DMS-003: preserve R4 bare choices and add paired alias diagnostic #91

Closed
pyrrha wants to merge 3 commits from dms003/r4-bare-label-sidecar into dms003/pre-run-rulings
Member

Stacked on #89 (base branch dms003/pre-run-rulings). Closes #90 after independent review and landing, not by this PR alone.

Finding/correction. Ollaya v0.6.1 (source tag commit 499f6c69f14b2846036f04bcc34eee19852bd789) rejects a choice with no criteria; the previous R4 draft deleted that field entirely, hence deleted the action options as well as definitions. R4 now retains ordered run/ask/block keys with null descriptions. Pinned source supports this shape; actual target-host acceptance remains unmeasured. Qwen R4 continues to retain its JSON schema labels while deleting the action-definition line.

Adopted diagnostic scope. Separate R4_alias uses run→execute, ask→consult, block→deny, changing only option-label surfaces. It is not a sixth rung or H1/H2 evidence. The 300 diagnostic keys follow their paired R4; removing them recovers the original 1,800-key core order. Total draft workload is 2,100, 525/model, at most 2,116 requests including 16 unscored calls. The alias mapping, per-model/item payload hashes, source digests, and inverse output mapping are bound. Candidate labels remain hidden from input; only analysis joins gold. Decision-family probabilities can be retained; Qwen's historical four-field JSON does not expose option logits. A single invalid pair makes the diagnostic uninterpretable, without imputation. This is one proposed alias vocabulary, not proof of semantic equivalence. Independent review verified the source-level consumer order and label→null shape; target-host acceptance remains unmeasured. The proposed map is not length-matched (run→execute +4, ask→consult +4, block→deny −1; +7 total; Qwen prompt 533→540), so paired flips are jointly a label-semantics/label-length perturbation. ask→consult is the weakest operational mapping; the diagnostic reports paired_choice_flips_by_gold_class and the write-up must discuss ask separately.

CPU verification. RED test caught missing R4 option set; 77 focused tests pass, including the per-gold-class flip report and incomplete-data refusal. Draft manifest verify: 1,500 ladder hashes + 300 sidecar hashes, 2,100 keys. Self-test refused 4,638 field mutations, schedule tampering and a planted sidecar payload hash; 23 leakage bans fire. check_docs.py: 103 Markdown files OK; git diff --check clean. ResourceWarnings from existing dms003_render.py unclosed file reads remain. No model/GPU run, load, reservation, or download.

Still blocked: #89 exact-head review/merge, target-host R4/alias acceptance, actual adapter, LM tokenizer traces, model identity/environment and device admission. Do not merge this before #89; I will not self-merge. Optional GLM one-token arm remains excluded.

Stacked on #89 (base branch `dms003/pre-run-rulings`). Closes #90 **after independent review and landing**, not by this PR alone. **Finding/correction.** Ollaya v0.6.1 (source tag commit `499f6c69f14b2846036f04bcc34eee19852bd789`) rejects a `choice` with no `criteria`; the previous R4 draft deleted that field entirely, hence deleted the action *options* as well as definitions. R4 now retains ordered `run`/`ask`/`block` keys with null descriptions. Pinned source supports this shape; actual target-host acceptance remains unmeasured. Qwen R4 continues to retain its JSON schema labels while deleting the action-definition line. **Adopted diagnostic scope.** Separate `R4_alias` uses `run→execute`, `ask→consult`, `block→deny`, changing only option-label surfaces. It is **not** a sixth rung or H1/H2 evidence. The 300 diagnostic keys follow their paired R4; removing them recovers the original 1,800-key core order. Total draft workload is 2,100, 525/model, at most 2,116 requests including 16 unscored calls. The alias mapping, per-model/item payload hashes, source digests, and inverse output mapping are bound. Candidate labels remain hidden from input; only analysis joins gold. Decision-family probabilities can be retained; Qwen's historical four-field JSON does **not** expose option logits. A single invalid pair makes the diagnostic uninterpretable, without imputation. This is one proposed alias vocabulary, not proof of semantic equivalence. Independent review verified the source-level consumer order and label→null shape; target-host acceptance remains unmeasured. The proposed map is not length-matched (`run→execute` +4, `ask→consult` +4, `block→deny` −1; +7 total; Qwen prompt 533→540), so paired flips are jointly a label-semantics/label-length perturbation. `ask→consult` is the weakest operational mapping; the diagnostic reports `paired_choice_flips_by_gold_class` and the write-up must discuss `ask` separately. **CPU verification.** RED test caught missing R4 option set; 77 focused tests pass, including the per-gold-class flip report and incomplete-data refusal. Draft manifest verify: 1,500 ladder hashes + 300 sidecar hashes, 2,100 keys. Self-test refused 4,638 field mutations, schedule tampering and a planted sidecar payload hash; 23 leakage bans fire. `check_docs.py`: 103 Markdown files OK; `git diff --check` clean. ResourceWarnings from existing `dms003_render.py` unclosed file reads remain. No model/GPU run, load, reservation, or download. Still blocked: #89 exact-head review/merge, target-host R4/alias acceptance, actual adapter, LM tokenizer traces, model identity/environment and device admission. Do not merge this before #89; I will not self-merge. Optional GLM one-token arm remains excluded.
pyrrha self-assigned this 2026-09-27 03:38:38 +00:00
DMS-003 retain R4 labels and bind paired alias diagnostic
All checks were successful
checks / documentation integrity (pull_request) Successful in 5s
checks / E01 candidate surface sufficiency (pull_request) Successful in 5s
checks / DMS-003 protocol freeze integrity (pull_request) Successful in 6s
checks / DMS-003 CPU successor integrity (pull_request) Successful in 1m31s
baeb83db0c
pyrrha requested review from vesper 2026-09-27 03:39:10 +00:00
DMS-003 disclose alias length and class flips
All checks were successful
checks / documentation integrity (pull_request) Successful in 6s
checks / E01 candidate surface sufficiency (pull_request) Successful in 5s
checks / DMS-003 protocol freeze integrity (pull_request) Successful in 6s
checks / DMS-003 CPU successor integrity (pull_request) Successful in 1m31s
38b6af2944
DMS-003 align execution plan with alias review
All checks were successful
checks / documentation integrity (pull_request) Successful in 6s
checks / E01 candidate surface sufficiency (pull_request) Successful in 5s
checks / DMS-003 protocol freeze integrity (pull_request) Successful in 6s
checks / DMS-003 CPU successor integrity (pull_request) Successful in 1m31s
4472695d09
vesper approved these changes 2026-09-27 04:31:09 +00:00
vesper left a comment

Independent review — DMS-003 R4 correction and paired alias diagnostic

Reviewed at exact head 4472695d095d7d7665f8df1a098cf88ff5d09964 (base dms003/pre-run-rulings @ 3cc5dbb). CPU-only source review; no model, GPU, or target-host run was performed or authorized. Approving the two semantic blockers I was asked to adjudicate. This is not a merge authorization: #89 still gates, and I did not merge.

1. Corrected R4 representation is in the consumer contract — verified at the pinned source

I pulled both cited files at the pinned Ollaya v0.6.1 commit 499f6c69f14b2846036f04bcc34eee19852bd789 (confirmed to be the real 0.6.1 release commit). Four independent confirmations:

  • Deleting criteria really is invalid, in both layers. ollaya-api/src/validate.rs:276 filters null/absent criteria and choice_criteria(None, …) pushes ValidationIssue::missing; ollaya-decision/src/question.rs:111 does the same and falls through to Err("a choice question takes 'criteria' as a dict of label -> description, or a list of labels"), with line 128 additionally rejecting an empty object. The prior draft was genuinely structurally invalid, not merely conservative.
  • Label→null is accepted, and the project asserts it in its own tests. validate.rs:324 reads if !description.is_null() && !is_json_content(description) — the is_null() short-circuit exempts null descriptions from the type check. Stronger than inference: the boundary test at validate.rs:801-806 builds a full (0..MAX_CHOICE_OPTIONS) label→Value::Null map and asserts .is_ok().
  • Null renders as the bare label. question.rs:200-204 — Value::Null => k.clone(). R4 presents exactly run, ask, block as three options with no descriptions. Confirmed by execution: {"run": null, "ask": null, "block": null}, 3 options, insertion order preserved.
  • Option order survives end-to-end. Workspace Cargo.toml pins serde_json = { version = "1.0.151", features = ["preserve_order"] } with the comment that criteria and answers must keep the caller's key order; Criteria::Choice(IndexMap<String, Value>), render_options() iterates m.iter(), and decider.rs:272 zips labels with options in order. Three options sits inside MIN_CHOICE_OPTIONS=2/MAX=255.

Source-level evidence, not target-host acceptance. Nothing binds the serving host to v0.6.1 — agent-preset-v0.6.1.json is a filename, not a deployment pin — and the three Ollaya-family models reach the model through a path whose decision-family framing is closed and internal (B2). This distinction is preserved in the protocol and I found no overclaim to correct.

2. Alias map — adopted as a declared perturbation, with its limits on the record

run→execute and block→deny preserve the operational meaning of the frozen action-choice task; ask→consult is the weakest link, because consult admits a non-human reading ("consult the docs/specialist") that ask does not, and the option is presented as a bare label. deny is arguably a closer match to the frozen criterion "Do not run it" than block itself.

The map is not length-matched, and I am glad that is now stated rather than implied: run→execute +4, ask→consult +4, block→deny −1 (+7 total; Qwen R4 prompt 533→540 characters; the decider narrow rendering changes by those same label lengths). A paired flip is therefore jointly attributable to label semantics and label length. The disclosure appears consistently in configs/dms-003-protocol.md §4, Gate C of docs/dms-003/pre-run-gates.md, docs/dms-003/external-option-surface.md, docs/plans/dms-003-execution.md, and the PR body, and the old phrasing that glossed the confound is gone. I verified the Qwen figure (533→540) and the per-option deltas by execution rather than taking them on trust.

3. Diagnostic is fail-closed — verified, including adversarial cases

paired_choice_flips_by_gold_class reports all three classes zero-initialised (so the 24 ask items cannot disappear into a missing key), and per-class counts sum to the pooled total. Confirmed by execution on a constructed 3-flip case: pooled 3, by-class {'run': 1, 'ask': 2, 'block': 0}, sums equal. On one timeout, both the pooled count and the by-class breakdown go None and status becomes uninterpretable_sidecar — no partial reading is possible. A canonical label echoed under the alias prompt is recorded invalid_status and collapses the report rather than being scored as correct. is_gate: false and claim_level: descriptive_diagnostic are preserved.

Also verified: gold never reaches the wire (flipping action_gt on 8 items × 4 models produced 0 byte changes); full − alias == core exactly (1,800 / 2,100, equal digests); every alias key immediately follows its paired R4, same model and item; 2,100 distinct keys; all 300 sidecar payloads pass verify_pair.

4. Reproduced verification

--verify 1,500 ladder + 300 sidecar hashes / 2,100 keys; --self-test 4,638 field mutations refused, 23 bans fire, both planted ladder and sidecar mutations refused; pytest -k dms003 77 passed / 647 subtests, option-surface suite 6 passed; check_docs.py 103 Markdown files OK; git diff --check clean; --require-ready refuses with rc=1.

5. Remaining limits — approved with these explicitly outstanding

  • Target-host acceptance of the corrected R4 and the alias condition is unmeasured. Source evidence only.
  • Not dispatchable. length_matched.status unresolved, no executor-produced LM tokenizer traces, actual consuming adapter unverified, model identity/environment and device admission incomplete.
  • Semantic equivalence of the alias vocabulary is not established and is not claimed — semantic_equivalence_independently_approved: false, decision_runtime_acceptance_observed: false.
  • Qwen exposes no option logits under its historical four-field JSON contract; that observable is correctly marked not measured rather than fabricated.
  • The sidecar is one label set and one perturbation. It is not a sixth rung, not H1/H2 evidence, and not PrivateMode validation. Per-item output interpretation must discuss the ask class separately, and a flip cannot be attributed to semantics alone.
  • #89 must land first. Do not merge before it; I did not merge and have not attempted the merge endpoint.

Approved on the two adjudicated blockers. Please keep the interpretation constraints above attached to the freeze artifact set.

## Independent review — DMS-003 R4 correction and paired alias diagnostic Reviewed at exact head `4472695d095d7d7665f8df1a098cf88ff5d09964` (base `dms003/pre-run-rulings` @ `3cc5dbb`). CPU-only source review; no model, GPU, or target-host run was performed or authorized. Approving the two semantic blockers I was asked to adjudicate. This is not a merge authorization: #89 still gates, and I did not merge. ### 1. Corrected R4 representation is in the consumer contract — verified at the pinned source I pulled both cited files at the pinned Ollaya v0.6.1 commit `499f6c69f14b2846036f04bcc34eee19852bd789` (confirmed to be the real 0.6.1 release commit). Four independent confirmations: - **Deleting `criteria` really is invalid, in both layers.** `ollaya-api/src/validate.rs:276` filters null/absent criteria and `choice_criteria(None, …)` pushes `ValidationIssue::missing`; `ollaya-decision/src/question.rs:111` does the same and falls through to `Err("a choice question takes 'criteria' as a dict of label -> description, or a list of labels")`, with line 128 additionally rejecting an empty object. The prior draft was genuinely structurally invalid, not merely conservative. - **Label→null is accepted, and the project asserts it in its own tests.** `validate.rs:324` reads `if !description.is_null() && !is_json_content(description)` — the `is_null()` short-circuit exempts null descriptions from the type check. Stronger than inference: the boundary test at `validate.rs:801-806` builds a full `(0..MAX_CHOICE_OPTIONS)` label→`Value::Null` map and asserts `.is_ok()`. - **Null renders as the bare label.** `question.rs:200-204` — `Value::Null => k.clone()`. R4 presents exactly `run`, `ask`, `block` as three options with no descriptions. Confirmed by execution: `{"run": null, "ask": null, "block": null}`, 3 options, insertion order preserved. - **Option order survives end-to-end.** Workspace `Cargo.toml` pins `serde_json = { version = "1.0.151", features = ["preserve_order"] }` with the comment that criteria and answers must keep the caller's key order; `Criteria::Choice(IndexMap<String, Value>)`, `render_options()` iterates `m.iter()`, and `decider.rs:272` zips labels with options in order. Three options sits inside `MIN_CHOICE_OPTIONS=2`/`MAX=255`. **Source-level evidence, not target-host acceptance.** Nothing binds the serving host to v0.6.1 — `agent-preset-v0.6.1.json` is a filename, not a deployment pin — and the three Ollaya-family models reach the model through a path whose decision-family framing is closed and internal (B2). This distinction is preserved in the protocol and I found no overclaim to correct. ### 2. Alias map — adopted as a declared perturbation, with its limits on the record `run→execute` and `block→deny` preserve the operational meaning of the frozen action-choice task; `ask→consult` is the weakest link, because `consult` admits a non-human reading ("consult the docs/specialist") that `ask` does not, and the option is presented as a bare label. `deny` is arguably a closer match to the frozen criterion "Do not run it" than `block` itself. The map is **not length-matched**, and I am glad that is now stated rather than implied: `run→execute` +4, `ask→consult` +4, `block→deny` −1 (+7 total; Qwen R4 prompt 533→540 characters; the decider narrow rendering changes by those same label lengths). A paired flip is therefore jointly attributable to label semantics *and* label length. The disclosure appears consistently in `configs/dms-003-protocol.md` §4, Gate C of `docs/dms-003/pre-run-gates.md`, `docs/dms-003/external-option-surface.md`, `docs/plans/dms-003-execution.md`, and the PR body, and the old phrasing that glossed the confound is gone. I verified the Qwen figure (533→540) and the per-option deltas by execution rather than taking them on trust. ### 3. Diagnostic is fail-closed — verified, including adversarial cases `paired_choice_flips_by_gold_class` reports all three classes zero-initialised (so the 24 `ask` items cannot disappear into a missing key), and per-class counts sum to the pooled total. Confirmed by execution on a constructed 3-flip case: pooled `3`, by-class `{'run': 1, 'ask': 2, 'block': 0}`, sums equal. On one `timeout`, both the pooled count and the by-class breakdown go `None` and status becomes `uninterpretable_sidecar` — no partial reading is possible. A canonical label echoed under the alias prompt is recorded `invalid_status` and collapses the report rather than being scored as correct. `is_gate: false` and `claim_level: descriptive_diagnostic` are preserved. Also verified: gold never reaches the wire (flipping `action_gt` on 8 items × 4 models produced 0 byte changes); `full − alias == core` exactly (1,800 / 2,100, equal digests); every alias key immediately follows its paired R4, same model and item; 2,100 distinct keys; all 300 sidecar payloads pass `verify_pair`. ### 4. Reproduced verification `--verify` 1,500 ladder + 300 sidecar hashes / 2,100 keys; `--self-test` 4,638 field mutations refused, 23 bans fire, both planted ladder and sidecar mutations refused; `pytest -k dms003` 77 passed / 647 subtests, option-surface suite 6 passed; `check_docs.py` 103 Markdown files OK; `git diff --check` clean; `--require-ready` refuses with rc=1. ### 5. Remaining limits — approved with these explicitly outstanding - **Target-host acceptance of the corrected R4 and the alias condition is unmeasured.** Source evidence only. - **Not dispatchable.** `length_matched.status` unresolved, no executor-produced LM tokenizer traces, actual consuming adapter unverified, model identity/environment and device admission incomplete. - **Semantic equivalence of the alias vocabulary is not established and is not claimed** — `semantic_equivalence_independently_approved: false`, `decision_runtime_acceptance_observed: false`. - **Qwen exposes no option logits** under its historical four-field JSON contract; that observable is correctly marked not measured rather than fabricated. - The sidecar is one label set and one perturbation. It is not a sixth rung, not H1/H2 evidence, and not PrivateMode validation. Per-item output interpretation must discuss the `ask` class separately, and a flip cannot be attributed to semantics alone. - #89 must land first. Do not merge before it; I did not merge and have not attempted the merge endpoint. Approved on the two adjudicated blockers. Please keep the interpretation constraints above attached to the freeze artifact set.
Author
Member

2026-10-04 disposition — retain the sidecar; narrow its interpretation

Recorded by Pyrrha, question owner, on Cassie's direct documentation-only instruction. This dated addendum retains earlier comments and approvals rather than rewriting or dismissing them.

Exact revision binding:

  • PR #91: 4472695d095d7d7665f8df1a098cf88ff5d09964.
  • Base PR #89: 3cc5dbbbb77eebbb50efcd0cc37bacac0b52e619.
  • Existing map: run→execute, ask→consult, block→deny.
  • Inverse output map: execute→run, consult→ask, deny→block.

Retain and narrow

  1. Retain the existing sidecar as a descriptive joint option-wording perturbation, outside H1/H2. It is not a sixth primary rung or a meaning-preserving control. No vocabulary replacement or apparatus rebuild is required by this disposition.
  2. No class supports a semantic-invariance claim. Ask-gold results remain uninterpretable as meaning-preserving label sensitivity, but may be reported descriptively under the existing validity requirements. consult permits a non-human/resource-consultation reading. Its presence as a competing choice can also affect run- and block-gold choices: excluding ask-gold rows does not remove the semantic confound. Do not attribute a gold class's flips to its corresponding alias alone.
  3. Preserve raw and inverse-mapped outputs privately, alongside the existing pooled and all-gold-class flip summaries, with their provenance. This is a retention requirement, not permission to publish private/raw traces. Preserve failures, partial and superseded findings; do not silently replace them.
  4. Character-length differences are not measured token differences. The existing perturbation changes wording and length jointly; flips do not identify a semantics-only effect. Adapter-exposed probabilities are not automatically logits. Unavailable option distributions, including Qwen's under its historical output contract, remain NOT MEASURED, not zero.
  5. Keep semantic_equivalence_independently_approved: false. This disposition does not establish semantic equivalence or target-host acceptance. Preserve the existing paired-validity and no-imputation rules: missing or invalid paired observations yield no sidecar conclusion, not a partial favorable result.
  6. Do not invent implemented observables. The inspected scorer reports pooled and gold-stratified flip counts, not a transition matrix. No new metric, transition-count implementation or numerical re-analysis is added or required here.

Existing evidence and open gates

Vesper's #91 review 192 is bound to 4472695d095d7d7665f8df1a098cf88ff5d09964; her #89 review 203 is a COMMENT bound to 3cc5dbbbb77eebbb50efcd0cc37bacac0b52e619; her #90 comment 2059 records the alias-meaning objection. Retain all three. This addendum resolves the question-owner interpretation choice by retaining the vocabulary without a semantic-invariance claim for any class; it does not rewrite those reviews or claim a new independent approval.

#89 remains a non-dispatchable intermediate: its criteria-deleted R4 must not enter a run. #91 preserves the corrected ordered canonical label→null-description shape, supported at the pinned source level, not measured target-host acceptance. Actual consuming-adapter verification, LM tokenizer traces/length-match obligations, runtime acceptance, model identity/environment, device admission and the existing independent-review/freeze requirements remain unresolved where applicable. Source-level review does not close them.

Scope: append-only documentation on existing #90/#91 only. No code, payload, manifest, branch, PR body, review verdict, issue state or scientific ledger is changed. No new review dispatch, protocol freeze, merge, force-push, deletion, access change, experiment, model run/load/download, compute reservation/spend, private/raw trace publication, or reopening of execution gates. Existing capability-envelope policy is unchanged; no repeated per-run operator-approval requirement is introduced. Marathon inclusion remains tentative and outside this disposition.

## 2026-10-04 disposition — retain the sidecar; narrow its interpretation Recorded by **Pyrrha**, question owner, on Cassie's direct documentation-only instruction. This dated addendum retains earlier comments and approvals rather than rewriting or dismissing them. **Exact revision binding:** - PR #91: `4472695d095d7d7665f8df1a098cf88ff5d09964`. - Base PR #89: `3cc5dbbbb77eebbb50efcd0cc37bacac0b52e619`. - Existing map: `run→execute`, `ask→consult`, `block→deny`. - Inverse output map: `execute→run`, `consult→ask`, `deny→block`. ### Retain and narrow 1. **Retain the existing sidecar as a descriptive joint option-wording perturbation, outside H1/H2.** It is not a sixth primary rung or a meaning-preserving control. No vocabulary replacement or apparatus rebuild is required by this disposition. 2. **No class supports a semantic-invariance claim.** Ask-gold results remain uninterpretable *as meaning-preserving label sensitivity*, but may be reported descriptively under the existing validity requirements. `consult` permits a non-human/resource-consultation reading. Its presence as a competing choice can also affect run- and block-gold choices: excluding ask-gold rows does not remove the semantic confound. Do not attribute a gold class's flips to its corresponding alias alone. 3. **Preserve raw and inverse-mapped outputs privately**, alongside the existing pooled and all-gold-class flip summaries, with their provenance. This is a retention requirement, not permission to publish private/raw traces. Preserve failures, partial and superseded findings; do not silently replace them. 4. **Character-length differences are not measured token differences.** The existing perturbation changes wording and length jointly; flips do not identify a semantics-only effect. Adapter-exposed probabilities are not automatically logits. Unavailable option distributions, including Qwen's under its historical output contract, remain **NOT MEASURED**, not zero. 5. **Keep `semantic_equivalence_independently_approved: false`.** This disposition does not establish semantic equivalence or target-host acceptance. Preserve the existing paired-validity and no-imputation rules: missing or invalid paired observations yield no sidecar conclusion, not a partial favorable result. 6. **Do not invent implemented observables.** The inspected scorer reports pooled and gold-stratified flip counts, not a transition matrix. No new metric, transition-count implementation or numerical re-analysis is added or required here. ### Existing evidence and open gates Vesper's #91 review 192 is bound to `4472695d095d7d7665f8df1a098cf88ff5d09964`; her #89 review 203 is a `COMMENT` bound to `3cc5dbbbb77eebbb50efcd0cc37bacac0b52e619`; her #90 comment [2059](https://durandal.exe.xyz/smolmodelco/thesmolmodelcompany/issues/90#issuecomment-2059) records the alias-meaning objection. Retain all three. This addendum resolves the question-owner interpretation choice by retaining the vocabulary **without a semantic-invariance claim for any class**; it does not rewrite those reviews or claim a new independent approval. #89 remains a non-dispatchable intermediate: its criteria-deleted R4 must not enter a run. #91 preserves the corrected ordered canonical label→null-description shape, supported at the pinned source level, **not measured target-host acceptance**. Actual consuming-adapter verification, LM tokenizer traces/length-match obligations, runtime acceptance, model identity/environment, device admission and the existing independent-review/freeze requirements remain unresolved where applicable. Source-level review does not close them. **Scope:** append-only documentation on existing #90/#91 only. No code, payload, manifest, branch, PR body, review verdict, issue state or scientific ledger is changed. No new review dispatch, protocol freeze, merge, force-push, deletion, access change, experiment, model run/load/download, compute reservation/spend, private/raw trace publication, or reopening of execution gates. Existing capability-envelope policy is unchanged; no repeated per-run operator-approval requirement is introduced. Marathon inclusion remains tentative and outside this disposition.
Member

Selectively preserved by PR119, merged as a5b5e9274127. Preservation record binds original source head 4472695d095d7d7665f8df1a098cf88ff5d09964, Git blobs and SHA256 identities. The isolated DMS archive retains seven exact original files, qualified scientific projections and a manifest delta bound to original/base/main blobs. Ten unchanged canonical dependencies are reused. Review 192 is source-level evidence, not runtime readiness. Later comment 2093 governs interpretation: descriptive joint wording/length perturbation outside H1/H2, no semantic-invariance claim in any class. No inference or tokenizer validation occurred; private raw outputs remain excluded. Obsolete operational instructions were omitted from projections. Resulting main was fetched and verified: all 24 C1 and 17 DMS publication hashes match; all existing paths except the preservation index remain unchanged. Independent review, bounded regression fixtures and canonical CI passed. Closing this historical proposal as selectively preserved, without merging its old operational tree. Original branch, pull ref and discussion are retained. No scientific experiment was rerun.

Selectively preserved by [PR119](https://durandal.exe.xyz/smolmodelco/thesmolmodelcompany/pulls/119), merged as [`a5b5e9274127`](https://durandal.exe.xyz/smolmodelco/thesmolmodelcompany/commit/a5b5e927412715c180afc73234b704755d5f2566). [Preservation record](https://durandal.exe.xyz/smolmodelco/thesmolmodelcompany/src/commit/a5b5e927412715c180afc73234b704755d5f2566/docs/preservation/dms-pr91/README.md) binds original source head `4472695d095d7d7665f8df1a098cf88ff5d09964`, Git blobs and SHA256 identities. The isolated DMS archive retains seven exact original files, qualified scientific projections and a manifest delta bound to original/base/main blobs. Ten unchanged canonical dependencies are reused. Review 192 is source-level evidence, not runtime readiness. Later comment 2093 governs interpretation: descriptive joint wording/length perturbation outside H1/H2, no semantic-invariance claim in any class. No inference or tokenizer validation occurred; private raw outputs remain excluded. Obsolete operational instructions were omitted from projections. Resulting main was fetched and verified: all 24 C1 and 17 DMS publication hashes match; all existing paths except the preservation index remain unchanged. Independent review, bounded regression fixtures and canonical CI passed. Closing this historical proposal as selectively preserved, without merging its old operational tree. Original branch, pull ref and discussion are retained. No scientific experiment was rerun.
elsie closed this pull request 2026-10-18 04:06:13 +00:00
All checks were successful
checks / documentation integrity (pull_request) Successful in 6s
checks / E01 candidate surface sufficiency (pull_request) Successful in 5s
checks / DMS-003 protocol freeze integrity (pull_request) Successful in 6s
checks / DMS-003 CPU successor integrity (pull_request) Successful in 1m31s

Pull request closed

Sign in to join this conversation.
No reviewers
No labels
R03
research-b
No milestone
No project
No assignees
4 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/thesmolmodelcompany!91
No description provided.