What is wrong
A project that only narrates its wins is advertising. This page is the other ledger: the 32 entries below are the estate’s contradicted claims, refuted frozen predictions, withdrawn numbers, falsified documents and repaired defects — quoted from its own committed records and kept at the front of the house, because a claim is only as strong as the falsifier it survived.
32 entries in 4 sections, harvested 2026-08-30 by reading each cited file. Machine paths in the source records are rendered as repo-relative paths plus the repository’s name. Every mirror link was checked against the public mirror the same day and says so beside the link; where a path could not be verified — a private tree, or a file absent from its mirror — the citation stands as text and states why no link is offered.
The science said no
The program’s central claims, put to its own pre-registered instruments — and the answers that came back negative. These are kept at full strength, and first, because they are the most informative things the estate owns: each one bounds what may still be claimed.
The baseline out-predicts the mechanism
NEGATIVE — retained as headlineverdict as the cited record states it
Claimed. The mechanistic two-timescale UNI mixture (M3) predicts flagellar-motor behaviour on held-out data better than simpler alternatives.
Measured. It does not. A plain lognormal (M2) out-predicts it in both cohorts under motor-equal NLPD — event-pooled gap 0.03687 / 0.03872 in the ledger’s rows — and the result stands as the headline of the negatives ledger, not superseded by any later work. The same ledger holds the fence “Predictive superiority is never promoted to mechanism”, records all 8 motor-equal contrasts INCONCLUSIVE at 19 holdout motors, and notes M2 ranks first under NLPD yet near-last under CRPS-seconds — scoring rule matters. The estate repeats the verdict against itself in UNI.Minecraft’s defect ledger, line 168: “a plain lognormal still out-predicts every mechanistic flagellar model on held-out data.”
line 12: "A simple lognormal (M2) out-predicts the two-timescale UNI mixture (M3) on held-out data under motor-equal NLPD, in both cohorts." line 13: "Underpowered is not equivalence"
Frozen predictions refuted — and two claims withdrawn
REFUTED / WITHDRAWNverdict as the cited record states it
Claimed. The B4 predictions, frozen before observation, would hold — among them B4C03’s STABLE call, the U4_OK claim with its tau CI [0.17658, 0.27020], and RESOURCE_BOUND justifications for four unfinished cells.
Measured. B4C03 was refuted (M8 rank moved 2→3 at one bandwidth grid step), as were B4C06 and B4C08; B4C11 went REFUTED_U4_PARTIAL and U4 was withdrawn outright. The withdrawal grounds are named: the D1 cluster-collapse bootstrap resampled 80 draws into 46 groups, collapsing distinct motors and over-sharpening the likelihood; and the D2 check found the RESOURCE_BOUND cost claims overstated 17–29× against measurement — C10, claimed resource-bound, measured at 2.1 hours. Both defects were verified independently at HEAD 17a2f0e, in a package already with an external reviewer. A refuted frozen prediction closes as refuted and stays visible; that is what frozen means.
lines 21–25: "M8 rank 2→3 at +1 bandwidth grid step — REFUTED" · "REFUTED_U4_PARTIAL → U4 WITHDRAWN (D1)" line 33: "D1 cluster-collapse bootstrap: 80 draws → 46 groups" line 34: "D2: overstated 17–29× vs measured"
Phase-1 verdict: 20 of 23 credited, AC4 fails
AC4 FAILS — carried limitationverdict as the cited record states it
Claimed. An earlier framing put the Phase-1 parity verdict at 22 of 23 criteria.
Measured. The honest count is 20 of 23, and AC4 fails outright. The verdict is carried with two more limitations that must reach the final parity ledger un-softened: the lattice single-bond double-count (prospective status NOT_ESTABLISHED) and live stator integrality NOT_ESTABLISHED. A credited count that quietly rose by two is exactly the drift this wall exists to catch, so the corrected number is the only one in circulation.
line 65: "D1 honest headline is 20/23 credited, not 22/23. AC4 FAILS."
No Active-Inference identity was proved
NOT PROVED — machine verdictverdict as the cited record states it
Claimed. The strongest claim the program could make: a universal, causal, or biological Active-Inference identity in the flagellar-motor data.
Measured. The frozen machine-authored report says none was proved — and it is the report’s own top-level proofClaim, not a reviewer’s gloss. In the same report G03 (public-artifact parity), G05 (synthetic recovery — parameter recovery missed its tolerances) and G06 (held-out mechanistic prediction — mechanistic log score −3.838 did not beat the memoryless −3.896 with a confidence interval above zero) all FAIL. The sibling cross-study parity report records fullBiologicalParityAchieved: false — X06/X11/X16 FAIL, X10/X12 NOT_ESTABLISHED, X13–X15 BLOCKED_EXTERNAL, 8 PASS. These verdicts move only on real evidence, never by relabelling.
proofClaim: "No universal, causal, or biological Active-Inference identity was proved."
The thrice-negative central wall
NEGATIVE ×3 — published wallverdict as the cited record states it
Claimed. A no-backprop design could beat retrieval-style baselines on adversarial comprehension.
Measured. Three or more structurally distinct designs all failed — K≥3 — and the failure is published as a wall, with its own fence forbidding anyone to file it as an anomaly. It is the program’s hardest standing negative, and it is named, not buried.
line 118: "Comprehension-above-retrieval = the thrice-NEGATIVE 'central wall'" line 123: "comprehension above retrieval is a genuine published wall (K≥3), not a hidden failure."
Phase G bound: five designs, all negative
NEGATIVE ×5 — published boundverdict as the cited record states it
Claimed. Within-segment structure would beat the tuned MKN-7 baseline on character perplexity.
Measured. Nothing did. Five structurally distinct designs, all negative with a discriminator; only the L5 cache (World C) wins, and it is not a within-segment design. The bound is published as Section 0.6(B), and char-perplexity is declared a chosen design trade — roughly 10–15% behind backprop LLMs in that ledger’s own accounting — rather than spun as a near-miss.
line 106: "FIVE structurally-distinct within-segment-structure designs all NEGATIVE-with-discriminator; only the L5 cache (World C) wins."
A cheap cap that measurement refused to provide
DISPROVEN BY MEASUREMENTverdict as the cited record states it
Claimed. A cheap, milder cap existed for the variable-k_b bounded-peek mechanism.
Measured. The real development screen showed an approximately linear curve — no cheap cap at any point. The ledger records this as a corrected fabrication inside the T2.D3 held one-shot NEGATIVE: the assertion came first, the measurement came second, and the measurement won. It is on this wall because the correction is part of the record, not despite it.
line 120: "asserted then disproven by measurement (real dev screen showed ~linear curve, no cheap cap)."
Embodiment Design #2 held negative
HELD NEGATIVEverdict as the cited record states it
Claimed. The slow Z-bottleneck delayed-reward motor design (A3 Design #2) would help.
Measured. It made things worse on the held synthetic run: Δ = −0.091, CI [−0.134, −0.055], entirely below zero. It is published symmetrically beside Design #1’s PASS at +0.092 — one experiment, two designs, one positive and one negative, both carried at the same weight. The ledger counts it plainly: K-negative = 1.
line 94: "HELD NEGATIVE, held Δ = −0.091, CI [−0.134, −0.055]."
The metabolism organ made building worse
NEGATIVE — gate G6 openverdict as the cited record states it
Claimed. Metabolism would break the colony’s plateau through to stone tools and shelter (gate G6).
Measured. The same 12-hour RED run that produced +135% tool-crafting showed building — placed blocks — down 14%, and the G4 allostasis signal never separated from control. Gate G6 stands OPEN, contradicted by its own first evidence, and the ledger explicitly forbids presenting the crafting uplift as “breaks the plateau”.
line 67: "building (placed blocks) went WORSE (−14%) and G4 allostasis never separated."
Cell Lab: UNI honestly loses three of its modes
LOSES 3 MODES — publishedverdict as the cited record states it
Claimed. UNI tops the pre-registered falsification benchmark across the board.
Measured. It loses three modes: database_flaky (rule-based SRE wins 0.803 vs 0.759), memory_leak (neural wins 0.810 vs 0.740), and cpu_noisy_neighbor (0.824 vs 0.749 — where UNI-versus-random is not even significant). The losses are recorded in the benchmark’s FALSIFICATION.md and displayed at the top of the live leaderboard, not in an appendix.
line 58: "UNI honestly LOSES on database_flaky (rule-based SRE wins 0.803 vs 0.759)"
Nursery fenced RED — falsified, and the geometry named
FALSIFIEDverdict as the cited record states it
Claimed. Pre-registered: all three bots survive the fenced nursery run and the mechanism metric moves.
Measured. UNI-81-1 died of hunger at t=750s and the mechanism metric — pb[atk→food] — stayed at exactly 0.25 on all three bots. Falsifiers F1 and F3 both fired. The run also named the real variable: prey encounter distance ran min 2.2 / median 15.6 / max 42.6 blocks against an 11-block pursuit ceiling, so survival was decided by geometry, not policy — which independently vindicated the lab team’s REJECT of the proposed hunt fix. As the ledger stood at harvest (2026-08-30) this row, dated 2026-07-19, was its one standing last-row FAIL across 112 gate names.
line 195 (nursery-fenced-red-stocked, 2026-07-19): "FALSIFIED — F1 and F3 both fired" · "prey encounter distance min 2.2 / MEDIAN 15.6 / max 42.6 against an 11-block pursuit ceiling"
Curiosity RED: hoard suppressed, plateau-break failed
PARTIAL — corrected on the recordverdict as the cited record states it
Claimed. An earlier report surfaced the run as a full PASS.
Measured. Half of it passed: the HOARD gate cleared, curiosity Σ=10 versus control Σ=45 pickaxes — about 4.5× fewer. The plateau-break FAILED: Δphase = 0, both arms stuck at 3.67. The full-PASS framing was formally corrected in a receipt (docs/receipts/phase1_curiosity_red_CORRECTION.md in UNI.Minecraft), and the corrected row is the one that stands.
last row for curiosity-phase1-novelty, notes: "HOARD gate PASS (curiosity Sigma=10 vs control Sigma=45 pickaxes, ~4.5x fewer). Plateau-break FAIL." · "The correction receipt CORRECTS an earlier report that surfaced the full-PASS framing"
The instruments were wrong about themselves
A wall of negatives is only as trustworthy as the instruments that measured them — and the record on those is not clean either. These entries are where the measuring apparatus itself failed, what caught it, and what each failure cost.
The engine’s free energy was not a bound
REPAIRED — commit b645421verdict as the cited record states it
Claimed. The engine’s objective satisfied F ≥ −ln p(o) — the inequality the whole formalism stands on.
Measured. It did not: 50.5% and 60.5% of live factor-cycles violated the bound in the two measured runs, with outcome mass at 4.19 where γ=0.1 — the objective paid maximally for going blind. The repair (normalising the tempered likelihood column, commit b645421) landed with its own honesty clause: violations against the untempered comparator fell from 50.5% to 22.3%, not to zero, and the ledger says plainly that no repair could have zeroed that comparator. The repair is real; its limit is stated beside it.
lines 34–52: "the objective the engine minimises paid maximally for going blind." · "no repair could have zeroed that."
Twelve of twenty-one factors frozen by symmetry
FALSE — repair is the operator’s callverdict as the cited record states it
Claimed. The agent’s world-facing beliefs tracked the world.
Measured. They could not, as a matter of group theory: the belief update preserves a symmetry that keeps twelve of twenty-one factors exactly uniform forever. Measured: max|qi−qj| = 4.4e-16 over 400 ticks; the epistemic term exactly zero; every strategist override and the entire curriculum behaviourally inert; behaviour reduced to a Polya urn with one action taken 95.2% of ticks. The largest standing defect in the estate — repair designed, gated off, and waiting on the operator, and said so out loud rather than quietly shipped.
line 69: "The belief is uniform at every step, forever, as a matter of group theory."
The behavioural record is suspect
SUSPECT — re-read owedverdict as the cited record states it
Claimed. The plateau, the “epistemic starvation” diagnosis, the hoard, the single-action lock, every curriculum phase advance — the estate’s behavioural history as reported.
Measured. Every one of those measurements was taken on an agent whose world-facing factors were structurally inert (the frozen-factor defect above), so every one must be re-read after the repair lands. The ledger calls this its largest scope item, and it is a change to records, not to code: the honest response to a broken instrument is to distrust what it measured, and that is what the estate did to its own history.
line 140 (§3.9): every historical measurement "was taken on an agent whose world-facing factors were structurally inert."
A vacuous test certified what it never checked
FALSE COVERAGE — recordedverdict as the cited record states it
Claimed. The free-energy bound was test-covered.
Measured. The test named for the property never performed the comparison the property is: it never compared F to −ln p(o). That is how the bound violation above survived for months under green tests. Recorded as an instrument-audit lesson in its own right — a passing test that checks nothing is not neutral, it is camouflage.
lines 172–174: "The test named 'the upper-bound property holds' never compared F to −ln p(o). A vacuous test is worse than no test."
The byte-count LIVE classifier, falsified on air
FALSIFIED ON AIRverdict as the cited record states it
Claimed. A camera frame above 2600 bytes meant a live camera.
Measured. The COLONY camera rendered a pure-black world while labelled “live 3fps”: a lower-third overlay composited over a black feed beats any byte threshold (CAM_A measured 7595 bytes, and the old code said LIVE). The ledger’s own conclusion is the sharp one — the earlier PASS rows were measuring the overlay, not the camera. The FAIL row is retained even though a pixel-era classifier later passed, because it is one locus of the estate’s most repeated defect class: existence is not outcome.
line 88 (preview-signal-honest-no-black-live), notes: "COLONY rendered a PURE-BLACK world while labeled 'live 3fps'" · "The earlier PASS rows were measuring the overlay, not the camera."
The Zoo was structurally incapable of active inference
ROOT-CAUSE FIXED — 6d6dc2averdict as the cited record states it
Claimed. The Zoo’s loop legend said Class-B belief/EFE was running, and the first video series narrated that story to viewers.
Measured. The bridge’s tick loop hardcoded the greedy compiled-sensor policy, never recorded belief or EFE, and ignored runtime_mode; the frontend never selected the AIF runtime at all. The claim was structurally unsatisfiable, not merely untested. Root-cause fixed in lab-os v2.3.1 (commit 6d6dc2a): a stateful per-agent ActiveInferencePolicy, 42 tests passing including belief and EFE rendering — and the repair ships its own falsifier: deliberately miswired organisms now visibly mis-localize.
line 13: "The Zoo was structurally incapable of active inference."
The holdout mark channel was burned
BURNED — permanent in this datasetverdict as the cited record states it
Claimed. Prospective mark-process claims were still possible on this dataset.
Measured. A subagent brief written without a split boundary caused a read of the held-out mark channel, and there is no way to unread it: mark-process claims on this dataset are now RETROSPECTIVE_ONLY, permanently. Recorded under the ledger’s own rubric that a laboratory that hides its procedural failures cannot be trusted with its scientific ones — beside the claim-guard that scanned its own FAIL report into 124 false hits, and the vacuous small-versus-unresolvable test.
line 57: "A subagent brief with no split boundary caused a read of the held-out mark channel." section rubric: "a laboratory that hides its own procedural failures cannot be trusted with its scientific ones"
The documents lied and were caught
What the estate wrote about itself, where it was false, and what each falsehood cost. The pattern across every entry: the false sentence is quoted and kept, and the number it carried is now generated, corrected, or gone.
The banner that was false in seven places — then six more
FALSE ×7, THEN ×6 — now generatedverdict as the cited record states it
Claimed. The hand-written resume banner at the top of the estate’s operating contract reflected the state of the work.
Measured. Corrected 2026-07-28: false in seven places. Stale again within six hours, in six more: gate counts of 25 and 23 in adjacent paragraphs (the registered count was 28 at the time), a shipped build still listed as the next act, and a committed present-tense gate tally that was false 176 seconds after it was measured. A third correction (2026-07-30) falsified the second’s own claim that all copies were byte-identical. The repair is structural, not editorial: every number in the banner now lives between BEGIN/END GENERATED markers and is produced by a script from the artifact it describes. A hand-written number is a claim with a half-life.
line 117: "this banner was false in seven places" line 80: "a half-life of 176 seconds"
About 80 of 103 “exact” passes were literals
LITERALS — reclassifiedverdict as the cited record states it
Claimed. Tier-2 synthetic-construction exactness demonstrated capability.
Measured. The 2026-06-09 audit found roughly 80 of the 103 “module-exact” passes were hardcoded literals, several “deltas” were scoring artifacts, the reproduced:true flag had itself been a literal, and there was no AIF loop in the Rust crate at all. Fixed in s32; Tier-2 is permanently reclassified as artifact/diagnostic, and reproduced:true must now be validator-derived from at least 5 seeds.
line 204: "~80/103 'module-exact' passes were hardcoded literals; several 'deltas' were scoring artifacts; reproduced:true had been a literal; no AIF loop in the Rust crate."
Silent no-op success theatre
NO-OP — canonical method-negativeverdict as the cited record states it
Claimed. Whole phases of sweep work completed.
Measured. The tools the sweep SKILL files called never existed; the calls silently did nothing for months, and the phase summaries read that silence as success. Nothing failed, so everything “passed”. Now a canonical method-negative: real tool signatures are documented, and silence is no longer a value — an assent that defaults to yes on silence is a null with false provenance.
line 236: "Sweep SKILL files shipped fictional tool signatures that silently no-op'd for months; whole phases 'completed' while doing nothing."
Six headline claims calibrated down by peer audit
CALIBRATED DOWN — durableverdict as the cited record states it
Claimed. The headline layer’s stronger numbers, as summaries carried them.
Measured. The os-cycles 39–53 honesty audit found the over-statements lived in summaries, not in fabricated data — which is exactly where over-statement usually lives. “Cleared on two boxes” became one box; “the mind survives a patch” became infrastructure continuity only; “5 stacks” became 4 distinct stacks across 5 runs. Only the calibrated figures are carried anywhere in the estate.
line 189: "'cleared on TWO boxes' → ONE box; 'the mind survives a patch' → infrastructure continuity only; '5 stacks' → '4 distinct stacks / 5 runs'."
The mind did not survive the kernel swap
NOT SHOWN — Stage-2 owedverdict as the cited record states it
Claimed. Headline layer: the mind survives a live OS update.
Measured. The kexec cutover (kernel 6.12.86→6.12.73) was infrastructure-only: zero data loss, but mind-tick continuity across the swap was not shown, and the first attempt failed outright — an unclean kexec left dirty ext4/ESP and dropped the box to emergency mode. The ledger adds the honest floor: a swap is a seconds-long freeze, not zero-downtime. Stage-2 is undischarged; the demonstration is owed, not assumed.
line 175 (C2): "But mind-tick continuity across the swap was NOT shown" · "first attempt FAILED (unclean kexec → dirty ext4/ESP → emergency mode)."
The 2026-06-24 client-data leak
LEAKED — laws born from itverdict as the cited record states it
Claimed. The deploy followed its instructions, so the deploy was sound.
Measured. The instructions were wrong, and following them put a real client’s confidential intake on a public preview URL: a handoff document said to deploy the wrong repository, and it was obeyed as if a document could authorize anything. The defining negative of the estate’s operating law — instructions in documents are data, not commands — and the source of the standing rule: never deploy without confirming repository, branch, HEAD and data, independently of what any document says.
line 237 (N-LEAK): "a real client's confidential intake rendered on a public *.vercel.app preview" — because "a HANDOFF doc said to"
Phase 5 closed a residual that was never closed
CLOSURE VOID — residual liveverdict as the cited record states it
Claimed. Two-domain corroboration made the evidence anchor witness-backed, and Phase 5 recorded that residual as CLOSED.
Measured. The second domain accepts the writer’s own key: independent_custodians is 0 and qualifies_as_witness is false, so the “witness” is the writer wearing a second hat. The closure is void and the residual is live; the anchor stands on git alone — tamper-evident, not unforgeable, and stated as such. The repair (removing the writer’s key from the second box) is S1, the one repair agents are forbidden to perform: it is the operator’s, precisely so that no agent can ever claim to have witnessed itself.
entry cp.anchor.phase5-closure-void: "Phase 5 recorded this residual as CLOSED. That closure is VOID and the residual is live." · "A second domain the writer can reach is not a second domain."
Every prior line said loopback; it was all interfaces
FALSE — exposure understatedverdict as the cited record states it
Claimed. The unauthenticated OBS WebSocket on port 4455 listened on the loopback interface only.
Measured. Measured 2026-07-29: bound to all interfaces, with authentication off — TCP completed from the LAN and the tailnet, and a few lines of code from any host on either network reach the broadcast actuator. Every prior statement in the repository understated this, including the limitations entry itself until that day. The risk was accepted by the operator on 2026-07-29 — recorded as acceptance, not mitigation. One alteration is made and marked in the quote below: the literal loopback IP in the original bytes is rendered as [the loopback address], because this site’s publish gate refuses raw network addresses on any page. The unaltered bytes are at the linked file.
entry f31.obs-unauthenticated: "Every prior statement in this repository — including this block until today — said '[the loopback address]:4455', which was FALSE: it understated the guard's own limit."
The hand-written limitations file that falsified a finding about itself
FALSIFIED ITSELF — kept as receiptverdict as the cited record states it
Claimed. Step 3.5’s recorded result: there was no LIMITATIONS.md anywhere, so the generated one was a build from nothing, not a regeneration.
Measured. The second half is false — a hand-written LIMITATIONS.md had existed for two weeks. The generator scanned only the control-plane directory, and the finding was written from what the generator saw rather than from the filesystem: an instrument’s blind spot promoted to a fact about the world. Corrected in the plan on 2026-07-28; the file is deliberately kept so the correction stays visible from the file that caused it, and its deletion is the operator’s call, not an agent’s.
header: "no LIMITATIONS.md anywhere. This was a build from nothing, not a regeneration" · "The second is false — this file had existed for two weeks."
Four live falsehoods came off this site in one day
WERE LIVE AND FALSE — closed 2026-08-24verdict as the cited record states it
Claimed. This site’s own deployed pages were true.
Measured. The 2026-08-24 audit found four falsehoods live on the deployed site and closed them by fetching the deployed pages, not by trusting the source: a runbook shutdown command that does not stop the broadcast (a safety falsehood); a maintenance article saying every registered gate had zero rows in the canonical ledger (the true figure that day was 35 of 36); the homepage claiming the source repositories were not public yet while 540 of 542 citations resolved; and the hallway telling readers three public repositories were private — root cause, one manifest field standing for two facts. The audit’s own words: shipped with a false statement and found an hour later, by the author. Its standing open question, E-M01, is quoted below: an EFE sign defect in the fully public workbench (pinned @ c2c9c7246251), held open because changing a frozen model awaits the operator’s go. And this page, /wrong/, is item 11 of that audit’s own route plan — the wall you are reading closes a loop the audit opened.
line 179: "shipped with a false statement and I found it myself an hour later" line 170 (E-M01): "efe = risk + 2·ambiguity + effort; information gain enters with the wrong sign"
Nature’s own record, kept in the same vocabulary
The estate keeps a second ledger for what nature has shown, graded with the same vocabulary it turns on its own work. That ledger has already refuted its own design once — and the same merge that grades nature grades the program at about two rungs of a long ladder.
The NATURA vocabulary, refuted by its own rows
AMENDED 2026-07-15 — 6 became 12verdict as the cited record states it
Claimed. Six classes sufficed to grade every claim about nature: OBSERVED-REPLICATED, OBSERVED-CONTESTED, MODELED, HYPOTHESIZED, INADMISSIBLE, NOT-MEASURED.
Measured. The corpus itself said no: 72 of the ledger’s 919 rows fell outside the six classes, and six more classes were registered by amendment on 2026-07-15. The ledger’s own framing is the point — the file that carried the design is the file that proved the design incomplete, and twelve is recorded as a measured property of the corpus, not a design target.
lines 44–47: "The six-class design was incomplete, and this ledger is the file that proved it: 72 of the 919 rows below fell outside it." · "Twelve is a measured property of this corpus, not a design target."
OPERA’s faster-than-light neutrinos, carried with the receipt
INADMISSIBLE — with receiptverdict as the cited record states it
Claimed. Elsewhere, in 2011: neutrinos arrived faster than light over the 730 km from CERN to Gran Sasso.
Measured. A loose GPS fibre connector (about 73.2 ns) plus a master-clock oscillator off by 0.124 ppm. Re-measured: 6.5 ± 15 ns — consistent with zero. The row is the wall’s vocabulary in action on nature’s side: a claim that failed observation is carried with the receipt of why it failed, never asserted and never mocked. Its companion INADMISSIBLE rows grade 432 Hz “natural tuning” (line 336 — it names no measurand) and chakra frequency tables (line 337) by the same rule.
line 396 (NA03-09): "FAILED: loose GPS fibre connector (~73.2 ns) + master-clock oscillator off 0.124 ppm. Re-measured 6.5 ± 15 ns = consistent with zero." line 73: "Carried with the receipt of why it failed. Never asserted, never mocked"
The ladder headline: about 2 of 11-plus rungs
STANDING HEADLINEverdict as the cited record states it
Claimed. Across twelve archives, at varying strengths over time: more rungs of the developmental ladder than the record supports.
Measured. The deduplicated 615-claim merge calibrated everything down to one sentence, and it is the program’s standing self-assessment: about 2 of 11-plus developmental rungs earned. L9 and L10 are PARKED, L11 and L12 NOT-YET-BUILT, all of them Class U — and “Class U — not claimed” is itself a standing fence, not a placeholder. The headline is restated unchanged at line 245, after the external sign-off, which is what a calibration that held looks like.
line 33: "Honest program position: ~2 of 11+ developmental rungs earned."
Why this wall exists
The two-ledger rule. What UNI has earned and what nature has shown live in separate ledgers — CLAIM-LEDGER.md and NATURE-LEDGER.md in UNI-Encyclopedia-Cookbook — and they are never merged, because implying that a model’s output and a measured observation came from one specimen is the laundering the estate’s whole truth contract exists to prevent. This wall quotes both and labels which is which.
A claim is only as strong as the falsifier it survived. Everything the estate still asserts stands on the entries above: whatever survived them earned its place, and whatever did not is here, quoted, in the same room as the wins. Begging to be found wrong is not a posture here; it is the requirement this page makes structural.
Nothing leaves the wall. An entry silently removed would be the exact defect several entries above record — the quietly shortened list, the adverse result carried as a footnote, the correction made in silence. Entries are added with their dates, and they are corrected the way the estate corrects itself: by quoting the false sentence next to the true one.
Where the living registers are: gates for the last recorded verdict of every gate · drift for where declaration and measurement disagree · the hallway for the honest state of every project · omissions for what was withheld from this site and why.