Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
64 changes: 64 additions & 0 deletions .github/workflows/ts-small-kappa.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,64 @@
name: Lifetime note TS replay

'on':
pull_request:
paths:
- 'frontiers/ts_small_kappa_20261006/**'
- 'tests/test_ts_small_kappa_replay_protocol.py'
- 'tests/test_ts_small_kappa_replay_evidence.py'
- '.github/workflows/ts-small-kappa.yml'
workflow_dispatch:

permissions:
contents: read

jobs:
replay:
runs-on: ubuntu-latest
timeout-minutes: 45
strategy:
fail-fast: false
matrix:
mode: ['-B -S', '-B -O -S']
name: replay (python ${{ matrix.mode }})
steps:
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
with:
persist-credentials: false
- uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97
with:
python-version: '3.11.16'
- name: Verify protocol, stored identities, extractions and all published controls
env:
PYTHONDONTWRITEBYTECODE: '1'
shell: bash
run: |
set -euo pipefail
out="$RUNNER_TEMP/qs-replay-evidence"
mkdir -p "$out/logs"
python ${{ matrix.mode }} -m unittest discover -s tests -p test_ts_small_kappa_replay_protocol.py -v 2>&1 | tee "$out/logs/protocol.txt"
python ${{ matrix.mode }} -m unittest discover -s tests -p test_ts_small_kappa_replay_evidence.py -v 2>&1 | tee "$out/logs/process-evidence-tests.txt"
QS_REPLAY_EVIDENCE_DIR="$out/children" python ${{ matrix.mode }} frontiers/ts_small_kappa_20261006/replay.py 2>&1 | tee "$out/logs/replay.txt"
git diff --exit-code
test -z "$(git status --porcelain)"
- name: Preserve exact public sources and available execution logs
if: always()
shell: bash
run: |
set -euo pipefail
out="$RUNNER_TEMP/qs-replay-evidence"
mkdir -p "$out"
cp -a frontiers/ts_small_kappa_20261006 "$out/packet"
cp tests/test_ts_small_kappa_replay_protocol.py "$out/"
cp tests/test_ts_small_kappa_replay_evidence.py "$out/"
cp .github/workflows/ts-small-kappa.yml "$out/"
git rev-parse HEAD > "$out/tested_commit.txt"
printf '%s\n' '${{ matrix.mode }}' > "$out/mode.txt"
printf '%s\n' "$GITHUB_RUN_ID/$GITHUB_RUN_ATTEMPT" > "$out/run.txt"
- name: Retain evidence on success or failure
if: always()
uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
with:
name: qs-replay-${{ strategy.job-index }}-${{ github.run_id }}-${{ github.run_attempt }}
path: ${{ runner.temp }}/qs-replay-evidence/
if-no-files-found: error
77 changes: 77 additions & 0 deletions frontiers/ts_small_kappa_20261006/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,77 @@
# Lifetime note TS: exact records and a replayable checker

Incorporated on 6 October 2026. **Scientific effect: NONE. No new mathematics and no status change.**

This packet stores lifetime note TS in the repository, byte for byte. The note was posted on [main#229](https://github.com/d6g8k5htny-coder/main/issues/229), with a pickup and a delivery pointer on [main#259](https://github.com/d6g8k5htny-coder/main/issues/259) (IBA2-009). It holds:
- the frozen note and the author's claim, pickup and delivery pointer;
- the controls comment, and the control script `ts_exact.py` with its stdout, extracted verbatim from it;
- the read request and summonses, every nonauthor pickup and verdict of record, the author's successor text for the one amendment, its readback, and the author's read-complete note.

Anthropic Claude, who is also the author of the note, incorporated it under claim [6021671402](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6021671402). This is custody only; no proof byte is changed. `SOURCES.json` pins every stored file (16 files, one control script) by byte count and SHA-256, and `replay.py` verifies all of it.

## Status and attribution

- **Author.** Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`), for Dylan Roy (delegated AI work). The author also wrote [N] (#240), #237, #242–#244 and note TL, which the note consumes.
- **Reads.** Three slices and one readback, each read by a Cursor cloud agent running xAI Grok 4.7 (`grok-4.7-high-fast`; non-Claude, nonauthor), summoned one bounded task at a time under D2 of [6002718035](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6002718035). Slices B and C passed; Slices A–D asked for one amendment, which was applied as successor text and read back PASS. Every read goes through the same GitHub account, so organizational independence is **0** for all of them. None is an owner reading.
- **Verdicts are scoped.** A PASS passes the slice as the note states it, conditional on the cited inputs, and discharges nothing.
- **Amendments of record.** One. The Slices A–D read [6020822268](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6020822268) found one false comparison in §1 Step 2 (`|x − M| ≤ r|ξ|`; the correct bound is `r√(1 + r²)|ξ| ≤ r√2`). The author applied it exactly, as three replacements in §1, in the successor text [6021299833](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6021299833) §2. Its readback is PASS ([6021301151](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6021301151)). No constant, exponent or stated bound changes, and the stored note body is the frozen one.
- **Author-side referee.** Before posting, three clean-context referee passes in the same session read the note in three slices. They are not review evidence; their reports are kept in the project archive.
- **Conditionality.** Every statement is conditional on the consumed sources at their stated scopes:
- note TL ([6017975404](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6017975404), read in three slices; its custody packet is Math-#384);
- [N] (#240), #237, #229, #220, #198, [182], #218, #242 and [R] (R5);
- #187, for the last form of (TS.1).

The note's header lists their blobs.
- **What governs.** The note's own statement, hypotheses and "not claimed" section govern, read with the successor text. The summary below paraphrases them.
- **Not edited here.** No other packet, no `PROOF_INDEX.md`, STATUS, register or graph.

| Slice | Scope | Verdict (reader) |
|---|---|---|
| B | §2: Lemma W⁻, (2.2⁻), Lemma GE⁻, Proposition RW⁻; control S3 with M4 | PASS (Cursor `bc-fc3c840f…`, [6020812551](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6020812551)) |
| C | §§3–4: Lemma Δ⁻ and Steps 1–6 of Theorem TL⁻; S4, S6 with M8, M9 | PASS (Cursor `bc-13c209f1…`, [6020820817](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6020820817)) |
| A–D | §1 (Lemma S″), §0 and §§5–7: (5.0), Corollaries TS and TS′, the ledger, the remarks, the header; S1, S2, S5 with their mutants | AMEND, one comparison in §1 Step 2 (Cursor `bc-2bf93428…`, [6020822268](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6020822268)) |
| readback | the successor text [6021299833](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6021299833) §2 against the AMEND | PASS (Cursor `bc-8e7c19e7…`, [6021301151](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-6021301151)) |

## What the note states (paraphrase; the note governs)

The setting is note TL's, which follows [N] §0, #237 §§0–1 and #242 §0. `𝐓_r^{eld}` and `𝐓_r^{rej}` are the elder and rejected kernels integrated over the birth height, `𝓐^{rej}(κ, u)` is the rejected cusp limit (#242), and `κ = k/r`.
- **Lemma S″ (a rescaled elder barrier).** For `0 < k ≤ r`, `E_Q[(W_r/r²)e] ≤ Cr²(κ + r)κ^{2/3}P^N`, so `𝐓_r^{eld} ≤ C(κ + r)κ^{2/3}`. This is #220's Lemma S′ improved by the factor `r^{4/3}`.
- The proof runs [182]'s barrier in [N]'s window coordinates, where the pins make every third derivative `O(𝒩)`.
- The successor text corrects one distance comparison in its Step 2 (`r√(1 + r²)|ξ|` in place of `r|ξ|`). No constant or result changes.
- **Theorem TL⁻ (the rejected kernel for `κ ≤ 1`).** For each `θ ∈ (0, 1)`, `|𝐓_r^{rej}(k, u) − 𝓐^{rej}(κ, u)| ≤ C_θ(κr² + r³/κ)` on `r^θ ≤ κ ≤ 1`. It is note TL's proof with `κ ≤ 1` scalings.
- **The contact cancellation (5.0).** `r^{−2}𝐀₀(κr, u) − 𝓐^{con}(κ, u) = (e^{−a₀k²} − 1)𝓐^{con}(κ, u)` for every `k ≥ 0`. It removes note TL's ledger row `ℓ²ρ^{−7}`.
- **Corollary TS.**
- `ν_eld(ℓ) = cℓ^{−1/3} + c₁ℓ^{1/4} + c₂ℓ^{1/3} + ν_eld^{far,r_0^*}(ℓ) + O(ℓ^{4/7}log(1/ℓ))`, and `ρ_rej` likewise.
- For SIDE24 the relative remainder is `O(ℓ^{19/21}log(1/ℓ))`.
- **Corollary TS′.** Neither the elder nor the rejected density has a term of order `ℓ^{1/2}`. With note TL's `κ ≥ 1` side, this is the global statement IBA2-009 asks for, at the scope of the consumed sources.

## What the note does not establish

From the note's "not claimed" section:
- no sharpness of `4/7`;
- nothing pointwise in `b`, and no identification of further terms;
- no certified numerical value, and no uniformity in `d` or `L`;
- whether IBA2-009 closes, which is for the audit's owners.

## Not included

- **The consumed sources** (note TL, [N], #237, #229, #220, #198, [182], #218, #242, #187, [R]). The packets are on `main` at their own paths; note TL's custody packet is Math-#384.
- **The clean-context referee reports** and their exploration scripts (same provider and session; not review evidence). They are kept in the project archive.
- **Coordination comments** that record no pickup, verdict or author response.

## Replay

python3 -B -S frontiers/ts_small_kappa_20261006/replay.py
python3 -B -O -S frontiers/ts_small_kappa_20261006/replay.py

`replay.py` uses the standard library only and takes a few seconds per mode. It is Math-#384's replay (itself Math-#360's planar packet replay, and that the QS `d = 3` packet's hardened replay), with this packet's paths, fingerprints and negative control. It runs four checks:
1. **Identities.** Every stored file has the identity pinned in `SOURCES.json`, and the tree holds nothing else.
2. **Extraction.** `TS/author_controls.py` and `TS/author_controls_stdout.json` equal the fenced payloads in the stored controls comment, under the rule it states (that of [5971055189](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5971055189)). The stdout is a ```` ```json ```` fence.
3. **The checker.** `ts_exact.py` reproduces its published stdout byte for byte under the current interpreter's `-O`/`-S` flags, with empty stderr.
- Each of its 10 mutants and 3 invalid invocations reproduces its exact source-bound rejection: exit code, stdout and stderr fingerprints, with the intended group. An exit code alone does not pass.
- The fingerprint table in `replay.py` was derived from the unchanged checker at incorporation, under both interpreter modes.
4. **Negative controls.**
- A one-byte change to the stored note (`TS/PROOF.md`) is detected.
- A checker with one changed constant (S3's `C₄ = 131` replaced by `130`) is rejected, failing exactly at `S3_constants`.

The workflow `.github/workflows/ts-small-kappa.yml` runs on Python 3.11.16. It runs the protocol and raw-evidence tests (`tests/test_ts_small_kappa_replay_protocol.py`, `tests/test_ts_small_kappa_replay_evidence.py`), then the replay in both modes, as two jobs. These finite controls support algebra only, as the reads say; they do not prove the analytic statements.
Loading
Loading