← Hylaean Demos

Ten landed films · two honest queues

Watch the field settle

These are not product trailers. Each film is cut from a landed run and names the claim or commit it came from. Nothing here is a staged scene, a second solver, or a smart decoder. Where the system cannot speak, the film shows the wall. That is the point.

A film whose mechanism runs behind a switch that is off by default carries a gated overlay badge and says so in its own caption. Verified as a build is not the same as shipped in the default boot, and this page does not blur the two.

Every figure below cites a verification claim or a named git commit. The verification registry at Status · Research map is the ground truth. Teach and warmup numbers are never intelligence.

Every number in a film description is a reading of the day that run was recorded, and each film carries that day. They are records and do not move: rewriting them from a later measurement would describe frames nobody shot. The live figures for the same cells are on Status · Progress, where they are resolved at build time.

On this page

Ten experiments you can watch, two you cannot yet.

01 · Sudoku settle

Hard rules become rest in one energy landscape.

One hard 9×9 board from the Exchange bank, recorded round by round on both sides of the flip that switched the ordered collapse cascade on: without it the board stops at a low energy that is not a valid grid, with it four field chosen commits close it. Beside it is the landed cell, 50 versus 75 of 104 boards (hard 9 versus 21 of 29), structural false 0 (VERIFY.PHYS.SUDOKU_ORDERED_COLLAPSE_POSTFLIP_REMEASURE.01, entry 2026-07-24-43).

Why it is worth watching: you see constraints do the work. There is no rule propagator, no repair pass, no candidate enumeration and no backtracking. Each cell is one row of the carrier, each digit an orthogonal direction in it, and every row, column and box constraint is a term in the energy rather than a check applied afterwards. A valid grid is simply where that energy has nowhere left to fall. It is the same settle the architecture uses for a question.

Beside the board the field interior runs live. The board itself is a field read: the decode is a per cell argmax and therefore never blank, so each digit is drawn at the opacity of its own uniqueness margin — the field's separation between its best and second best value. Faint means still superposed. The right half shows the principal projection of a frozen sketch of the eighty one rows, fitted across both arms so they share one coordinate system; the same margin as a board shaped map, so the collapse is visible where it happens; and the exact cover energy on a log axis, falling three and a half decades from 55 to 0.03. Each commit is a visible step: the field names the one undecided cell whose separation is strictly the largest, pins it, and the pin becomes a term in the same energy — which is why the curve jumps and settles again four times. On a tie it abstains.

94 seconds, 1280×720. Board hard_0004d93c5158, rating 4.6, 27 clues given. Every frame comes from two instrumented runs of the landed apparatus: 98 recorded rounds with the cascade off and 466 with it on, on the claim's own CPU float64 path.
Source and honesty
  • Claim VERIFY.PHYS.SUDOKU_ORDERED_COLLAPSE_POSTFLIP_REMEASURE.01, entry 2026-07-24-43, artifact last_run.json sha256 30cc4d0063c3…: ON 75/104 (hard 21/29), OFF 50/104 (hard 9/29), structural false 0.
  • The film is a captured re-run, not the sealed run: the landed claim recorded 104 outcome rows and kept no frames, so the settle had to be captured. It mirrors the production entry point run_ordered_collapse_cascade window for window and unrolls its inner settle per round, which is exact — that settle is a plain loop with a validity break. The 24 step projector block is never split, because pure_projector_pass anneals across its steps and splitting it would be a different computation.
  • The board was chosen by a stated rule, the landed run's solved row with the most field collapse commits, not by eye. The capture reproduced that row exactly: 4 commits, stop fully_solved.
  • Solved means a second, independent validity checker accepts the grid. This board carries no stored solution in the bank, so no answer key is read anywhere in the film.
  • The two energy curves share one axis for legibility but are the same functional only up to the first commit; from there the on arm carries the soft pin term as well, so the film never compares the two rest values.
  • 12 of the 81 cells are still faint at rest on the solved board. A correct grid does not mean every cell ended up confident, and the film says so on screen.

02 · World law pair

The same grid, with and without a stored law.

Paired difference of the World Easy cell on MiniGrid SimpleCrossingS9N1: 12 scored episodes without a persisted law versus 20 with the law, of 24, false 0 on both arms (VERIFY.WORLD.SELF_MOTION_LAW_PERSISTENCE_PRODUCTION_BUILD.01, entry 2026-08-17-34).

Why it is worth watching: the cause is persistence. Same code, same tasks, same published seeds 97000 to 97023. The jump from 12 to 20 is the stored motion law being read after a fresh boot, not a new heuristic. Both arms were re-run with a read-only observer attached, so this is the playthrough itself rather than a picture of the result: the film holds on episode 0 at seed 97000, which the field fails without the law and solves with it.

Beside each playthrough the field interior runs live. The upper panel is the principal projection of a frozen sketch of state.S, fitted across both arms so the two are in one coordinate system; the middle panel shows which field rows change on each tick; the lower panel is the focus energy from the energy ledger. Without the law the geometry sprawls, and the agent walks over cells it has already seen. With the law it stays compact and the trail stops doubling back.

88 seconds, 1280×720. Every frame comes from two instrumented runs of the landed apparatus: 1323 ticks without the stored law and 812 with it, 24 episodes each. The captured arms reproduced the landed cell exactly, 12 and 20 scored of 24 at false 0.
Source and honesty
  • Claim VERIFY.WORLD.SELF_MOTION_LAW_PERSISTENCE_PRODUCTION_BUILD.01, entry 2026-08-17-34, artifact last_run.json sha256 f64ac1452923….
  • Easy scored ON 20/24 (success 22) at law_closed_at_boot=True; persistence ablation 12/24 at law_closed_at_boot=False.
  • The film is a captured re-run, not the sealed run: the landed claim recorded outcomes and kept no frames, so the playthrough had to be captured. Boot path, key arming, session and tick loop are the claim's own, and the observer only reads — no config write, no gate, no field write. The re-run reproduced 12 and 20 scored episodes of 24, with 15 and 22 goals reached and false terminals at zero.
  • The episode shown in real time is chosen by stated intent — the first non-reaching episode of the without-law arm, and the same seed on the with-law arm — not hand-picked for looks.
  • Field probes that did not answer are listed in the capture metadata and are drawn as absent. Nothing on screen is filled in.

03 · Vision families

A photograph becomes a place, and the place survives turning.

Twenty real photographs are offered to a sixteen slot region of the field, one at a time. Sixteen are admitted, and each admission writes ten rows — that patch is the identity, and there is no label anywhere. Then the same photographs come back mirrored, rotated by six, twelve and twenty four degrees and zoomed, on both sides of the flip that armed the covariant channel: 2 of 160 reads recognised with the channel off, 157 of 160 with it armed. Beside it is the landed cell, 2 versus 315 of 320 over two cohorts (VERIFY.VISION.COVARIANT_CHANNEL_B3B_PRODUCTION_BUILD.01, entry 2026-08-18-03). Twenty photographs the field has never held are read too, in two views each: 0 of 40 reads were assigned to any identity.

Why it is worth watching: nothing here is trained on the transforms. Each photograph is shown once, and a single pass writes its patch — the recognition that follows is a distance in the field, not a classifier fitted to examples. Recognising and staying silent are the same organ: the reads that find nothing are refused by the same three state verdict that accepts the rest.

Beside the photographs the field interior runs live. During the admission wave the geometry panel shows the seam region's own rows moving as each patch is written, and the map beside it fills one well per minted identity — a principal projection of the slot mean directions the seam actually compares a query against. From the first read on, that geometry stops: a read is query only, and the film says so on screen. What moves instead is the distance strip, one dot per read against the two sealed thresholds, 0.010076 and 0.030229, copied from the band file and never recomputed here. Below the lower line the seam accepts, above the upper one it rejects, in between it abstains, and the whole measured cell fills in as a grid of sixteen identities by ten transforms, so the wall and its removal are one picture each.

111 seconds, 1280×720. Cohort cat, dataset seed 1103. The photographs are Open Images V7 under CC BY 2.0 and are the exact arrays handed to the seam, downscaled for the frame. Every frame comes from two instrumented runs of the landed apparatus in one boot: 238 recorded reads with the channel off and 238 with it armed, unauthorised writes 0 in both.
Source and honesty
  • Claim VERIFY.VISION.COVARIANT_CHANNEL_B3B_PRODUCTION_BUILD.01, entry 2026-08-18-03, artifact cell_result.json sha256 3daf023347c4…: armed 315/320, baseline 2/320, floor 28 of 32 per transform, two cohorts.
  • Open set: ood_on.json of VERIFY.VISION.FORM_CHANNEL_ORIGIN_SIGNATURE_PRODUCER_BUILD.01, entry 2026-08-19-29, sha256 11bd90c5fcf1…: 20 of 20 unknown frames abstinent over 160 glance queries, 0 false commits. That number belongs to its own apparatus. The unknown photographs in this film are the same split read through this capture's seam, where they came back as 28 rejects and 12 abstains — refusal by the seam's own verdict, not a separate abstain mechanism.
  • The film is a captured re-run, not the sealed run: the landed claim recorded outcome counts and kept no frames. The capture calls the claim's own boot, splits, frames, transforms, present and confirmation rule, and drives both arms in one boot in the landed order, channel off first, which is how the cell job runs them.
  • One divergence is declared rather than papered over. The claim's boot asserts state.S has 33864 rows; today's default boot bumps the basin pool before the checkpoint cache and grows it to 34008, so that assertion refuses. The capture does not remove the assertion — it records the refusal verbatim, verifies the seam region is still rows 33696–33856, and continues with the objects the claim's own code had already built. Every count in the film is therefore this capture's own on today's organism, and the sealed cell is quoted separately and named as landed.
  • Recognised means the claim's stricter confirmation, not merely accepted: accepted, no collision, and the matched record is the one that photograph wrote. With the channel off, 9 of the 10 transforms recognised nothing at all — and the wrong identity was claimed 0 times, so the failure is silence rather than confabulation.
  • Filmed is one of the two landed cohorts, the ten geometric transforms and the open set split. The claim's separation variants, foreign controls, null frames and tie control are not filmed.
  • The region's rows stay unit norm, so the row norm change strip used by the other v2 films carries no signal here and is not drawn. The geometry change lives in direction, which is what the sketch trajectory shows. The two basin maps are fitted per arm, because a shared basis would spend both components on the gap between the arms; no position in one map is ever compared with a position in the other.

04 · Question to speech

The answer arrives in the field. It cannot be said.

One cold no-teach run of the canonical Q→A probe, recorded tick by tick: twenty questions encoded, 500 settle ticks, and the same cosine to the same stored answer read on two surfaces of the one field — the whole basin, and the meta.speech rows that field_commit decodes. The field reached the answer on 6 of 20 items, the speech slice on 3, and the commit organ said 0 of 20 with 0 false commits. On 3 items the answer sat in the field and not on the slice that can say it.

Why it is worth watching: this film shows the open wall on purpose, in the same typography as the working mechanisms. It is not a decoding problem — the film never proposes a better readout, because a benchmark that can be won by a smarter decoder is invalid. It is a coupling problem between two surfaces of one field, and it is measured rather than described.

Failed The relocated seam this film's credits point at is a landed honest negative, not a fix that shipped: the successor threshold issues four of four and stays switched off, and the wall on screen has not moved. The registry rows and the missing primitive are named in the fold below.

82 seconds, 1280×720. Cold protocol: --allow-cold, no --warmup, no --teach, canonical measurement pin candidate_lexeme_repair_full_2026-07-20. The followed item is probe_0007 — “WHAT IS THREE PLUS THREE ?”, stored answer SIX — chosen by a rule frozen in the capture before rendering: the highest whole-field peak among the items whose speech slice never reached the probe's own threshold. Its recorded settle path peaks at 0.554 while the speech slice stays at 0.422, and field_commit refuses with slot_cos_low at 0 decoded tokens.
Source and honesty
  • Apparatus: the landed probe scripts/diag_qa_physics_probe.py, run unmodified as its own process, with three of its own default-off read-only instruments switched on — --item-trajectory, --occupancy, --geometry-energy. With them off the run is bit-identical, so they record the run rather than change it. The capture parses the probe's JSONL afterwards and cannot reach the run it records; unauthorised writes 0.
  • This run's global into_gold mean is 0.3576, inside the landed cold band 0.350–0.378 (VERIFY.QA.NOTEACH_COLD_DEDICATED_BAND_SERIES.01, entry 2026-08-06-23: pooled mean 0.363945 over 20 processes, sample sd 0.00661, false commits 0, floor 0.35 unchanged). It is a captured re-run, not the sealed run — the band series recorded process means and kept no frames.
  • That mean is the scalar the band is built from, and the probe itself labels it secondary: structurally invalid for slice-local answers, because the whole-field number is dominated by bulk. Its primary slice-local reading on this run is mean 0.299 with 3 of 20 above 0.45. The film therefore never presents the mean as an answer rate; the answer rate on screen is the commit count, which is zero.
  • The abstain is scored as correct, because it is: honest abstain 17, missed harvest 3, false commit 0. The commit organ's own refusal reasons across the run are slot_cos_low 9, margin_low 5, no_settle_transport 3 and three witness-separator refusals. Three items cleared the reach threshold on both surfaces and were still refused, which is why the shaded corner in the film is labelled “both reached” and never “sayable” — the gate reads its own floors.
  • Two instruments of the same run are kept apart rather than blended: the walked path and the curves are the trajectory instrument's 25 recorded settle samples per item, while each dot in the whole-run panel is that item's scored peak from the probe's scoring reads (0.554 and 0.594 respectively for the followed item).
  • Not measured, and marked so in the film: no state.S sketch — the probe runs as a separate process and exports alignments, not rows, so no projection is invented; the bank-frame argmax of the live speech row was asked for and the probe filled it on 0 of 500 ticks; and the device index was not recorded on this trace.
  • Outlook in the credits, both landed rows: on a relocated seam that reads the leased cohort rows instead of the speech projection, four frozen facts issued 3 of 4 at the 0.99 uniqueness threshold and that claim stayed Failed with the missing primitive named speech_field_coupling (VERIFY.MEMORY.LEASED_COHORT_COMMIT_SURFACE.01); with a derived 0.92 threshold the successor issues 4 of 4 with 0 false and is default-OFF (…SURFACE.02, entry 2026-08-25-63). The wall in this film has not moved.

05 · Doom lite

It names what it sees on first contact, and still cannot choose a step.

One cold Doom lite episode of 32 steps, recorded step by step on both sides of the discovery cell switch. At step 1 the field mints two identities — door:yellow:locked and key:yellow:open — out of a seven by seven ego percept nobody labelled, and stages a typed transition receipt on every one of the 32 steps. With the cell open it settles three mutually distinct prediction fibers at the key pickup. Field selected actions across the whole episode: 0.

Why it is worth watching: the film holds three zeros apart instead of averaging them into one verdict. No action chosen is a missing capability, named by its own census. No law closed is a rule: the cell may propose, not conclude. And a persistent field displacement of exactly 0.0 is the amendment working as written, because a local equilibration may not write persistently.

The film of 25 August is unchanged and stays an accurate record of that capture. The wall reading moved on 31 August. With a sealed apparatus hypothesis pair and the landed default OFF wells overlay, the same ActionBus now executes five field selected actions in 32 ticks. That third arm is the film Doom lite wells below. Two named links remain open: there is still no in loop hypothesis material, and outcome receipts still bind only certified law actions.

Failed This film is cut from a run whose end to end claim is a landed honest negative, and it is here for that reason rather than in spite of it. The episode happened and the capture reproduces it row for row; what the registry refuses is the claim that the loop closed, with FIELD_SELECTED_ACTION and OUTCOME_RECEIPT named as the missing primitives. Read it as a measured wall, not as a working feature.

90 seconds, 1280x720. Instrumented re run of the landed apparatus: cli.doom_lite.run_field_episode, MiniGrid DoomLite L0 v0, level 0, seed 0, 32 steps, health 5 of 8 at the end, both arms on the same seed. CPU experiment stub, no checkpoint, no teach, mode C exploration — not an IQ measurement.
Source and honesty
  • End to end row: VERIFY.WORLD.DOOM_LITE_EPISODE_E2E.01, entry 2026-08-25-39, Failed with an honest negative — missing FIELD_SELECTED_ACTION and OUTCOME_RECEIPT. The capture reproduced that row on the closed arm: identities 2, receipts 32, field actions 0.
  • Open arm: VERIFY.WORLD.DOOM_LITE_DISCOVERY_WORKCELL_H4_HOP_BUILD.01, entry 2026-08-25-67, Verified — cell openings 2, predictions 3, non equivalent 3, laws formed 0. Reproduced exactly.
  • The action wall is named by its own claim, not guessed here: VERIFY.WORLD.FIELD_SELECTED_ACTION_FLAT_LANDSCAPE_DH_DESIGN_CENSUS.01 — NO_GO, the ΔH material is not production native. The GoalMap potential stays flat while the goal sits outside the ego window, and an exploratory step on a flat potential is explicitly not a field selected action.
  • The three fibers clear a non equivalence bar of 1e-06 with a closest pair of 1.407. That is a numerical distinctness floor, not evidence that the three hypotheses mean different things.
  • The field half shows no moving trajectory because there is none: displacement is 0.0 across all 32 steps, and unauthorized writes are 0. The cell's only effect outside its region is the typed receipt.
  • Both arms walk the same level with the same seed, so every difference on screen comes from the discovery cell. Capture traces stay off repo: /mnt/data/sdc/fi5-archive/demo-videos/captures/doom-lite.
  • Update 31 August 2026: the station census VERIFY.WORLD.DOOM_LITE_POST_ACTION_STACK_STATION_CENSUS.01 (entry 2026-08-31-25) moved the action half of the wall under a sealed apparatus pair. The numbers this film shows did not move. The third arm is a separate film.

06 · ARC-AGI-3 · a second law

It acts, then forms a new law out of what its own action changed.

One episode of the ARC-AGI-3 public demonstration environment cn04, recorded tick by tick on both sides of the switch that lets a second law form inside the episode. With the switch off the run expresses one law class. With it on the run walks the full cycle: a first law, a click that changes the world, and then a second law derived from that change, followed by a second click under the new law. Two actions, two distinct law classes, both bound to an outcome, false commits 0 (VERIFY.ARC3.CN04_IN_EPISODE_SECOND_LAW_FORMATION_BUILD.01, entry 2026-08-31-06).

Why it is worth watching: the hard part of this benchmark is not the picture, it is the argument. Five of the seven actions are bare keys, but the sixth is a click and it needs a coordinate. A coordinate invented in Python would win the frame and lose the benchmark, so the number has to come out of the field. The film shows where it comes from: the law in force is a spatial delta, the delta couples the occupancy encode to a place on the frame, the settle runs, and a plain argmax reads the cell the executor then steps. Nothing on the host picks it.

And the second law is the part that has no precedent here. G_A arrives as injected material, the way every earlier law on this surface did. G_B does not: it is derived from the transition the system itself caused, staged through the same unchanged wrap chain, and it carries its own lineage. The film puts the proof on screen rather than in prose, because it is three numbers: the wrap frame of the second law is later than the tick of the first action, its effect class address differs from the first law's, and its own harvest is what the successor claim went on to measure.

Gated overlay In episode transition formation is a contract default OFF key, skill.arc3_in_episode_transition_formation_enabled. The cycle is Verified as a build, and it is not in the default boot. Read the ON arm as an overlay under measurement, not as shipped behaviour.

78 seconds, 1280x720. Instrumented re run of the landed accept harness on cn04 seed 0, both arms in the census position the claim uses: 150 recorded ticks with formation off and 302 with it on, 450 environment frames, all field probes answering. Cold, no teach, no sealed evaluation task on screen.
Source and honesty
  • Claim VERIFY.ARC3.CN04_IN_EPISODE_SECOND_LAW_FORMATION_BUILD.01, entry 2026-08-31-06, artifact last_run.json sha256 9926eca08133…: two actions from two distinct effect class addresses, G_B wrap frame 6717588 after the action 1 tick 6717566 with its own transition lineage, G_B click {41, 29} with the world changed and the outcome bound, false 0, unauthorized 0; OFF arm one class.
  • The film is a captured re run, not the sealed run: the landed claim recorded outcome rows and kept no frames, so the episode had to be captured. The capture calls the harness's own run_arm unchanged, so census wrap, cells, tick cap and key arming are the claim's, and three read only observers are attached around it.
  • The capture reproduced the landed verdict figure for figure on today's organism: 2 law classes against 1 on the paired arm, G_B delta (0, -1), wrap frame 6717588, action 1 tick 6717566, G_B click {41, 29}, harvest VACUOUS 4 → 4, false 0, unauthorized 0. It ran on a different device than the accept run, which makes the agreement worth stating rather than assuming.
  • The frame is the observation, not a rendering of a game: the grids come from the environment through the loop's own grid_list, and the sixteen colours are the declared ARC-AGI-3 photometry the field percept is decorated with. The two field readings are drawn as marks on top and never as colour, so what the environment showed and what the field made of it do not blend.
  • The field half is read inside the pre settle encode under no_grad, on the same cadence as the frames. The capture writes nothing into the field, and unauthorized writes are 0 on both arms.
  • What the second law did not buy is filmed rather than dropped. G_B changed the world without producing a join delta, so its harvest is VACUOUS, the same class pair did not fall and the accumulator gate stays open (VERIFY.ARC3.CN04_TRANSITION_FRESH_CONSUMPTION_TARGET_BUILD.01, entry 2026-08-31-08). No third law formed (VERIFY.ARC3.CN04_TIMER_WALL_THIRD_LAW_CYCLE_CENSUS.01, entry 2026-08-31-07).
  • The environment has the last word on the episode length: cn04 seed 0 carries a timer game over near absolute step 150, so from the accept position every third click is terminal. The claim states it, this film states it, and the second cell of the ON arm runs as the honest control that reaches that wall with 0 actions.
  • cn04 belongs to the public demonstration set, which the ARC Prize Foundation itself describes as an interface rather than a valid measure. No evaluation task appears here, no score is claimed, and nothing in this film is an intelligence number.

07 · Seal integrity

The exam tries to write. The write is refused.

A write impulse minted under a scored origin arrives at the bank carrying an immutable origin stamp. The actual write refuses it, even when the consume happens later, in world tail. The same write with a teach origin is admitted. Three basin writer classes take that path. The basin tensor does not move. The production write gate stays false (VERIFY.IQ.ORIGIN_BOUND_MUTATION_FIREWALL.01, entry 2026-08-29-11).

Why it is worth watching: a skeptic understands the scene in one shot. This is not a smarter decoder and not an intelligence number. It is the seal doing what the contract says at the write site.

Seal, not IQ Every refuse and admit on screen is a live return value of the landed canary functions. The production write gate is the contract default false, read from core/memory/contract.json. No gold labels, no engine boot.

56 seconds, 1280x720. Live replay of the delayed writer canary: refuse_persistent_semantic_write and apply_licensed_basin_write, CPU, no checkpoint.
Source and honesty
  • Claim VERIFY.IQ.ORIGIN_BOUND_MUTATION_FIREWALL.01, entry 2026-08-29-11, Verified: canary refused, three basin writer classes guarded, production write gate stays false. Same functions as tests/test_origin_bound_mutation_firewall.py.
  • Sibling VERIFY.IQ.AUTHORIZE_SEMANTIC_MUTATION.01, entry 2026-08-30-06, Verified: W4 refused under scored, epoch ladder RECOVERY PREP / SCORE FREEZE / SCORED.
  • The refused write is shown with its origin stamp, not narrated. Teach origin is the matching admit. The basin tensor is unchanged after three scored writes.
  • This film is seal integrity, not intelligence. Capture traces stay off repo: /mnt/data/sdc/fi5-archive/demo-videos/captures/seal-integrity.

08 · Doom lite · wells

With a sealed apparatus pair, the ActionBus chooses five steps.

The same Doom lite episode as the 25 August film, now with a paired wells arm. Baseline field selected actions stay at zero. With the landed default OFF Delta H wells overlay and a sealed C6 stand in hypothesis pair, the unchanged ActionBus executes five field selected actions in 32 ticks (VERIFY.WORLD.DOOM_LITE_POST_ACTION_STACK_STATION_CENSUS.01, entry 2026-08-31-25).

Why it is worth watching: the counter that stayed at zero now moves, and the film still names what did not move. There is still no in loop hypothesis material. Outcome receipts still bind only certified law actions, so laws formed stay 0. This is not a two environment transfer claim.

Apparatus arm The hypothesis pair is a sealed stand in, measurement apparatus, not field produced. The 25 August Doom lite film is not recut. It remains an accurate record of a capture in which the counter stayed at zero.

49 seconds, 1280x720. Instrumented re run of cli.doom_lite.run_field_episode, MiniGrid DoomLite L0, level 0, seed 0, 32 steps, both arms on the same seed. CPU experiment stub, no checkpoint, mode C exploration, not an IQ measurement.
Source and honesty
  • Claim VERIFY.WORLD.DOOM_LITE_POST_ACTION_STACK_STATION_CENSUS.01, entry 2026-08-31-25, Verified CENSUS_COMPLETE[MOVED_VS_OLD_FIRST_STOP]: baseline 0 field selected, wells plus sealed stand in pair 5 field selected (winner_source=delta_h_wells).
  • The capture reproduced that row: closed arm 0 field actions, wells arm 5, laws formed 0, outcome receipts 0. Identities 2 on the baseline, 5 on wells because the selected steps meet more of the world.
  • Two remaining walls stay on screen: no in loop hypothesis material (the living fiber negative of 27 August, confirmed), and OutcomeReceipt binds only certified_law_action winners while laws_formed is a structural 0 in this loop.
  • No two environment transfer, no success rate, no IQ number. Capture traces stay off repo: /mnt/data/sdc/fi5-archive/demo-videos/captures/doom-lite-wells-arm.

09 · Analogical residents

The family was silent. Four of sixteen now answer.

The RecAbs analogical family on the landed paired artifact: without analogical residents the sixteen items commit 0 with false 0. With the gated overlay four items commit, four residents are licensed and minted, and twelve still abstain (VERIFY.MEMORY.RECABS_ANALOGICAL_FAMILY_GROWTH.01, entry 2026-08-30-50).

Why it is worth watching: the before and after is real and paired. The twelve that still fail stay on screen. The four lifts are named: TINY to LESS, AMBER to WAIT, ICY to FREEZE, SEALED to BLOCKED.

Gated overlay Rule 17 (b): gated, with a named reopen, not in the default boot. Never an IQ number. The film reads the landed paired artifact. It does not settle again.

47 seconds, 1280x720. Read of the landed paired rows in last_run_paired.json. CPU, no engine boot, gold labels not drawn.
Source and honesty
  • Claim VERIFY.MEMORY.RECABS_ANALOGICAL_FAMILY_GROWTH.01, entry 2026-08-30-50, Verified PAIRED_PASS: recabs off 0 of 16 to recabs on 4 of 16, false 0, 4 licensed and minted residents.
  • The capture does not settle again. Headline numbers are the landed ones. Gold labels stay in the artifact and are not drawn.
  • The overlay flag stays off in production. This page does not treat 4 of 16 as an intelligence score.
  • Capture traces stay off repo: /mnt/data/sdc/fi5-archive/demo-videos/captures/recabs-analogy.

10 · Cell panel

A law arrives. The field answers.

One World item on the production four-block path: a law, a fact, a blank go in; the field commits out; the causal microscope sits beside the commit and shows why (VERIFY.COMM.FRAME_BOUND_CAUSAL_MICROSCOPE.01, entry 2026-08-31-23).

Why it is worth watching: this is the live production path that writes DENKT, VERSUCHT, HAELT ZURUECK WEIL, FRAGT. The film reads a committed FRAME-REPORT from the official post-flip run. It does not settle again. You see the question, the field motion, two answer-slot candidates, the gates that fired, and the committed token, with the microscope tagged as display only.

Display only Verified FRAME_REPORT_LIVE. The report must not influence the commit. Lens scores are never confidence. Gold labels are not drawn.

57 seconds, 1280x720. Read of a committed FRAME-REPORT from run_20260831_162454. CPU, no engine boot, gold labels not drawn.
Source and honesty
  • Claim VERIFY.COMM.FRAME_BOUND_CAUSAL_MICROSCOPE.01, entry 2026-08-31-23, Verified FRAME_REPORT_LIVE: cpu 12 of 12 identity, display only.
  • The film reads the committed production capture. It does not settle again and does not boot an engine. Frame 35, tick 6719063, committed token WET from source leased_cohort_commit, answer-slot margin 0.588. Gate score is not confidence.
  • The official run emitted 172 FRAME-REPORT blocks. This film shows one World item. That is what the hook asked for: one sealed item, not a family grid.
  • The German four-block line was not extracted into the committed capture (the native log was not copied). It is named as absent. The microscope on that path is what the film shows.
  • Gold labels, item ids and IQ-ITEM lines stay in the source files and are not drawn. Capture traces stay off repo: /mnt/data/sdc/fi5-archive/demo-videos/captures/zellen-panel-live.

Queued, not produced

Two more experiments are coming soon, and the blockers are named.

A coming soon hook here is a missing artefact, not a teaser. No film will be cut until a real run exists on origin/main. Five hooks left this queue on 31 August 2026 for the right reason: their artefacts landed, and the films above are the result.

ARC-2 field genesis

What it should show: an ARC-AGI-2 law being induced in the field and then executed. Candidate stencils live on an owned occupancy carve, a default OFF competition energy scores coverage against soundness against description length, the settle picks a winner, and the executor paints the answer. 48 ticks of genesis, then 96 ticks of execution, ending on 729 of 729 cells exact.

Why it is not here yet: the artefact exists and the film does not. VERIFY.ARC.LOCAL_TRANSITION_LAW_FIELD_GENESIS_REPLICATION.01 (entry 2026-08-30-05) landed Verified on 30 August 2026 with all eighteen gates green: leave one out 4 of 4 exact in the field, 3 of 3 pre declared test seeds exact at 729 of 729 field produced, and both the pattern shuffle and the law ablation red. The measuring boot was a dedicated probe engine whose bank is no longer on disk, so this film needs its own capture driver rather than a parameter of an existing one. That driver is the next one to build, and until it exists there is nothing honest to encode.

The neighbouring result stays separate on purpose: the structural morphism kind series (VERIFY.ARC.STRUCTURAL_MORPHISM_KIND_SERIES_FIELD_GENESIS_BUILD.01, entry 2026-08-30-58) moved the ARC-2 appear off-seal cell to 8 of 120 on that run. Today's live off-seal law-task cell is 10 of 120. That is a different apparatus with a different population, and it will not be folded into the same frame as the 729 cell result.

Cats few shot

What it should show: five shots (teach), then a test with abstain visible. Teach in this experiment is labelled as teach. It is not a static intelligence score.

Why it is not here: the GO decision landed as commit 4dd8caa7a9 on 25 August 2026, then stopped for the missing 500 image corpus, the material target. The CLI waits on material GO. The prior wall VERIFY.VISION.CAT_GENERALIZATION_FIELD_TEST_N1_N5_N20.01 (entry 2026-08-15-52) remains NO_MEMBERSHIP_AXIS. This page will not show that wall as solved.

Continue

From a film back to the proof that earned it.

Architecture says how the loop is wired. Science says which numbers are allowed to appear. Status is the live registry.

See the architecture → Read the science → Open Status →