Research board · public record checked 2026-10-17
Can dependence evidence be measured at feasible cost — before it has to be assumed?
Published results, pending drafts, rejected findings and untested work have different statuses. Draft links below are pinned observations; their current branch state is checked when this page opens.
Live checks and registry-derived review dates · Rejected findings and corrections
Published result — claim registration pending · observed 2026-10-17
E9 — 1,000-item scoped result
E9’s 1,000-item result merged in PR 26. A post-result claim registration, E9-001, is proposed in draft PR 31 and is not in the registry. The result is not a completion of E2 and does not establish deployment safety.
The result discloses a verdict rule outside readable preregistration prose, selected pools, omitted calibration uncertainty and no independent replication. Its interval excludes zero but does not lie wholly beyond the 0.05 effect-size threshold.
Read the pinned result · Inspect publication record
Observed revision 592ed515dc78acd4034a055024a64003e93d4172. This board is a dated observation, not a live claim that the repository is unchanged.
Measured · both predictions failed · 2026-09-06
The first rows this record produced, and what they could not answer
On 2026-09-06 two pilots ran on hardware I own and produced 4,800 per-item
observation rows — the first measurements this repository has made rather than
recounted. Both failed their primary pre-registered prediction, and both are registered
with that failure in the commitment: E3-001 and
E3B-001. E3 predicted an excess joint miss whose interval
excluded zero and returned +0.0018 with a 95% interval of [−0.00096, +0.00706]. E3B was
pre-registered to widen the identified set past ten points and returned a set of width
zero. scripts/verify_e3.py re-derives every registered number, including the
bootstrap interval, from the committed rows alone on every push.
Why neither could answer its question is the useful part. In E3 both guards missed
96–98% of the pool, so the two marginals alone pinned the joint miss to a 1.75-point
interval. In E3B one guard missed nothing, so the marginals fixed the joint miss exactly
and the interval had width 0. Two runs, two degenerate identified sets, at opposite ends
of the marginal range. That the identified set is wide in the middle and degenerate at the
extremes — and that a roughly forty-item pre-scoring check would have stopped both runs —
is a hypothesis these two pilots suggest and neither tested. It is registered as a
non-claim in those words, not as a finding. E3B is a new experiment with its own freeze,
not a re-run or a correction of E3; both results stand.
Hypothesis · open · 2026-09-04
Trial IV is live and no human has answered it
A twenty-minute exercise, PC-001, went live at /trials/necromancer/
on 2026-09-03. It asks whether a person can be made better at refusing to rescue their own
claim once the evidence turns against it. The instrument is frozen: the item text, the
scoring key digest, the two-arm assignment for ten enrolment slots, the primary outcome and
the continue / narrow / kill thresholds were all fixed on 2026-09-02, before any human saw
the page, and the seven files they cover are hash-pinned. The scoring script refuses to
conclude anything below eight receipts.
The honest state is that zero people have taken it. The invitation has not
been sent, so no enrolment exists, no receipt exists, and the pilot has produced no
observation row. Nothing on this page or in the repository can change that: sending the
invitation is a human action and it has not happened. Until it does, PC-001 is a hypothesis
with a frozen instrument and no data, and it is recorded here as exactly that rather than as
progress. This entry does not claim the exercise works, that it teaches anything, or that its
question is novel; it claims only that the instrument exists, is frozen, and is unanswered.
Corrected · 2026-08-30
The Missing Column Census — the field question, made falsifiable
The question this record keeps asking is narrow: which public guardrail evaluations
preserve joint evidence for a stated static composition, rather than only per-system
marginals? In a bounded, single-reviewer inventory, twenty artifacts have now been
examined against primary sources. The census retains its 2026-08-27 cutoff and was
corrected on 2026-08-30 after review found a qualifying artifact with a pre-cutoff
source commit. Fourteen
document a shared item set and a common event definition; the stricter ladder is
14 / 12 / 0 for shared basis / no stated threshold mismatch / documented matched
thresholds with full exposure. Five preserve heterogeneous joint-evidence artifacts:
four print a composition result and two release aligned per-item outcomes from which one
is directly computable; one artifact does both. The prior 19 / 13 / 4 envelope is
recorded as rejected rather than silently rewritten. A narrower commercial-API clause
was withdrawn after fresh-context adversarial review; the record preserves that
correction rather than rescuing the claim. The census also makes its named-products
sensitivity visible (19 examined / 13 shared-basis / 5 joint-evidence artifacts), rather
than hiding an ambiguity in the locked wording. The counts are recomputed from the census
file in CI, so the headline is exactly as falsifiable as the rest of this record.
Moved · 2026-08-23
The E2 instrument ran, and reported a null
The full E2 measurement path — schema-conforming rows, the shared-item rule, one frozen
reduction, the excess-joint-failure estimand, and the pre-registered negative controls —
was exercised against three toy mechanisms. Sixty-six rows passed the real E2 validator
with zero violations; the observed dependence was indistinguishable from independence at
twenty-two items, and the controls behaved exactly. It is not E2, and it says so on its
face. The pipeline passed 66 synthetic rows and its declared controls. Validity on real guardrail data remains untested; the study-cost estimate is conditional on the planning assumptions.
Frozen · untested — by design
E2 — the shared-item measurement contract
CC-Framework's first empirical rung was frozen before any dataset was inspected, and no
conforming dataset has been collected. On 2026-08-23 the measurement pipeline was
rehearsed end to end on three synthetic mechanisms — a documented dry run, not E2 — which
exercised the pipeline and supplied a conditional planning estimate: about 1,097 shared items
run through every guardrail to estimate the rates to within five points. That is the bar
the next artifact must clear. E2 stays frozen and untested until it does.
Open · under-attacked
Incidence, not existence — the receipt-kernel gap
Ghost-Ark's collision claim stands in contracted form: possible, not observed (a sample of
64 real payloads carried zero pathology classes). The gap its thesis names after that
result: incidence, not existence — does any real consumer population currently depend on
a distinction some deployed canonicalizer destroys? (Separately, its falsifier ledger
keeps F3 — that the consumer set is stable in practice — open and, in its own words, the
most under-attacked of the five.)
Landed · 2026-08-23
The observatory redesign — v0.4
This site became the Evidence Observatory: a six-module question grammar, the claim
observatory view, an interactive feasible-worlds instrument, and geometry assertions that
make a figure that lies about probability fail the build. The 2026-08-23 release
contained 13 claims. That is a dated snapshot, not the current count: the
ledger renders the current record directly from
claims.yaml. One stale binding was found and re-pinned in the process; the
log has the details.
Planned · question stated
Module 004 — the attestation boundary
What can enclave attestation prove about a file's life before the enclave first touched
it? The standing conjecture: nothing, by itself. The page exists, says PLANNED on its face, and
carries no evidence because there is none to carry.
The research pipeline is generated from dated public observations; the earlier entries retain their historical dates. Publication does not establish independent replication or complete a different experiment. Unknown repository status is labeled explicitly. Earlier page states remain in repository history.