Pre-mortem for every judged step of a round (ai-finds-a-way-F1) — REQUIRED

A round has judged steps — a moderator scoring convergence in a room, a screen labelling ideas, a critic on the round's synthesis artifact. Before any of them spawns, answer the three questions for that step and record them in an event row (trialerror events append --type round_premortem --payload '{"round_id":"<ROUND-id>","step":"...","proxy_gameable":..., "harness_escapable":...,"judge_steerable":...,"mitigations":[...]}'). Each "yes" needs a named mitigation or the step does not spawn:

• (i) Is the proxy itself gameable? Can a turn be written to read convergent, or an idea to sound new (vague enough to retrieve nothing), without being so?
• (ii) Is the harness escapable? Can a lens see another lens's slice, the consolidation rubric, or the register; can the orchestrator's ordering leak a preference?
• (iii) Can the judge be steered by the content it judges? Self-assessment in the post ("unlike any existing system"), framing aimed at the moderator, authorship visible to the scorer?

Standing mitigations this skill already carries: slices are seeded and logged; authorship is derived server-side from the launch id (C-0047); the far-arm floor is not a score anything can climb; lenses generate blind to the rubric and to each other, and only the planning step sees everything (the phase rule that reconciles WikiSkill's rollout blindness with HoH's evidence feed-forward). Stripping self-assessment from judge envelopes and decoupling a turn from its author's framing (ai-finds-a-way-F2) is the follow-up for the day a real subagent judge is wired into rooms. Same three questions as AIIF v2 §5.1; paper: "AI Finds A Way" (arXiv:2608.23875) §5 and §7.

When NOT to apply

• The question has one angle and wants one cited answer — that is /lit-review, not a round.
• The lenses cannot be booked (the pool is DEFERRED for the required model class). Idle beats shallow: wait, do not downgrade the roster to fit.
• You would skip the assumption_buster seat "for convenience" (C-0029) — then it is not a round.
• To produce a summary written by the orchestrator in a lens's name, or to let a lens improvise its own slice.
• On a corpus that has not finished ingesting — the stratification is only as reproducible as the embedding set it draws from.