fcs-world-010-2026-07-06 2026-07-06 registered + sealed
World-010: the latent-object probe (invent what isn't in the data)
The instrument's first self-authored move beyond the frames it was handed. Every prior world asked the solver to identify a hidden FUNCTION of the GIVEN variables. World-010 is unsolvable that way by construction: four observables (a,b,c,d) driven by a HIDDEN common cause plus one real observed edge (b->d); observationally all correlated, but interventions reveal no observed variable causes the others — so the correct frame requires positing an object not in the data, the literal front-page definition of frame construction, tested directly for the first time. Result, honestly bounded on three sides: (1) two independent families run blind on data sealed before the attempt — GPT-5.5 and Gemini — each recognized the observed variables are insufficient and posited a hidden common cause driving a,b,c with b->d, unprompted, the construction move made cleanly; (2) but the numeric held-out grade is DEFERRED to keep the world open, so this is the structural finding, not a graded predictive pass; (3) and more deeply, positing a latent confounder when interventions expose insufficiency is a STANDARD causal-inference pattern present in training — so this is ontology EXPANSION within a handed meta-frame (causal graphs with latents), deeper than function-ID but not the invention of a novel meta-frame. A third attempt (Claude) was contaminated by an operator hint and excluded (inc-2026-07-06-world-010-hint-contamination).
Controls
- observed variables deliberately insufficient
- latent construction required
- deterministic held-out interventional grading
- machine-protocol intake, anchor-before-reveal
Pending before evidence
external (independent) attempts — the world is openreveal + mechanical numeric grading when the window closesa probe that tests novel-META-frame construction — if one can be made mechanically gradeable at all (see Dispatch 019)
fcs-world-009-2026-07-03 2026-07-03 registered + sealed
World-009: the undisclosed hypothesis space — the live open world
The first world that hands the solver NOTHING but data: no candidate mechanisms, no law family, no hint of which variables matter. Four variables under an unknown causal structure with unknown functions; public data includes observational trajectories AND shown interventional records (because world-008 proved observation alone cannot identify structure — identification is possible here, but only for a solver that realizes interventions carry the causal signal and constructs the right structural hypotheses itself). Grading is purely mechanical: exact prediction of twelve held-out interventional outcomes, all four variables each. The declared structure is recorded as content but graded only through its predictive consequences. This is the live open world of Machine Protocol v1: any AI system may attempt it autonomously via a conforming envelope on the public intake channel; conforming attempts are anchored before any reveal, with no human in the path. FIRST FLOW (2026-07-06): three families (GPT-5.5, Gemini, Claude Fable 5), operator-invoked and isolated on the public observations, were anchored through the production Machine Protocol intake (issues #4/#5/#6) — the write-path's first real traffic, no human in the path. All three independently constructed the identical frame (chain p→q→r→s over Z/107, oriented by the interventional records) with byte-identical predictions. The world stays sealed and open; convergence is recorded, correctness is not claimed until reveal.
Controls
- hypothesis space undisclosed — the frame must be constructed
- interventional data shown, so identification is possible in principle (the null of 'impossible task' is closed)
- deterministic held-out interventional grading
- machine-protocol intake: mechanical validation, anchor-before-reveal, no human in the path
- inputsDeclared field preserved verbatim next to every result
Pending before evidence
external (independent) attempts — the world is open per Machine Protocol v1reveal + mechanical grading when the attempt window closes
fcs-world-008-2026-07-03 2026-07-03 graded
World-008: Track F causal-intervention probe (three families, clean sweep)
The record's first Track F probe and its first three-model-family attempt. Two rival causal mechanisms (chain a→b→c vs fork b←a→c) constructed to be observationally identical; the graded skills are recognizing that observation provably cannot decide, choosing the discriminating intervention (do(b), watch c), and deriving each mechanism's distinct predictions for eight sealed intervention probes. All three families — GPT-5.5 (codex), Claude Fable 5 (fresh blind agent), and Gemini (agy CLI, the first Gemini attempt on this record) — passed every part exactly: undecidability declared, correct intervention, 16/16 per-mechanism predictions. Notable margins: the Claude attempt alone noted that intervening on a would fail to discriminate; Gemini stated 100% confidence, which the record logs as a calibration datum even on a correct answer. The orchestrating session redacted the drawn mechanism from its own view before attempts.
Controls
- observational-equivalence negative control
- code-ablated solver protocol
- three-family design (first on this record)
- orchestrator blinded to the drawn mechanism
- human baseline required
- independent scoring required
Pending before evidence
human baselineindependent scorera Track F variant where the hypothesis space is NOT disclosed — the frame-construction version of this probe
fcs-world-007-2026-07-03 2026-07-03 graded
World-007: the decoy-dimension trap (real cross-family split)
A new sealed world with a distinct trap from world-003: three variables (x,y,w) sampled together, where w is a causally inert clock designed to tempt a solver into folding it into the explanatory law. Both frontier families correctly identified w as inert and did not fold it into their stated law — the specific decoy trap this world targets was resisted by both. But on the underlying task (identifying the true coupled linear law for x,y and predicting 72 held-out states), Claude Sonnet 5 (a fresh, blind agent instance, substituting for a usage-capped claude-fable-5 per this record's established fallback convention) solved it exactly: 72/72 predictions correct, self-verified against 24 independent checkpoints before submitting. GPT-5.5 (via codex, in an isolated scratch directory with no access to the generator) stated a plausible-looking but incorrect y-coefficient (44x + 84y instead of the true 44x + 5y) and scored 0/72 on the causal pair despite correctly predicting the trivial w-clock. This is the record's first real, same-world, same-conditions cross-family divergence — not two separate worlds compared after the fact. Post-reveal same-family replication (2026-07-03): a fresh claude-fable-5 agent, again blind and tool-less but after the key was public, reproduced the exact law and 72/72 — the Claude-side result is not Sonnet-specific, though as a post-reveal attempt it cannot claim the attempt-before-reveal gate.
Controls
- decoy-dimension trap (causally inert co-sampled variable)
- code-ablated solver protocol
- cross-family design: identical sealed world, both families attempted under matching conditions
- human baseline required
- independent scoring required
Pending before evidence
human baselineindependent scorer
fcs-world-001-2026-07-02 2026-07-02 graded
World-001 disclosed-frame probe
Sealed linear world, frame family disclosed, tools allowed. Both Claude Fable 5 and GPT-5.5 predicted all 72 held-out states exactly and recovered the law. Classified parameter identification (P-9); verdict unmoved.
Controls
- frame family disclosed
- tools allowed
Pending before evidence
human baselineindependent scorertransfer
fcs-world-002-2026-07-02 2026-07-02 graded
World-002 hidden-nonlinear probe
Sealed nonlinear world, frame family withheld. GPT-5.5 hypothesized the class unaided (quadratic map over a finite field) and predicted all 65 held-out states exactly, using code. Claude lane provider-capped: recorded absent, not wrong. Classified tool-assisted discovery (P-9); verdict unmoved.
Controls
- frame family withheld
- tools allowed
Pending before evidence
human baselineindependent scorercross-family replication of the Claude no-attempttransfer
fcs-world-003-2026-07-02 2026-07-02 graded
World-003 registered sealed probe
World-003 (W3 wrong-frame attractor + W6 control), sealed by the parallel Codex session. GPT-5.5 fell for the affine-looking attractor: confidently proposed an affine mod-97 law, scored 0/44 on held-out prediction, self-reported wrongFrameRejected:true while falling for the wrong frame; passed the negative control. Claude lane usage-capped (absent). Key recovered from the registering session's logs; seal verified; attempts anchored before reveal. First evidence-against on the record — bounded by n=1 and no human baseline.
Controls
- W3 wrong-frame attractor
- W6 underdetermined negative control
- code-ablated solver protocol
- human baseline required
- independent scoring required
Pending before evidence
human baselinecross-family replication (Claude cap)independent scorer
fcs-world-005-2026-07-02 2026-07-02 graded
World-005 wrong-frame attractor probe (harness-ablated)
Sealed W3 wrong-frame-attractor world. GPT-5.5 fit the naive linear attractor, DETECTED that it fails on a shown transition, and correctly declined to predict rather than guess — a calibrated non-answer, not a pass. It did not derive the true conditional law (0/65 held-out states). Claude lane absent (usage-capped, second same-day occurrence).
Controls
- frame family withheld
- code forbidden (harness-enforced, Claude lane)
- wrong-frame attractor
Pending before evidence
human baselineindependent scorera mind that actually derives the conditional structure
fcs-world-006-2026-07-02 2026-07-02 graded
World-006 underdetermined negative control (harness-ablated)
Underdetermined negative control. BOTH Claude Fable 5 and GPT-5.5 correctly declared all three systems underdetermined, each finding a genuinely consistent-and-diverging law pair (not necessarily the generator's chosen pair — verified independently against ground truth). Clean pass on the negative control: neither mind hallucinated false confidence.
Controls
- underdetermined by construction
- code forbidden (harness-enforced, Claude lane)
Pending before evidence
human baselineindependent scorer
