MSH6 and LDLR explanation case studies

This supplementary experiment asks GPT-6 Astra at max reasoning effort to justify its predictions for one deliberately selected panel: OpenSplice MSH6 exon 7 (E01, ENSE00003677325). All 50 benchmark alleles are retained. Selection was based on mechanistic interpretability after reading the paper and measurements, before inspecting this panel's model explanations. It is not a representative performance sample or a case selected for model failure. On September 17, the user also requested the older LDLR promoter example below, retaining its full 50-allele panel before model-trace inspection.

Complete prompts and responses

The exact final responses, including their original Markdown and all 50 final predictions, are committed with this article. Expand either response to read it here. The downloads also include the exact prompts, baseline answers, and data needed to recompute both correlations. The separate provider-exposed summaries are partial records, not complete or verified accounts of internal reasoning.

The full experiment archive (9.8 MB) preserves all 32 original files, including complete provider payloads, question snapshots, request settings, usage, attempt receipts, and baseline provenance. Its manifest records each file's size and SHA-256 digest; checksums also cover the downloadable archive. It is published separately from official benchmark versions and leaderboard results.

Reference interpretation, fixed before model explanations

The OpenSplice preprint shows the mutational landscape in Fig. 2f and the regulatory map in Fig. 6a, group 6. The Results paragraph beginning “Similar diversity is also observed among highly included exons” interprets MSH6 exon 7 specifically. The authors describe a 90-nt exon with roughly 97% wild-type inclusion and a sparse regulatory architecture: 78.4% neutral and 21.6% enhancer-state sequence across the 185-nt native insert. They locate an upstream enhancer-state segment at insert positions 47–50, a 21-nt acceptor-region segment, and a 15-nt segment across the exon end and donor. The reported states summarize substitution and short-deletion effects; they do not identify bound proteins. The 21-nt deletions below were not used to construct those maps. No direct branchpoint or RNA-binding-protein validation for MSH6's upstream segment was identified in the retrieved main text. Calling that segment a proven branchpoint would exceed the available evidence.

That paragraph records the reference available before model inspection. The expanded literature audit below, added on September 18, identifies earlier functional evidence for the branchpoint motif.

The assay uses a FAS three-exon minigene in HEK293T cells. The supplied sequence is in construct orientation. The 185-nt native insert occupies cassette bases 147–331; the tested exon is 217–306. Add 146 to native-insert coordinates to obtain cassette coordinates. Thus the upstream 47–50 segment maps to 193–196, the acceptor AG to 215–216, and the donor GT to 307–308. Junction recognition is a sequence-grounded interpretation; exact effects and the regulatory-state map require measurements.

Deletion intervals describe removed bases, excluding each normalized VCF anchor. Tiny positive measurements are described as tolerance, not evidence of a specific inclusion-enhancing mechanism. These illustrations were fixed before model inspection; the scored experiment still contains all 50 alleles. Measurements come from the pinned benchmark source panel.

The model can locate the junctions, read alleles and deletion lengths, and use the supplied cell and reporter context. It is not given the gene name, wild-type PSI, measured effects, paper, map labels, or these illustrative candidates. Specific protein binding, branchpoint usage, altered RNA structure, and cryptic splice-product identities would need additional evidence; motif plausibility alone does not establish them. Reporter results also do not establish an effect in endogenous tissues or a clinical consequence.

LDLR secondary reference, fixed before model explanations

Kircher et al. (2019), Fig. 1d and the LDLR Results section, describe effects clustered at SP1 and SREBP1/SREBP2 sites. The established sterol-responsive sequence ATCACCCCAC occurs at reporter positions 161–170 in the pinned LDLR panel. The primary functional study identifies that sequence in its reporter-construct Methods. The five selected substitutions within it all reduce activity:

These measurements are consistent with disrupting a positive regulatory element; they do not demonstrate altered protein binding separately for every allele. Sequence localization and effect direction are assessable from the model's explanation. Specific TF occupancy, affinity changes, and precise effects require additional evidence. The benchmark supplies the complete insert in construct orientation and an episomal reporter in HepG2 cells, measured after 24 hours with no added treatment. It does not supply gene identity or a sterol-depletion condition. Predictions premised on an unprovided treatment will be identified.

The source is the study's LDLR replicate 2, as recorded in the pinned satMutMPRA source. All 50 alleles remain in the scored panel. Both the 2019 study and the underlying mechanism predate Astra's cutoff, so this secondary case supplies mechanistic interpretability without the primary case's later-release rationale.

Availability evidence

Astra's documented cutoff is April 30, 2026. The publisher version record dates the first preprint to May 23, 2026, before the Figshare initial release on May 24. The September 16 verification retained the publisher XML and version record and traced the study's predecessor-release evidence; see the provenance audit. The selected source uses Figshare v5. Earliest-supplement versus v5 byte identity for these exact 50 measurements has not been established, and the supplementary TSV was not independently checked. A post-cutoff first verified study date is evidence about public availability, not proof of absence from training. General splice-site knowledge predates this assay.

Explanation request and comparison

The experiment replaces only the optional explanation sentence in the original prompt. It requests substantive sequence-grounded biological justifications, candidate IDs, proposed mechanisms, exceptions, alternatives, and uncertainty, without a fixed word limit. It permits established biological knowledge but asks the model not to substitute assumed gene identities or remembered experimental results for a justification grounded in the supplied input. It also explicitly permits saying that the input does not support a mechanistic explanation. This instruction guides the stated justification; it cannot establish whether recall influenced the prediction. The sequence, geometry, VCF, prediction target, and required final JSON remain unchanged. A distinct template version and question digest identify this supplementary prompt. The full question snapshot retains reference measurements for scoring; only its prompt goes to the model.

Each panel used openai/gpt-6-astra, max effort, seed 20260829, and the existing 128,000-token output ceiling including reasoning. Sampling controls are omitted. These settings follow the requested effort and existing benchmark controls; they are not a provider-endorsed biology recipe. The ceiling is a resource limit, not a requested explanation length.

The experiment has a separate $15 user-authorized budget, unrelated to issue #72. The OpenAI route is pinned with fallbacks disabled and all request parameters required. Prompt and completion prices are capped at $10 and $50 per million tokens, with no per-request fee. The recorded preflight bounds both requests together even before an anticipated 50% batch discount. There are no automatic retries. Exact request bodies, price assumptions, and submission receipts are retained beside the result files.

The September 17 batch returned malformed upstream results for both requests, with no usable model output and $0 charged. A September 18 direct attempt was interrupted locally during the first request, without a saved response or cost receipt. The user then clarified that both panels should run directly. The restart reserves the interrupted request's full possible cost and checks the remaining budget before each request. Prompts and inference settings remain unchanged; the existing streaming transport is used to keep the connection alive. All attempts and the transport changes are recorded separately from the frozen selection. A missing response is not treated as a free request.

The completed MSH6 direct response cost $2.381404. Its receipt left insufficient room for LDLR's full standard-price ceiling while retaining the interrupted request's $6.45879 reserve, so LDLR was withheld locally. The subsequent LDLR request pins OpenAI's openai/flex direct endpoint and caps prices at $5/$25 per million input/output tokens. The prompt, max reasoning effort, seed, and 128,000-token ceiling are unchanged. Including that new request's ceiling, the conservative total is $12.065004. The route snapshot and all earlier records remain beside the experiment artifacts.

The September 15 publication's strongest completed configuration was Astra High (overall Spearman 0.598976), as recorded in the original freeze. The September 18 publication includes the user's completed Astra Max baseline, now strongest at 0.610478. This later availability does not change the selected panels. The matching baseline runs are gpt-6-astra-max-opensplice-snv-20260916 and gpt-6-astra-max-satmut-mpra-20260916, both served by OpenAI with max effort, seed 20260829, and a 128,000-token ceiling. The supplementary experiment adds provider-routing price restrictions and uses direct requests. No duplicate baseline is requested; provider and inference-setting differences are retained in the comparison artifacts.

The comparison distinguishes sequence localization and allele/orientation handling, predicted direction and magnitude, proposed causal links, unsupported specificity, contradictions, and unresolved claims. Explanations are stated justifications, not verified descriptions of internal reasoning. Agreement with the paper is not mechanistic proof. Existing deterministic Spearman and Pearson scoring applies to all 50 predictions. A single completion per condition cannot establish a benefit from requesting explanations or separate prompt effects from sampling variability. Supplementary results do not enter the leaderboard.

Primary result: MSH6

The explanation completion returned valid predictions for all 50 alleles without truncation. Spearman was 0.803453, compared with 0.799818 in the matching Astra Max baseline; Pearson was 0.713871, compared with 0.740322. These small, mixed changes from one completion per condition do not establish an improvement from requesting an explanation.

The six illustrations fixed before model inspection show why the overall score does not settle the mechanistic interpretation:

The explanation correctly localizes the canonical acceptor and donor, distinguishes the final exonic donor base from the intronic GT, and describes V44's mutant donor as CT rather than GC. Its graded V41 prediction closely matches the measured effect. It also excludes retained VCF anchors when describing deletion intervals.

For V19 it reconstructs the newly created AG216–217 and proposes a shifted acceptor, predicting only −10 points despite the measured −87.466. This is a plausible sequence hypothesis with an unsuccessful prediction, not evidence that the alternative junction forms. The model explicitly identifies uncertainty about whether shifted products count toward PSI. A post-response check of the paper's Methods, “PSI per barcode”, finds that isoforms were classified by exact matching to expected inclusion or skipping sequences, with expected exonic deletions incorporated into the inclusion sequence. Arbitrary boundary shifts therefore cannot simply be assumed to rescue the scored inclusion target. Such products can be excluded from both inclusion and skipping counts: the reported PSI is inclusion divided by inclusion plus skipping, not inclusion divided by all possible products. A low measured PSI does not prove that no shifted product formed. Our prompt gives the geometry and PSI target but omits this exact-match counting rule. This target ambiguity limits attribution of the discrepancy to biological reasoning alone; the frozen experiment is not changed after seeing the result.

For the interior, the model proposes enhancer loss for V24 and silencer loss for V28/V30, tying these to short SR-associated, RBFOX-like, and PTBP-like motifs. It explicitly calls binding and activity unproven. The measured near-neutral effects do not support the predicted magnitudes or establish those mechanisms. The paper's largely tolerant interior remains the stronger empirical account. Neither these measurements nor the model's caveats establish which proteins bind.

The completed response used 47,597 output tokens, of which the provider reports 42,838 as reasoning tokens, plus 1,527 input tokens. The $2.381404 direct charge is separate from the unresolved interrupted attempt. The baseline's per-panel cost is allocated from a batch receipt, so a cost difference cannot be attributed solely to the explanation instruction.

What the full MSH6 panel reveals

This expansion is a post-response audit, added on September 18. The six examples above remain the original illustrations. Here we examine every allele, with coordinates in the submitted cassette and deletion anchors excluded. The exon spans 217–306. The upstream enhancer-state segment from OpenSplice maps to 193–196; the sequence-defined acceptor AG|G is 215–217 and donor TAG|GTAAGA is 304–312. These short recognition sequences are not claimed to be the full boundaries of the paper's 21-nt and 15-nt enhancer-state segments. An enhancer state inferred from mutagenesis also does not identify an SR protein.

The following groups partition all 50 alleles by affected bases. Acceptor and donor overlaps take priority over interior or intronic location. A deletion can extend beyond its assigned region. These are descriptive means for the selected panel, not estimates of how all mutations in each region behave.

The model captures the broad concentration of damaging variants at the junctions, which is consistent with its high rank correlation. It substantially compresses the acceptor losses, and the explanation completion moves both junction means closer to zero than the baseline. Interior means conceal opposing predictions of enhancer loss and silencer loss; individual measurements are more informative.

Earlier experiments on this exon

Additional literature changes the evidential status of the proposed branchpoint. Canson et al. (2022), Fig. 3D tested MSH6 exon 7 in pSPL3 minigenes in Ishikawa cells. Within TTCAT, T191G, C192G, and A193G correspond to c.3557−26, −25, and −24; they reported 35%, 1%, and 6% skipping, respectively. Combined T191G/A193G gave 46%. They also discussed alternative branchpoints, including an experimentally inferred site at −28. This supports the motif's functional relevance and possible redundancy, without establishing A193 as the sole branchpoint in OpenSplice's different reporter and cell line. None of these three substitutions is in our 50-allele panel.

OpenSplice cites Mercer et al. (2015) and Taggart et al. (2017) as references 34 and 35. Canson connects those mapping datasets to this exon and supplies the targeted functional experiment; Canson itself is not cited in the retrieved OpenSplice reference list. Thus the literature connection is through the earlier branchpoint studies, rather than a prior MSH6 experiment cited in its exon example.

There is also evidence in endogenous patient RNA. Šetrajčič Dragoš et al. (2022), Table 1 and Fig. 1G reported 6.8% exon-7-skipping reads for c.3646+5G>A, with none detected in controls. Its deleted transcript interval, r.3557_3646, is the same 90-nt exon; the donor substitution maps to cassette G311A. That exact substitution is absent from our panel, although several deletions remove G311. This independently supports donor sensitivity outside the GT dinucleotide. The patient read fraction is not a mutant-only ΔPSI, and should not be substituted for an OpenSplice effect or used to calibrate a different allele. Neither study validates the model's specific SRSF2, RBFOX, or PTBP assignments in the exon interior.

These exon-specific experiments predate Astra's cutoff. The newer OpenSplice measurements do not imply that this exon's regulatory mechanisms were previously unknown, or establish whether earlier knowledge influenced this response.

Upstream losses cannot be assigned to one branchpoint

V07 removes 184–204 and V09 removes 186–206. Both remove the TTCAT motif, including A193, and the entire 193–196 enhancer-state segment. Yet their measured losses differ by 54.747 points: −29.249 versus −83.996. Astra predicts −22 and −30, getting the order but missing most of the separation. The reference poly-T run occupies 201–213; these deletions leave nine versus seven of its Ts. That offers a plausible contributor, alongside changed spacing and deletion junctions. The two alleles do not isolate tract length or a single branchpoint, so their difference cannot establish any one of those mechanisms.

New AG sequences do not demonstrate scored rescue

Reconstructing each mutant confirms four nearby AGs discussed by the model. Positions below refer to original cassette bases even when deletion brings them together. Sequence creation is directly checkable; splice usage is a hypothesis.

The shared failure is treating a plausible nearby site as sufficient rescue of the assay's expected inclusion product. The counting limitation described above prevents deciding from these ΔPSIs whether shifted products actually formed. V21 provides a complementary example: G217T preserves the canonical intronic AG but yields −60.673, versus Astra's −6. Acceptor recognition extends beyond the dinucleotide, and correctly finding the AG does not quantify its strength.

The interior tolerates the model's proposed motif losses

All 11 interior alleles have measured effects between −1.120 and +0.624. V24, V28, and V30 each remove 21 bases, leaving a 69-nt exon, yet measure +0.297, +0.488, and +0.415. Astra instead predicts −16, +8, and +6, invoking an SR-associated enhancer (CCAG, 253–256), an RBFOX-like motif (TGCATG, 265–270), and a PTBP-like pyrimidine motif (TCTCT, 282–286). These net effects do not support the predicted magnitudes. They do not prove absence of binding, or exclude compensating effects of a long deletion. The 21-nt deletions were excluded from fitting the paper's SpliceMap, but still come from the same assay; they are not independent biological replication.

The paper's approximately 97% wild-type PSI also leaves only about 3 points of possible inclusion gain. Predictions of +8 and +6 exceed that available range. Wild-type PSI was withheld from the prompt, making this a limitation in effect-size calibration with missing information. It does not by itself negate the ranking score or establish why the model chose those values.

Mutant donor reconstruction is informative, but incomplete

Some of the strongest reasoning is about the actual sequence left after deletion. Here the bar marks the original exon–intron boundary, retained through the edit.

Astra correctly orders the one-base V40 deletion as worse than the three-base V39 deletion, and the six-base V46 deletion as worse than the 21-base V49 deletion. In V49, donor position +5 is now the C at original cassette position 332, in the fixed FAS intron. This is a concrete reason to supply the complete reporter context: the deletion draws on sequence outside the native insert. The sizeable V47 miss, despite retaining GT, shows the limit of assigning a moderate penalty from a recognizable donor alone. The broad spatial pattern is convincing; several detailed rescue and motif-strength predictions remain wrong.

Secondary result: LDLR

The direct Flex completion returned valid predictions for all 50 alleles without truncation. Spearman was 0.498629, compared with 0.472678 in the matching Astra Max baseline; Pearson was 0.450955, compared with 0.607670. The explanation therefore gives a slightly better ordering but weaker correspondence between the numerical effect sizes. One completion per condition does not establish a prompt effect, and the service routes differ.

The model correctly identifies ATCACCCCAC at 161–170 and predicts decreased activity for all five preselected substitutions within it. This agrees with the direction of the measurements and the paper's SREBP-site interpretation. It does not establish altered occupancy for any individual allele.

The model acknowledges that motif-core recognition supports effect direction more strongly than fine ordering. The ordering of V28 and V29 is nevertheless reversed: it predicts A164T to be more damaging than A164C, whereas A164C has the larger measured loss. The C/T-rich flanking elements are also correctly localized, with Sp-family binding presented as a hypothesis requiring evidence.

What the established LDLR elements reveal

This expanded audit was added after inspecting the responses, on September 18. It does not change the five preselected illustrations or the 50 scored variants. The reference is now more specific than a search for plausible short motifs: satMutMPRA Fig. 1d and Supplementary Table 12 connect the activity landscape to previously studied promoter elements and alleles. The GPN-Star paper, Fig. 4D labels four blocks FP1, SREBP1, SREBP2, and SP1 and examines dependencies within and between them.

The nomenclature needs care. The satMutMPRA supplement, Table 12, pp. 92–93 defines the numbered labels as SRE repeats (with a duplicated “SREBP1” in the caption). They should not be treated as proof that SREBP-1 and SREBP-2 bind different sites. Classic functional studies identify a distal Sp1 site in repeat 1, the SRE within repeat 2, and a proximal Sp1 site in repeat 3. For example, Peeters et al. found that a repeat-1 deletion abolished Sp1 binding in a footprint assay and reduced reporter activity to about 10% of normal. Astra's Sp-family hypothesis for the upstream C/T-rich sequence is therefore compatible with established biology, despite the figure's SREBP1 label.

The following mapping uses the submitted insert's 1-based, inclusive coordinates. Genomic GRCh38 position is reporter position + 11,089,230; the corresponding NM_000527.4 upstream coordinate is position − 319. The latter is relative to translation, not a transcription-start coordinate. These offsets are checked against every selected source allele; V34, for example, is reporter C177T = chr19:11089407 C>T = c.-142C>T. FP1 uses the experimentally footprinted sequence. The other three intervals are the precise cores discussed by Astra, within the known regulatory modules; they do not define complete binding footprints or classify their flanks as inert.

FP1 is a missed regulatory element, not merely a missing name. Mehta et al. footprinted this exact 20-base sequence, tested substitutions in reporters, and observed sequence-specific nuclear-protein binding. Their induction experiments involved sterol depletion; we do not transfer that treatment or its effect size to this untreated MPRA, nor infer the identity of the bound protein. The independent benchmark measurements nevertheless show losses for all six selected FP1 variants. V07 (G81A) is the most damaging allele in our entire 50-variant panel at −3.784, but both baseline (−0.03) and explanation (0.00) effectively miss it. V09–V12, the four substitutions in the tract Astra discusses at 93–99, measure −1.692 to −2.083 despite predictions of only −0.04 to −0.10. Declining to call this a canonical TATA box is reasonable; treating that lack of a TATA match as insufficient evidence for a substantial effect misses FP1.

Recognition of the SRE is better than recognition of the whole promoter. Every selected variant in each of these four intervals reduces activity. The means below include every selected allele in each interval, including the C128 deletion in repeat 1; the deletion's retained VCF anchor is excluded when assigning its position. These are descriptive summaries of this deliberately selected panel, not new benchmark scores or evidence of a prompt effect.

The explanation singles out the SRE as its strongest loss region, although the sampled FP1 and Sp1-associated variants can be equally or more damaging. It also moves the repeat-1 and repeat-3 means closer to zero than the baseline. That pattern helps explain the weaker Pearson correlation without claiming a causal effect of the explanation instruction. “Other” includes regulatory flanks and does not mean biologically unimportant sequence.

Cooperation changes the interpretation of motif gains

Sanchez et al. experimentally showed that SREBP and Sp1 cooperate, that Sp1 orientation matters, and that other tested TF sites could not simply replace the adjacent Sp1 site. An intact SRE therefore does not imply preserved promoter activity after damage to a cooperating element. This is a more constrained reference model than adding independent benefits for every newly created short motif.

The following post-response illustrations examine the complete group for which the explanation proposes beneficial ETS-site changes. They were not part of the preselected five motif substitutions above, and do not change the scored 50-allele panel:

The new short cores are present in the reconstructed sequences, but their presence does not establish productive factor binding. V18/V34 have the wrong predicted sign, and the proposed compensation in V36 misses a large measured loss. V40/V43 agree in direction; that agreement does not demonstrate the proposed ETS mechanism. The explanation itself cautions that binding, cooperation, and competition require experimental evidence. These are uncertain mechanistic assignments and unsuccessful predictions, rather than fabricated sequence motifs.

V34 is particularly informative because the literature tests the same allele. Table 12 identifies c.-142C>T in the SP1 region and reports −3.09 for LDLR.2, matching our pinned value. In the older transcription-start numbering this is −49C>T: Mozas et al. measured an 80% reduction in HepG2 reporter activity and severely impaired Sp1 binding in gel-shift assays. Those older measurements use a different construct and are not interchangeable with our numerical target, but they independently support loss of an established activator site. Astra notices the new TTCC, yet its proposed +0.15 net effect fails to account for that documented loss mechanism. There is no evidence here that ETS binds the new sequence or rescues activity.

What GPN-Star adds, and what it cannot establish

GPN-Star's Fig. 4D and Supplementary Fig. 14 provide a useful complementary analysis: the model's nucleotide-dependency map resolves the four named regions and connections between them, including the SREBP2-labelled region and SP1. It changes the question from “does this short motif look plausible?” to “has the model captured the interacting regulatory architecture?” Astra's final answer mentions cooperation but misses FP1 and gives unsupported compensation substantial weight at known activator elements. This is our interpretation of its stated explanation, not an observation of its internal computation.

The GPN-Star map is model-derived, not a new binding or double-mutant assay. Its MPRA track also reuses the Kircher measurements: the authors filter by barcode support and significance, average absolute allele effects by position, fill missing positions with zero, and smooth over five positions. The archived LDLR notebook preserves that construction. The track therefore neither supplies independent experimental replication nor tests signed predictions for our 50 alleles. We use it to motivate checking elements and interactions, not to claim a quantitative GPN-Star-versus-Astra performance comparison or experimental epistasis.

Assay handling is generally careful: the model retains deletion anchors, recognizes that no separate minimal promoter is present, and does not assume sterol depletion. It declines to assert TATA-box, upstream-ORF, or methylation mechanisms from insufficient context. We checked its 17 stated reference-sequence features against the supplied insert, as well as the five ETS-related mutant contexts above. These checks establish sequence consistency, not factor occupancy.

The final answer is grounded in sequence positions, but the provider-exposed summary explicitly refers to the “classic LDL receptor promoter.” The prompt's mechanistic framing therefore does not demonstrate removal of recognition or recall. This partial provider summary is not a verified account of internal reasoning, and cannot establish how much recognition affected the predictions. The study and underlying mechanism predate the model's cutoff.

The response used 41,624 output tokens, including 37,639 reported reasoning tokens, plus 1,076 input tokens. Its reported cost was $1.04732125. Together, the two completed explanation responses cost $3.42872525; the original failed batch cost $0, and the interrupted direct request's charge remains unresolved. Keeping its full $6.45879 reserve gives a conservative total of $9.88751525, below the authorized $15. No additional completions are planned.

Reproduction

From the repository root, with the evaluator installed:

uv run --locked python projects/blog-analysis/scripts/mechanism_case.py prepare-pair \
  --output .vepbench/experiments/issue97-msh6-ldlr-20260917

Preparation is offline and refuses to overwrite the experiment directory. It verifies the selected question against the frozen digest and saves both question snapshots, the exact proposed request, and the selection record before inference. Production records use the existing evaluator and retain complete provider payloads, usage, reasoning, failures, and costs outside Git. The small, exact readable responses and comparison extracts are committed beside this article. Preparing these files does not submit a request.

The explicit run --experiment PATH subcommand submits the two-request batch once, reading only OPENROUTER_API_KEY for authentication. Use the existing vepbench batch status and vepbench batch collect commands with the saved batch-state.json to retrieve it. An existing execution plan prevents repeat submission even if no final result was collected.

The recover-direct --experiment PATH subcommand is reserved for the reviewed unbilled batch failure. It checks the receipt and preserved failures, reproduces the original request bodies, and reserves a single direct attempt before sending anything. An existing recovery plan blocks repetition even if no response was saved; unknown or billed outcomes require a separate budget review.

The reviewed September 18 restart uses resume-interrupted-direct. It preserves the interrupted attempt, records each new request before contacting the provider, and bounds total spending including the interrupted request's unresolved charge. It refuses to repeat an existing restart plan.

After both result files are available, reproduce the deterministic comparison:

uv run --locked python projects/blog-analysis/scripts/mechanism_case.py compare \
  --experiment .vepbench/experiments/issue97-msh6-ldlr-20260917 \
  --panel msh6 \
  --baseline PATH_TO_BASELINE_RESULTS.jsonl \
  --results PATH_TO_EXPLANATION_RESULTS.jsonl \
  --output .vepbench/experiments/issue97-msh6-ldlr-20260917/msh6-comparison.json

Repeat with --panel ldlr and a distinct output path for the secondary case.

The committed evidence is generated from those saved records, without model calls:

uv run --locked python projects/blog-analysis/scripts/mechanism_case.py export-evidence \
  --experiment .vepbench/experiments/issue97-msh6-ldlr-20260917 \
  --baseline .vepbench/experiments/issue97-msh6-ldlr-20260917/baseline-20260918/results.jsonl \
  --output /tmp/vepbench-mechanism-evidence

Use a fresh output directory; the exporter refuses replacement. The committed evidence is sufficient for offline score and element-audit replay via uv run --locked pytest projects/blog-analysis/tests/test_mechanism_case.py. This checks the exact prompts, file hashes, all predictions, both correlations, allele mappings and interval summaries without access to private local records. The annotation intervals are defined in projects/blog-analysis/config/ldlr-elements.json and projects/blog-analysis/config/msh6-elements.json; no literature labels were added to the model prompt after inference. The linked HF archive preserves the full provider payloads and attempt receipts separately from the readable extracts.