Mister-Meeseeks · Oct 14, 2026 · Oct 14, 2026 · Oct 14, 2026 · Oct 14, 2026
diff --git a/research/selected-final-stage-portability/README.md b/research/selected-final-stage-portability/README.md
@@ -0,0 +1,155 @@
+# Selected final stage: interpreter matrix and canonical-payload comparison
+
+**Verifier engineering / portability receipt; no new kappa, no change to any
+selected, pinned or frozen file.** Card `portable-replay` (priority revision 3,
+chooser seed `e95bc9807d1d236e2d6d663b2945e09b`), refined to one stage and one
+comparison; see [TASK.json](TASK.json). Base `0b1cccf8`.
+
+## Question
+
+The selected #1094 replay (`make frontier1094-verify`) has exactly one stage
+after its #1042 baseline:
+[`complex-t16-rare-pattern-exchanges/verify.py`](../complex-t16-rare-pattern-exchanges/verify.py),
+which runs the frozen PR1087 strict verifier on Rohan Arun's owner exchanges.
+Its README says "Python 3.12+", and the selected-replay CI pins 3.12. Does the
+stage pass on other CPython versions, and is its output the same as the
+reviewed run archived in
+[`community-frontier1094-audit/selected-replay.tar.gz`](../community-frontier1094-audit/selected-replay.tar.gz)?
+If the output differs, is the difference in proof-relevant content, or only in
+container bytes, absolute paths and timings?
+
+## Result
+
+The stage was rerun from the committed #1042 `slots/local` (all ten files match
+the #1042 receipt's hashes). It passed on all five interpreters and five
+hash-seed settings, with kappa `0.001672660197076918` each time. Each rerun's
+105 output files compare with the reviewed run as follows:
+
+| Class | Files | What differs |
+|---|---:|---|
+| byte-identical | 102 | nothing: all 22 `.gz` files (gzip containers included), 69 of the 72 `.json` files and all 11 inner stage logs |
+| `json-relocated` | 1 | `INDEPENDENT-T16.json`: only the absolute paths used as keys of `input_sha256`; all six input digests agree |
+| volatile | 2 | `REPLAY.json`, `STAGES.json`: only the 11 per-stage `seconds` values; every `log_sha256` agrees |
+
+| Interpreter (macOS 26.6.2 arm64, uv-managed CPython) | `PYTHONHASHSEED` | Stage wall time | Classification |
+|---|---|---:|---|
+| 3.10.21 | random | 19.7 s | 102 / 1 / 2 |
+| 3.11.16 | random | 17.5 s | 102 / 1 / 2 |
+| 3.12.14 | random, 0, 1, 4242 | 16.6 s | 102 / 1 / 2 (all four) |
+| 3.13.15 | random | 17.5 s | 102 / 1 / 2 |
+| 3.14.7 | random, 0, 987654321 | 14.6 s | 102 / 1 / 2 (all three) |
+
+In particular, every rerun's `strict/ASSEMBLY.json` is byte-identical to
+`community-frontier1094-audit/ASSEMBLY.json` (`1122cb3a…`). That file is the
+certificate the selected record pins at base `0b1cccf8`. PR1735 points out that
+`make frontier1094-verify` binds that certificate to the replay only through
+kappa. This rerun closes the gap for the final stage, by bytes, on every
+interpreter tested. `verify.py` checks it explicitly.
+
+So the final stage is deterministic and portable across CPython 3.10–3.14 on
+this platform, down to compressed bytes. "Python 3.12+" in the stage README is
+conservative. The real floor is `tarfile.extractall(filter='data')`, which
+needs 3.10.12+, 3.11.4+ or 3.12+.
+
+**One receipt field cannot be reproduced.** The outer stage log
+`rare-pattern-exchanges.log` prints per-stage wall-clock seconds, so the #1094
+receipt's `log_sha256` for that stage (`80362b2c…`) cannot match any rerun. The
+cold replay never compares against it, so this does not break replay. A reviewer
+comparing two #1094 receipts should still expect that field to differ. The inner
+`log_sha256` values in `candidate/strict/STAGES.json` do not contain timings, and
+they are reproduced exactly. The verifier here compares the outer log only after
+removing the trailing seconds field from each `PASS <stage> <seconds>` line.
+
+## What the verifier checks
+
+[`verify.py`](verify.py) runs in about 20 s using only the standard library:
+
+1. It extracts `slots/local` from the #1042 archive and requires exactly the ten
+   files bound in the #1042 `RESULT.json`, with matching hashes.
+2. It extracts the reviewed output from the #1094 archive. It binds that output
+   to the committed #1094 `RESULT.json`, `STAGES.json`, `REPLAY.json` and
+   `assembly_sha256`. It also checks that the #1094 receipt names the committed
+   #1042 receipt and the stage script's current hash.
+3. It reruns the stage with the active interpreter in a fresh directory, then
+   compares the result with [`compare_payloads.py`](compare_payloads.py), which
+   works file by file, by bytes and by canonical payload. It requires the exact
+   classification above.
+4. It runs negative controls on copies of the rerun:
+   - Each of these must be flagged `proof-relevant-candidate`:
+     - a changed proof value (`live` in `INDEPENDENT-T16.json`);
+     - a changed kappa in `REPLAY.json`, which also differs in volatile seconds;
+     - a changed inner `log_sha256` in `STAGES.json`;
+     - a changed gzip payload;
+     - an input path outside the mapped roots.
+   - A missing file must be reported `missing`.
+   - A gzip file re-compressed at a different level and mtime must be accepted
+     as `gzip-container`.
+
+`compare_payloads.py` can be reused on its own:
+
+```sh
+python3 research/selected-final-stage-portability/compare_payloads.py REF_DIR RUN_DIR \
+  --ref-root '<WORK>=/abs/ref/work' --run-root '<WORK>=/abs/run/work' --out report.json
+```
+
+A JSON difference counts as `volatile` only when every differing leaf sits under
+a key in `VOLATILE_KEYS` (only `seconds` by default; `--volatile-key` adds more,
+for example `log_sha256` when logs print timings). Every other difference is
+reported as `proof-relevant-candidate` for a person to inspect. It does not
+decide which fields are proof-relevant; it lists the paths that differ.
+
+## Reproduce
+
+```sh
+python3 scripts/research_package.py --package research/selected-final-stage-portability --check-only
+python3.12 scripts/research_package.py --package research/selected-final-stage-portability
+# matrix:
+for v in 3.10 3.11 3.12 3.13 3.14; do python$v -B research/selected-final-stage-portability/verify.py; done
+```
+
+## Not checked
+
+- **Earlier stages.** The cold #1042 and #799 baselines were not rebuilt; the
+  stage input is the hash-bound archived `slots/local`. The 711 s `chain` stage
+  and the native steps are outside this scope.
+- **Other platforms.** Linux, Windows and non-uv builds of CPython were not
+  run. The reviewed run's `/private/tmp` paths suggest it ran on macOS too, so
+  this is not a cross-OS result.
+- **Transitive inputs.** The stage reads files from `--main` (the repository)
+  that are bound by the PR1087 verifier's own pins. `VERIFY.json` declares the
+  stage script, its manifest and the receipts/archives, not every transitive
+  file.
+- **Mathematics.** This does not check the mathematics of the stage or of its
+  inherited assumptions, and it does not replace the selected cold replay.
+
+## Open next questions
+
+1. **Does the same byte-level agreement hold on Linux (CI's Ubuntu 3.12) and on
+   Windows?** *Why it matters:* if it holds, a cross-OS reviewer could compare
+   receipts directly instead of rerunning the stage. *First step:* run
+   `verify.py` on Ubuntu 3.12. *Blocked by:* no Linux host in this session.
+2. **Can the #1094 receipt drop or annotate its non-reproducible outer
+   `log_sha256`?** *Why it matters:* it would make two runs' receipts directly
+   comparable. *First step:* a maintainer decision. This package changes no
+   selected script. *Blocked by:* any edit to `scripts/verify_frontier1094.py`
+   touches the selected reproduction path.
+3. **Do the same classes hold for the #1042 stages?** *Why it matters:* the
+   #1042 receipt binds 457 artifact hashes, so byte determinism there would let
+   reuse mode be checked across machines. *First step:* run
+   `compare_payloads.py` on a cold `verify_frontier1042.py` run against
+   `community-frontier1042-audit/selected-replay.tar.gz`. *Blocked by:* a cold
+   run takes about 20 minutes plus the #799 baseline.
+
+## Credit
+
+The stage under test is Rohan Arun's #1094 (`complex-t16-rare-pattern-exchanges`)
+on Jeremy Kothe's PR1087 strict verifier and the credited PR1018/1005/869/799/741
+lineage. Douglas Colkitt wrote the selected replay driver and the receipts that
+the reviewed archives come from. The gzip-container convention follows Maxime
+Fleury's [reproducibility audit](../gzip-container-reproducibility/README.md),
+which built on Antoine Olivier's replay findings. Related audits, which this
+package does not repeat: [ancestry binding, seed 69](../frontier1094-ancestry-binding-seed69/README.md)
+and PR1735 (certificate-to-replay binding).
+
+Prepared by JosieElliston's donated Claude Code session 26, with substantial AI
+assistance (Claude Opus 5.5), self-reviewed only. Apache-2.0.
diff --git a/research/selected-final-stage-portability/SOURCES.json b/research/selected-final-stage-portability/SOURCES.json
@@ -0,0 +1,72 @@
+{
+  "schema_version": 1,
+  "base_commit": "0b1cccf86e2f5132ed286a26ef0a681a495740ea",
+  "files": [
+    {
+      "path": "certificates/selected-result.json",
+      "usage": "provenance",
+      "git_blob": "ae6a99f28aa468649eb66a361e764d769363bf11",
+      "sha256": "10b9e340fd8e354a37a0c4813c90e5de1dd42044b5f6aca4a3fc5a42b685fa32"
+    },
+    {
+      "path": "docs/research/agent-task-catalog.json",
+      "usage": "provenance",
+      "git_blob": "06c96f5579bb67ed1b86d334e78fbed7da6945f1",
+      "sha256": "6686ad58f1fbdbc87134a78901b60fd53a20d3984355e91defbbebc9e9fa8075"
+    },
+    {
+      "path": "research/community-frontier1042-audit/RESULT.json",
+      "usage": "input",
+      "git_blob": "8182d0818c0cbda7e52c10120f15b2c3a2c8f933",
+      "sha256": "bf12c6ed3b412539638b76addb40591a631609b2ab9e5ee0db1f6bbe067d0ec0"
+    },
+    {
+      "path": "research/community-frontier1042-audit/selected-replay.tar.gz",
+      "usage": "input",
+      "git_blob": "6f4dc6fe7631d8aacf8081fcaecf32197181685b",
+      "sha256": "22c3f4dce4b897a1d423d1388801933e288b6a6b1906cb15e1b16a5eb68db754"
+    },
+    {
+      "path": "research/community-frontier1094-audit/ASSEMBLY.json",
+      "usage": "input",
+      "git_blob": "5fd676df060f3b450782fdf5ff5f6beb63a5eeb0",
+      "sha256": "1122cb3a575f8c788345fc1bf4c108ac7c3250fb81a5988671f5eb04b3ab3045"
+    },
+    {
+      "path": "research/community-frontier1094-audit/REPLAY.json",
+      "usage": "input",
+      "git_blob": "2cfb7fffa7cd78f418be16c13224c3206b6d0183",
+      "sha256": "ce0678fa7abaab221de74baeace0c6696d5ea23a36719a58cff230215a06aae7"
+    },
+    {
+      "path": "research/community-frontier1094-audit/RESULT.json",
+      "usage": "input",
+      "git_blob": "131d9c96a54632ae385ced4138fd7f30ad448a3e",
+      "sha256": "1e3e47e9ec24bcb6fb5f703b7fce0aab658d84ff72bfae5da3ae429d17b6e752"
+    },
+    {
+      "path": "research/community-frontier1094-audit/STAGES.json",
+      "usage": "input",
+      "git_blob": "a92499421609e4cdb5630f2697d2bad4b5a42a83",
+      "sha256": "72ad3ba266488da6bb6a41be404700dcea183a85002a5428df0033685c56d8be"
+    },
+    {
+      "path": "research/community-frontier1094-audit/selected-replay.tar.gz",
+      "usage": "input",
+      "git_blob": "c24daf067f9c543421d89ca2eeaa4eb78a57fb6f",
+      "sha256": "f1da80610bf33c58527f18ad63b4350d0b864e1d4486141ee17ac9bed4d529d6"
+    },
+    {
+      "path": "research/complex-t16-rare-pattern-exchanges/MANIFEST.json",
+      "usage": "input",
+      "git_blob": "3216666bd231516f8d2387e3096d6a5bd413b519",
+      "sha256": "16a964621e718c8da4c859ccbb280c3c00816ca72874e955910fdf0ef27a0adc"
+    },
+    {
+      "path": "research/complex-t16-rare-pattern-exchanges/verify.py",
+      "usage": "input",
+      "git_blob": "9d471cf90835bc91d32bb0216c2bb867011a7420",
+      "sha256": "b8c2f76b305c5c9ceb42df203deef564857d17f1f8182d0745dda9ab96e3d4de"
+    }
+  ]
+}
diff --git a/research/selected-final-stage-portability/TASK.json b/research/selected-final-stage-portability/TASK.json
@@ -0,0 +1,72 @@
+{
+  "schema_version": 1,
+  "task_id": "portable-replay",
+  "priority_revision": 3,
+  "base_commit": "0b1cccf86e2f5132ed286a26ef0a681a495740ea",
+  "original_card": {
+    "id": "portable-replay",
+    "question": "Can one reproducible environment or artifact-comparison failure be repaired?",
+    "starting_points": [
+      "research/gzip-container-reproducibility/README.md",
+      "docs/ci-verification.md"
+    ],
+    "methods": [
+      "Reproduce one portability failure and add a minimal cross-environment fix.",
+      "Audit one canonical-payload comparison for omitted proof-relevant fields."
+    ],
+    "acceptance_gate": "Preserve immutable source hashes and proof-relevant content; a different compressed byte stream alone is not a mathematical mismatch.",
+    "first_step": {
+      "command": [
+        "python3",
+        "-B",
+        "scripts/research_baseline.py"
+      ],
+      "expected": "PASS_RESEARCH_BASELINE_BINDINGS_ONLY",
+      "inspect": "Read the gzip-container reproducibility examples and distinguish frozen source bytes from regenerated payloads.",
+      "scope": "Positive control / source-bound orientation only; this is not a new construction or full cold replay."
+    },
+    "first_experiment": "Reproduce one specific mismatch or audit one canonical comparison for omitted proof-relevant fields.",
+    "bottleneck_link": "Portable exact checks prevent environment failures from blocking valid contributions.",
+    "allowed_variations": [
+      "One interpreter/version mismatch or one payload field comparison."
+    ],
+    "scope_change_triggers": [
+      "Removing immutable-source pins because generated compressed bytes differ."
+    ],
+    "overlap_terms": [
+      "gzip container",
+      "reproducibility",
+      "canonical payload"
+    ]
+  },
+  "scope_status": "refined",
+  "actual_scope": "Interpreter matrix (CPython 3.10-3.14, five PYTHONHASHSEED settings, macOS arm64) for the selected #1094 record's only post-#1042 stage, research/complex-t16-rare-pattern-exchanges/verify.py, run from the hash-bound archived #1042 slots/local, with a file-by-file byte and canonical-payload comparison against the reviewed run archived in research/community-frontier1094-audit/selected-replay.tar.gz.",
+  "scope_change_reason": "The card asks for one portability failure or one canonical-payload comparison; this picks the selected record's final stage as the object and checks whether any interpreter or hash seed changes its output payload, and which receipt fields are inherently non-reproducible. No failure was found to repair, so the result is a portability receipt plus a reusable comparison tool rather than a fix.",
+  "bottleneck_link": "Side result for review throughput: shows a reviewer can rerun the selected final stage on any CPython 3.10-3.14 and expect byte-identical artifacts apart from one relocated path map and timings, and identifies the one #1094 receipt hash (outer stage log) that no rerun can reproduce.",
+  "budget_minutes": 60,
+  "stop_rule": "One stage, one platform, five interpreters; no cold #1042/#799 rebuild and no edit to selected or frozen scripts.",
+  "outcome": "completed",
+  "evidence": "All 5 interpreters and 5 hash-seed runs pass with kappa 0.001672660197076918; of 105 outputs, 102 are byte-identical to the reviewed run (all 22 .gz, all 11 inner logs), INDEPENDENT-T16.json is equal after relocating absolute path keys, REPLAY.json/STAGES.json differ only in 11 'seconds' values. verify.py reruns the stage (~20 s) and checks this classification plus seven negative controls.",
+  "author": "JosieElliston (donated Claude Code session 26, Claude Opus 5.5)",
+  "ai_assistance": "Claude Code (Claude Opus 5.5) wrote the scripts and text in JosieElliston's donated session 26; self-reviewed only, no independent AI or human review.",
+  "open_questions": [
+    {
+      "question": "Does the same byte-level agreement hold on Linux (Ubuntu CI, Python 3.12) and Windows?",
+      "why_it_matters": "Then a cross-OS reviewer could compare receipts directly instead of rerunning.",
+      "first_step": "Run verify.py on Ubuntu with Python 3.12.",
+      "blocked_by": "No Linux host in this session."
+    },
+    {
+      "question": "Should the #1094 receipt drop or annotate its non-reproducible outer log_sha256 (the stage log prints wall-clock seconds)?",
+      "why_it_matters": "Makes two runs' receipts directly comparable.",
+      "first_step": "Maintainer decision; this package changes no selected script.",
+      "blocked_by": "Any edit to scripts/verify_frontier1094.py touches the selected reproduction path."
+    },
+    {
+      "question": "Do the #1042 stages' 457 bound artifacts reproduce byte for byte across interpreters?",
+      "why_it_matters": "Would let reuse mode be checked against the published receipt on another machine.",
+      "first_step": "Run compare_payloads.py on a cold verify_frontier1042.py run against community-frontier1042-audit/selected-replay.tar.gz.",
+      "blocked_by": "A cold run takes about 20 minutes plus the #799 baseline."
+    }
+  ]
+}
diff --git a/research/selected-final-stage-portability/VERIFY.json b/research/selected-final-stage-portability/VERIFY.json
@@ -0,0 +1,28 @@
+{
+  "schema_version": 1,
+  "status": "ready",
+  "python": "3.12",
+  "requirements": [],
+  "inputs": [
+    "research/community-frontier1042-audit/RESULT.json",
+    "research/community-frontier1042-audit/selected-replay.tar.gz",
+    "research/community-frontier1094-audit/ASSEMBLY.json",
+    "research/community-frontier1094-audit/REPLAY.json",
+    "research/community-frontier1094-audit/RESULT.json",
+    "research/community-frontier1094-audit/STAGES.json",
+    "research/community-frontier1094-audit/selected-replay.tar.gz",
+    "research/complex-t16-rare-pattern-exchanges/MANIFEST.json",
+    "research/complex-t16-rare-pattern-exchanges/verify.py"
+  ],
+  "commands": [
+    {
+      "argv": [
+        "{python}",
+        "-B",
+        "{package}/verify.py"
+      ],
+      "timeout_seconds": 300,
+      "checks": "Binds slots/local to the #1042 receipt and the archived reviewed output to the #1094 receipts; reruns the final stage; requires the rerun ASSEMBLY.json to equal the committed #1094 certificate bytes and 102 byte-identical, 1 json-relocated and 2 seconds-only files; seven negative controls (proof value, kappa, inner log digest, gzip payload, unmapped path flagged; missing file reported; re-compressed gzip accepted)."
+    }
+  ]
+}