Evidence ledger
Generated from agents/shared/decisions.md, agents/shared/findings.md and threads.md by build_ledger.py. Regenerated rather than edited.
The paper argues a claim and the narrative tells the story. This is the complete record behind both — including the entries that went nowhere, which are the majority.
41 decisions · 22 findings · 67 threads (1 active, 52 answered, 3 dropped, 3 open, 8 parked)
Decisions and results
Engineering decisions and experiment outcomes. Each states its falsification before the run, and corrections keep their own entry rather than editing the original.
E-002 · grounds are evidential, not deductive — AGM's ordering half, not its closure half
Decision · 2026-09-10 — Decision: Model the edit-to-grounds relation as evidential support, not entailment. orphan becomes a graded quantity: the joint implausibility an edit…
O-004 · raise the model to GPT-J-6B via NDIF; hand-build deductive grounds for the pilot
Decision · 2026-09-10 — Decision: Amend CLAUDE.md's blanket size cap. The pilot moves from GPT2-medium to GPT-J-6B via NDIF, and the pilot's ground sets are hand-built and…
E-007 · candidates come from a reference vocabulary, not the edit set
Decision · 2026-09-13 — Decision: candidatepool takes an explicit reference set, separate from the edits being filtered, and every unscoreable item is reported in skipped with a…
O-005 · large models only — the size floor replaces the size ceiling
Decision · 2026-09-15 — Decision: work only with very large models. From NDIF's ungated set that is Llama-3.1-70B (possession 79%) and Llama-3.1-405B (85%). GPT-J-6B and the 8B…
E-009 · the gate clears at 70B, and the three legs are not independent
Result · 2026-09-15 — Decision: the edit arm runs on the 70B set. 74 usable chains is enough for an existence claim with room for edit-stage attrition; 20 was not. [O-005]'s…
E-009b · the outer deficit is answer surface form, not pool concentration
Correction · 2026-09-15 — RCA. [E-009] closed with "outer stays the worst leg at both scales, which rules out small-model thinness and leaves candidate-pool concentration —…
E-011 · the possession measure has no operating point for modal answers
Result · 2026-09-15 — The pre-stated Deny branch fired, and it was the wrong dichotomy. [E-009b] predicted that rendering countries naturally would recover the article-taking…
O-006 · the edit arm runs at Llama-3.1-8B
Decision · 2026-09-15 — The gate at three scales, same 136 chains, same pools, same seed: Restricted to the domain [E-011] shows the measure is defined on — the outer answer is…
E-012 · the two defects are orthogonal, and fixing them beats 60B parameters
Result · 2026-09-15 — Usable chains, same 136 chains, same model, same seed: Neither fix alone moves anything: +1 and +1. Together, +20. This is an interaction, not two…
E-013 · the covariance term is not used by the reference implementation
Gate 0 resolved · 2026-09-15 — The problem as stated. ROME's update needs C⁻¹k, the inverse second-moment matrix of keys. Measured from the model configs: Llama-3.1-8B dmlp = 14336, so…
E-013 · the mandated control is too weak, and this is why
Smoke test, n=1 · 2026-09-15 — Vlaminck: born in Paris, Paris in France, therefore born in France. Edited the conclusion to Germany. v optimisation drove NLL 5.68 → 0.02 over 25 steps.…
E-014 · the edit displaces the premise, country-specifically, at 67%
Result · 2026-09-17 — Written late. This result was referenced by [E-015], [E-016] and the published paper for three days without an entry of its own — found when cross-linking…
E-015 · DENY — the relocation carries no inference
Result · 2026-09-18 — Resolves agents/shared/disputes.md [E-014]. Both arms push the same country-valued vector through the same subject's key; only the birth edit licenses…
E-016 · whitening is a near-no-op by cosine and a 50x change in targeting
Gate · 2026-09-18 — distribution (NOT ROME's Wikipedia — divergence recorded) Anisotropy — the gate clears. Spectrum of the layer-5 key second moment: The keys live on an…
E-016 · whitening cannot reduce same-subject leakage, and the reason is analytic
Result · 2026-09-18 — The experiment. Re-ran [E-015]'s paired test with u = C⁻¹k, where C is the low-rank-plus-ridge key second moment (Woodbury, N=2048, CounterFact prompt…
E-017 · same-subject leakage is structural — no probe form escapes it
Result · 2026-09-19 — Tests [T-074], the threat [E-016] left standing: its analytic account required the probe to share the edit prompt's prefix, which our probes do by…
O-007 · four "nnsight constraints" were one flaky NDIF node
Correction · 2026-09-20 — What I claimed. A commit on 2026-09-20 documented four constraints discovered while building the [T-075] layer sweep: that a list comprehension inside a…
T-075 · the pinning is exact for prefix-sharing probes at every depth, and
Result · 2026-09-20 — Mean coefficient at the subject's last token against the edit's k (1.0 = full delta): Worst case across subjects is sharper still: the possessive form…
E-018 · the different-subject floor is 0.082 — [E-017] stands, and the floor is now measured
Result · 2026-09-20 — Tests [T-076], the control [E-017] never ran. Every prior control held the subject fixed and varied the form; this one varies the subject and holds the…
O-008 · one of [O-007]'s four retracted constraints is real
Correction to a correction · 2026-09-20 — What [O-007] said. Four "nnsight constraints" documented while building the [T-075] sweep were all wrong, and all four were the same thing: NDIF rejects…
E-019 · the analytic half is layer-invariant; the empirical half is an early-layer fact
Result · 2026-09-20 — Runs [T-075] per design.md Part IV. DENY in the mid-deep range and NULL at layer 31 — both pre-stated outcomes fired, at different depths, which is…
E-020 · the decay is direction, not magnitude — and layer 31 is convergence, not noise
Result · 2026-09-20 — Runs [T-077]. [E-019] measured the decay and explained nothing about it. The coefficient is not a cosine — c = (|k|/|k|) · cos(k, k) — so it can fall two…
E-021 · the subject-token delta is sufficient and necessary — the mechanism is causal
Result · 2026-09-20 — Runs [T-067], the hole §6 of the paper names first. Everything from [E-013] to [E-020] is behavioural — vary an input, read an output. [E-016] showed the…
E-022 · possession IS a fact-level property — and MUTE has two causes, not one
Result · 2026-09-20 — Runs [T-062] and [T-064] from one measurement. [T-062] — CONFIRM, and it refutes the expectation this ticket was opened on. Modal-cell share per fact,…
E-023 · the closed pool undercounts relocation badly — and the Washington attractor was ours
Result · 2026-09-20 — Runs [T-066]. Every number in this project comes from ranking a closed type-matched pool. This asks what that hides by letting the edited model generate…
E-024 · an edit reaches the whole subject, but damages in proportion to type overlap
Result · 2026-09-20 — Tests whether a birthplace edit disturbs attributes that have nothing to do with birth. Two probes give a relatedness gradient; the control is the same…
E-025 · a floor plus type-matched displacement — not a relatedness gradient
Result · 2026-09-20 — Separates the two accounts [E-024] could not: language is semantically related to country of origin but typed as a language, so it shares relatedness with…
E-027 · the pinning is a ROME fact — and the different-subject floor is not
Result · 2026-09-20 — Runs [T-080]. Every coefficient in this project had been measured on one model. Worst deviation from 1.000 across all 216 prefix-sharing cells: 0.00e+00.…
E-026 · the experiment failed, its gate was mis-specified, and one existence claim survives
Result · 2026-09-20 — Runs [T-079]. The pre-stated question is UNANSWERED and the first headline this run produced was wrong. Both are recorded before the surviving finding,…
E-029 · the ES caveat does not exist at the standard edit layer — [T-085] closes
Result · 2026-09-20 — Runs [T-085]. [E-026] left an observation its failure did not touch — edits passing ROME's Efficacy Score while leaving the belief standing and…
E-028a · T-084 unanswered — and target lift is worthless as a measure of edit strength
Result · 2026-09-20 — Runs [T-084]: is [E-026]'s weak layer-20 edit under-optimisation, or depth? Grid and bar were fixed in the source before the run and every cell is…
E-028c · T-084 answered — and equal delivery still does not give equal effect
Result · 2026-09-21 — Target was 6.26 nats — ROME/EasyEdit at layer 5 with 25 steps. Layer 20 crosses it at ~353 steps, and the held-out six confirm the budget transfers (gap…
E-028d · it is the LAYER — and the coefficient does not predict effect
Result · 2026-09-21 — The fourth cell, bought to break the collinearity [E-028c] built in. The coefficient at the subject's last token is exactly 1.000 in every cell below —…
E-030 · the GPT-J floor is architectural — [E-027]'s narrowing is confirmed
Result · 2026-09-27 — Runs [T-082]. §4.3 of the paper narrows [E-018]'s "different subjects are near-orthogonal at the subject position" to a Llama-3.1-8B statement, on the…
E-031 · NO DIP — and the observation dies for a better reason than significance
Result · 2026-09-27 — Runs [T-083]. [E-027] noticed both models' different-subject floors dipping near depth fraction 0.15 — roughly where ROME edits in both — on 4 and 6…
E-032 · the pinning is exact from 6B to 405B — the floor is not a scale law
Result · 2026-09-27 — Runs [T-080] at scale. Coefficient measurement needs no gradients, so this cost ~24 traces where an edit at 405B would have cost days. Four models, a 68×…
E-033 · editing at scale was blocked by a one-line device bug, not by NDIF
Decision · 2026-09-27 — Decision: editing experiments may now run at 70B and 405B. The blocker was ours. computevbatch's loss accumulated weightdecay (Dd[i] Dd[i]).sum() —…
E-033 · type matching is the mechanism — the "floor" was one pairing
Interim · 2026-09-27 — Runs [T-081], the third and last of [T-065]'s never-varied dimensions. Recorded as interim: the 8B birthplace replication and the 70B occupation arm at…
E-033 · all six arms — type matching is the mechanism, the floor is one cell
Final · 2026-10-03 — All six arms complete. The interim conclusion holds and sharpens. Type matching: 6 of 6, no exceptions. Across two models, two relations and two depths,…
E-034 · the loop owns execution, a person owns judgement
Decision · 2026-10-14 — Decision: an unattended runner for this project, in four pieces — preflight, triage, report, watch — with a hard boundary at judgement. It owns: running…
E-036 · UNDIFFERENTIATED — and two of the six criteria could not have answered
Result · 2026-10-14 — First end-to-end run under the harness: queued, preflighted in 15s, run, triaged, reported, no intervention. Zero retries. The verdict fired as…
Findings
Literature reads and measurements that are not decisions. Two of these closed whole directions at the prior-art gate.
R-001 · RippleEdits' Logical Generalization is NOT a backward probe — but claim 1 must be narrowed
2026-09-10 — Verified metadata (session note §9 required this). Cohen, Biran, Yoran, Globerson, Geva, "Evaluating the Ripple Effects of Knowledge Editing in Language…
R-002 · Lens 2 clears — not scooped, but must be repositioned against knowledge-conflict and deferral
2026-09-10 — Verdict: PROCEED. No existing work partitions an edit's affected knowledge by whether logic determines the outcome and surfaces the underdetermined part…
R-005a · mined-rule artifact IS available — but on DBpedia and MQuAKE/MLaKE, not Wikidata/CounterFact
2026-09-10 — Gate result: §0 method (e) is FEASIBLE. The critical-path dependency clears. Two corrections to what design.md v0.5 asserted, both material. the…
R-003 · 35.4% of CounterFact edits use a rigid relation — T-023's gate passes
2026-09-10 — The fear was unfounded. T-023 worried that CounterFact is so dominated by mutable relations that a naive 50-edit sample would draw almost entirely from…
E-001 · §0 method (e) FAILS — the mined rules are alias tautologies
2026-09-10 — Verdict: method (e) does not work on the shipped artifact, and the failure looks structural rather than incidental. Fall back per design.md §0.…
E-002 · Wikidata carries real grounds — but they are evidential, not deductive
2026-09-10 — Verdict: method (e) is alive on Wikidata. Steps 1 and 2 pass. But what we found are not AGM kernels, and that has consequences. 55 rigid-relation edits, 5…
E-003 · possession is real, scale fixes it — and the field's standard test overstates it
2026-09-10 — Verdict: [T-039] confirmed and [O-004] empirically vindicated. Possession is a genuine gate, GPT-2 fails it, and a large model clears it. Scaling within…
E-003b · GPT-J-6B possession is 73% — usable as the edit target, with filtering
2026-09-10 — Answers [T-045]. The hard fork is avoided: we do not need ROME on a 70B model. GPT-J sits between the two: +11pp over GPT-2, -20pp under Llama. So it is…
E-004 · joint possession is 69% — the pilot has a pool, but the ground number is confounded
2026-09-10 — Verdict: the gate passes. Joint possession — head held and at least one ground held, same subject, both top-1 by constrained rank — is 61/89 = 69%,…
E-005 · possession re-measured — 56/74/79/85 across 6B-405B, and scale buys only the hard relations
2026-09-11 — Replaces the superseded E-003/E-003b/E-004 numbers. Possession = the true answer ranks first among 50 type-matched candidates and ranks higher with the…
E-006 · ROME on Llama-3.1-8B via NDIF is feasible — both required primitives work
2026-09-11 — Verdict: option 1 of [T-050] is available. The blocking concern was that NDIF hosts one shared copy of each model, so a weight edit cannot be persisted.…
R-006 · the diagnostic framing is FALSE — CounterFact does verify possession, with a weak test
2026-09-11 — Verdict: drop "editing benchmarks measure propagation into knowledge they never verified was there." It is not true, and a reviewer would refute it from…
R-007 · 2505.18690 does not take T-054 — and hands us one adversary point
2026-09-11 — Verdict: the nearest competitor does not claim our result. T-054 stands. "Benchmarking and Rethinking Knowledge Editing for Large Language Models" — He,…
T-055 · coordination lift is a set-level type diagnostic, not a per-candidate filter
2026-09-11 — Motivation. The possession measure assumes its distractors are type-matched; that assumption was never checked. Coordination should test it — same-type…
D-002 · CounterFact has NO possession filter — R-006 was wrong, verified from the primary source
2026-09-14 — This reverses [R-006]. I concluded there that "CounterFact does verify possession, with a weak test", from §3.3's sentence about counterfactuals starting…
R-010 · the WHY gate narrows sharply — two of three proposed contributions are taken
2026-09-18 — Run before building the probing taxonomy the project was about to pivot to. Three prior-art hits, two of them direct. Holtzman, West, Shwartz, Choi,…
R-009 · two ROMEs circulate, and they differ on the term the method is built around
2026-09-20 — Opened by [E-013] gate 0, which found EasyEdit ships mom2adjustment: false in all fourteen of its ROME configs — including gpt2-xl and gpt-j-6B, the two…
T-065 · pairs bound, crossings decompose — and narrowing comes from unfixing a dimension
2026-09-20 — Tests the thread's two structural claims against the record rather than reasoning about them. Every result in decisions.md hand-classified by the shape of…
R-011 · the depth result has prior art, and T-086 has a candidate
2026-10-03 — Re-ran prior art against where the project now is rather than where it started. Four entries added to the reading list, each verified field-by-field…
R-011b · the project's founding citations all resolve — two were described wrong
2026-10-03 — CLAUDE.md has carried eight arXiv identifiers since day one under its own warning: Seven of the eight had never been checked and were absent from the…
R-012 · reading the two papers moved both claims we made from their abstracts
2026-10-03 — [R-011] put two papers into §2 on the strength of their abstracts. [R-012] read them. Both pre-registered falsification conditions fired, in opposite…
R-013 · RippleEdits has NO subject-initial prompts — §7's premise is false
2026-10-14 — §7 item 2 of the write-up — the project's one outward-facing consequence, priced at a week and unrun since September — rests on this sentence: It is false…
Confidence levels and full evidence are in the repository entries; these are one-line pointers, not summaries.
Threads
Open questions, tracked as a tree. A thread is a unit of inquiry; a ticket is a unit of work. Parked is not dropped — a parked thread carries enough context to resume cold.
Open (3)
| id | question | parent | status |
|---|---|---|---|
T-078 | What reorients the subject key with depth? — [E-020] ruled out magnitude and named attention mixing as the leading candidate without testing it. The cheapest cut is the one… | T-077 | open |
T-085 | What does ROME's Efficacy Score fail to distinguish? — The one thing [E-026] produced that its failure does not touch. ROME's ES is P(target) > P(true) on the edit prompt. At layer… | T-079 | open |
T-086 | What actually sets how far a layer-5 edit travels? — [E-029]'s by-product, and larger than the question it was answering. Belief displacement on the edit prompt and ES margin each… | T-085 | open |
Active (1)
| id | question | parent | status |
|---|---|---|---|
T-054 | The CounterFact filter outlived the model it was calibrated against — [R-006] established that CounterFact filtered records on P(true) > P(counterfactual) pre-edit — so possession is checked. But… | T-047 | ACTIVE |
Answered (52)
| id | question | parent | status |
|---|---|---|---|
T-001 | Should edit-slice merge into rome-neighbors, and in which direction? — rome-neighbors is now KEEP (a parametric<->retrieval consistency certifier — a method, aimed at the RAG seam). edit-slice… | — (root) | answered |
T-002 | Does Cohen et al. (TACL 2024) Logical Generalization already constitute a backward probe? — RippleEdits' Logical Generalization test type covers inverse and symmetric relations. If that is already probing grounds rather… | T-001 | answered |
T-004 | Does the backward-probe pilot survive the ten design lenses? — No implement ticket may open until it passes. The WHY gate is the stop condition, and lens 2 (prior art) is where T-002 lands —… | T-001 | answered |
T-005 | What exactly may edit-slice import from rome-neighbors without importing its scope? — T-001 permits code reuse but not scope merge. Where is the line? ripplekit is ~490 lines (config, data, reps, predictors,… | T-001 | answered |
T-006 | Is the grounds relation decidable in the CounterFact/Wikidata setting at all? — Declaration 5 defines grounds as the facts that were premises for the edited fact. In code, the compiler decides that. In… | T-002 | answered |
T-007 | Is forward/backward really expansion/contraction? — Grounds are conjunctive (a conclusion needs all its premises); consequences are not. So editing a conclusion leaves an intact… | T-002 | answered |
T-008 | Does the backward direction terminate without k? — Forward closure is infinite, which is why declaration 4 needs an explicit k. But a fact has a small finite set of minimal… | T-007 | answered |
T-009 | Kernel count as the backward out-degree variable — Over-determination — a fact with multiple independent justifications — is what makes contraction fail, since expansion needs one… | T-007 | answered |
T-010 | Metric sign error — should grounds fall or hold? — CLAUDE.md says of grounds probes "log-prob of pre-edit correct answer should not fall," treating grounds as bystanders. Under the… | T-007 | answered |
T-013 | Do kernels survive transfer to a graded setting? — Kernels are a logical notion with sharp membership. In a probabilistic system support is graded — everything weakly moves… | T-012 | answered |
T-014 | Salvage from the cut material — §11 cut the self-reinforcing mechanism and backward-pass sync as method design — correct. But the motivating intuition (p. 12:… | T-007 | answered |
T-031 | Inputs/outputs: the grounds test must be directed — A fact is a node with inputs (justifications, the a priori) and outputs (consequences). A one-sided intervention detects a link,… | T-012 | answered |
T-032 | In-degree and out-degree are different variables — We have been conflating two branching factors. | T-031 | answered |
T-033 | The oracle is the unedited model's own in-context reasoning — Direction is established in-context on the unedited model; orphaning is measured parametrically on the edited model. Different… | T-031 | answered |
T-034 | Show structure, never ranking — In-degree and input-depth are informative about which ground is least a priori — but using them to select what to retract is… | T-025 | answered |
T-035 | Mined Horn rules read backwards give the kernels — 2606.10554 mines body => head rules from Wikidata with AMIE and reads them forward (edit touches body, does head update?).… | T-012 | answered |
T-036 | Probe-generation-from-a-graph is not ours — session-2026-09-08 §5 argued "read questions off the oracle — the graph is the question generator." AMIE-mined rules over… | T-035 | answered |
T-023 | CounterFact relation inventory — rigid vs mutable — Orphaning requires the edited relation to be rigid over time. If CounterFact is dominated by mutable relations, a naive sample… | T-010 | answered |
T-037 | Time-rigidity blocks relocation only — it never establishes contradiction — The rigid/mutable 2x2 asserted that a rigid edited relation forces a contradiction with its grounds. Does it? | T-013 | answered |
T-038 | The termination argument is partly given back — [T-008] argued backward needs no k because kernels are finite while forward closure is infinite — the asymmetry that made the… | T-013 | answered |
T-039 | Possession — does the MODEL hold the ground? — Wikidata says Perec was born in Paris. Does GPT-2? A ground the model never held cannot be orphaned — nothing is left standing to… | T-012 | answered |
T-041 | Does contraction EVER occur, on hand-built deductive grounds? — E-002 found grounds evidential as Wikidata expresses them — a fact about the schema, not about knowledge. Hand-build small… | T-013 | answered — no |
T-042 | Amend the charter's no-scaling rule — CLAUDE.md ruled out scaling; T-039 makes model size a construct requirement. Amend or violate? | T-039 | answered |
T-044 | The standard pre-edit condition overstates possession — The editing literature selects edit sets with P(targettrue) > P(targetnew). Is that a valid possession test? | T-039 | answered |
T-045 | Possession of the EDIT TARGET, not just the audited model — Possession was measured on Llama-70B, but ROME would be applied to GPT-J-6B. If GPT-J sits nearer GPT-2 than Llama, the… | T-039 | answered |
T-046 | Reframe the primary deliverable as the possession/structure map — The strongest results so far are edit-independent: the measurement critique (75/61/12 on identical items) and the possession… | T-039 | answered — reframed, but SPLIT |
T-047 | Lens 2 second pass — factual probing, not editing — R-002 searched the editing literature because the claim was about editing. If the headline becomes "what does the model hold",… | T-046 | answered — [R-006] |
T-049 | Type-matching is not cue-matching — E-003/E-004 controlled distractors by type — when the relation wants a place, every candidate is a place. Is that the right… | T-039 | answered |
T-050 | P19 is the pilot's pinch point — [E-005] — P19 place of birth tops out at 53% even at Llama-405B, and GPT-J holds 13%. P19/P20 carry the richest grounds (burial,… | T-045 | answered |
T-051 | Candidate-set size, not cue contamination, drove the inflation — [T-049] claimed type-matching is not cue-matching and that surface cues inflated E-003/E-004. How much did that actually… | T-049 | answered |
T-055 | Zeugma as a type-matching validator for candidate sets — The possession measure depends on candidates being type-matched — that is what stops CounterFact's ambiguous templates from… | T-044 | answered — shipped at reduced strength |
T-057 | Was CounterFact filtered against a model at all? — [D-002] found the ROME paper never names a model used to score records during CounterFact construction, and describes a… | T-054 | answered — from the paper, no run needed |
T-058 | Transitive containment as the deductive ground family — [E-002] found grounds as Wikidata expresses them are evidential — burial place supports place of death without entailing it. To… | T-041 | answered |
T-059 | Is the outer deficit answer surface form rather than knowledge? — [E-009b] found the outer possession deficit is two strings — United States 6/40 held and United Kingdom 2/14, against France… | T-058 | answered |
T-060 | Does the gate clear at 8B, dissolving the 70B editability problem? — The whole ROME-at-70B covariance problem (~3.3 GB over dmlp 28672, uncollected and uncosted, recorded as the unpaid consequence… | T-059 | answered — yes |
T-061 | Is subject familiarity the variable behind the leg dependence? — [E-009] measured that chain legs are not independent (inner1 & outer lift 2.10 at GPT-J, 1.15 at 70B) and explained it as… | T-058 | answered — no |
T-062 | Is possession a property of a fact, or of a (fact, template) pair? — notes/definitions.md declaration 1 is binding — "a fact is a behavioral unit defined by its probe set" — and every possession… | T-059 | answered — yes, a fact-level predicate exists |
T-064 | MUTE — known but inexpressible in this phrasing — Crossing [E-012]'s two tests gives four cells, and one of them has no name in the literature: row test fails, column test passes… | T-062 | answered — two causes, not one |
T-065 | Is the instrument's atom a contrastive pair, with content only at crossings? — Raised by the user, on noticing that nearly every measurement here is a minimal contrastive pair — subject vs placeholder, bare… | T-062 | answered — claim 1 sharpened, claim 2 refuted |
T-066 | What does a closed candidate pool hide that free generation would show? — Every measurement in this repo ranks over a closed, type-matched pool. That is what makes it controlled — and [E-014] had to… | T-065 | answered — a great deal |
T-067 | A causal measure of relocation, not a behavioural one — [E-015] asks whether the model behaves as if it inferred "born in Germany ⇒ born in a German city", by holding subject and target… | T-065 | answered — both |
T-073 | Pool attractor mass as a required artifact field — Washington, D.C. absorbed 33% of all birth-arm destinations in [E-015] — it is the pool's high-prior city and the default sink… | T-066 | answered — implemented |
T-074 | Does the pinned coefficient survive a prompt that mentions the subject late? — [E-016] established that ROME's u·k normalisation pins the update coefficient to exactly 1 at the subject's last token, so any… | T-067 | answered — yes, it survives |
T-075 | Is the subject key context-robust at every layer, or only at layer 5? — [E-017] measured the coefficient at the subject's last token to be 0.93-1.00 across probe forms at layer 5 — the layer ROME edits… | T-074 | answered — only at shallow layers |
T-076 | Is the different-subject floor high enough to weaken [E-017]? — Split out of the T-075 design because it is not a layer question and does not need the sweep. One measurement at layer 5:… | T-075 | answered — no, the floor is 0.082 |
T-077 | Is the depth decay a direction change or a magnitude change? — [E-019] measured reformulated probes decaying 0.95 → 0.58 with depth and explained nothing about it. The coefficient is not a… | T-075 | answered — direction |
T-079 | Does the decayed coefficient at depth actually produce less displacement? — The joint [T-075] explicitly left open: "Whether the decayed coefficient at L20 actually produces less displacement is untested;… | T-075 | answered — DENY: the coefficient bounds delivery, not effect |
T-080 | Is the pinning a ROME fact or a Llama-3.1-8B fact? — Every coefficient in the record is measured on one model. The derivation behind the pinning depends only on ROME's u·k… | T-076 | answered — the pinning is ROME's; the floor is Llama's |
T-081 | Is the floor-plus-type-displacement structure a birthplace fact? — [T-065] found every narrowing in this project came from varying a dimension previously held fixed, and named the three never… | T-072 | answered — type matching generalises; the floor is one cell |
T-082 | Is GPT-J's high different-subject floor real, or just unfamiliarity? — [E-027] measured GPT-J's different-subject floor at 0.155 at the edit layer against Llama's 0.082, and at 0.524 at the final… | T-080 | answered — real; the floor is architectural |
T-083 | Why is the different-subject floor lowest near the layer ROME edits? — [E-027] noticed, without looking for it, that both models' floors are non-monotonic with a minimum near depth fraction 0.15 —… | T-080 | answered — the premise is false; it is not lowest there |
T-084 | Is a weak edit at depth under-optimisation, or a property of depth? — [E-026]'s layer-20 edits suppressed the true answer by only 1.81 nats against layer 5's 9.70, while achieving comparable target… | T-079 | answered — under-optimisation, not depth |
Parked (8)
| id | question | parent | status |
|---|---|---|---|
T-003 | Can RippleBench-Maker's distance function take a code dependency graph? — RippleBench (2512.04144) advertises a substitutable distance function including graph path length. If a compiler-derived… | T-001 | parked |
T-011 | Is orphaning structural to expansion operators as a class? — An LM has no retraction primitive — gradient editing can only move mass toward a target, and "stop believing what implied the old… | T-007 | parked |
T-012 | Grounds by intervention rather than enumeration — — | T-006 | parked |
T-043 | The dimension space of a knowledge bit — Nodes have more axes than we are using: direction (done), in-degree, out-degree, support type (deductive/evidential), modality… | T-031 | parked |
T-048 | Grain confound in the head-vs-ground comparison — [E-004] found ground possession (82%) above head possession (73%), contradicting the prediction. The cause looks like… | T-039 | parked |
T-052 | Grounds are a STAR, not a chain — We have said throughout that "the graph gives grounds". It does not. Every ground we have used — P19 birth, P27 citizenship, P119… | T-043 | parked |
T-053 | Generic question generation from property metadata — [E-004] found template coverage, not ground availability, was the binding constraint — 26 hand-written templates covered 13… | T-052 | parked |
T-056 | Dual-facet nouns as a probe of representational structure — Does a model hold "ledger" as simultaneously physical matter and informational content? Co-predication is the standard diagnostic… | T-055 | parked — candidate NEXT project, not this one |
Dropped (3)
| id | question | parent | status |
|---|---|---|---|
T-063 | Are entailed conclusions less robust than their premises? — The move-4 variable from the Part II dry-run, and the one quantity in this design nobody else has asked. If a conclusion is… | T-062 | dropped |
T-070 | An edit as an entrenchment probe — From the premise dry-run: if an editor systematically yields the defeasible premise and retains the necessary one, it implements… | T-063 | dropped |
T-071 | Does declaration 6 need amending? — Declaration 6 says the choice of what to retract "belongs to a human and not to the method". If the editor makes that choice… | T-070 | dropped |