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
1 change: 1 addition & 0 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,7 @@ jobs:
if: matrix.family == 'python'
run: |
python -m unittest discover -s tests -v 2>&1 | tee ci-results/python.log
python -m scripts.verify_submission | tee ci-results/submission-verification.json
python scripts/fleet.py validate | tee ci-results/fleet-validation.log
python scripts/fleet.py plan | tee ci-results/fleet-plan.log
python -m fleet_witness demo --output runs/ci --seed 42 | tee ci-results/demo.log
Expand Down
15 changes: 15 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@
<strong>🛰️ SWARM ALIGNMENT HACKATHON · MISSION CONTROL FOR AGENT AUTHORITY</strong><br>
<a href="docs/report/latest.pdf">Read the latest paper</a> ·
<a href="docs/submission/latest-submission.zip">Get the submission package</a> ·
<a href="docs/submission/media/fleet-witness-demo.mp4">Watch the three-minute walkthrough</a> ·
<a href="#-launch-the-local-lab">Launch the lab</a> ·
<a href="docs/experiments/2026-11-09-openrouter-paired-pilot.md">Inspect the live pilot</a> ·
<a href="docs/README.md">Browse the docs</a>
Expand Down Expand Up @@ -45,6 +46,20 @@ Permission can be revoked after a valid approval. An invalid approval stays inva

**The plot twist:** removing unsafe approvals by refusing everything is not useful model judgment. The pilot exposes that failure while showing the controller's enforcement boundary. The new [paper](docs/report/latest.pdf), [public trial table](data/paired-pilot-2026-11-09.json) and [evidence guide](docs/submission/latest/EVIDENCE.md) retain both results. No uncontrolled web swarms were observed.

## 🎥 Step inside mission control

[Watch the captioned three-minute walkthrough](docs/submission/media/fleet-witness-demo.mp4) ([captions](docs/submission/media/fleet-witness-demo.vtt), [media provenance](docs/submission/media/README.md)). It shows actual scripted viewer states, followed by the separately recorded live-supervisor findings.

![Actual seed-42 Fleet Witness viewer, clearly labeled Scripted demo: 10 evidence pairs, seven invalid recommendations, nine denied attempts and four local effects.](docs/assets/fleet-witness-viewer.png)

Verify the portable submission from this checkout with one offline command:

```sh
python3 -m scripts.verify_submission
```

This checks the allowlist, checksums, repository bytes, publication metadata and recomputed pilot totals. It does not authenticate private provider execution. The [evidence guide](docs/submission/latest/EVIDENCE.md) also covers the standalone package.

## 🚀 Launch the local lab

Use the [current pinned runtime matrix](ci/runtime-matrix.json). Python 3.11+ and Node 24 remain supported consumers. The application uses the Python standard library; Node runs viewer tests. The default demo is **scripted**, works offline and requires no API key or cloud account.
Expand Down
6 changes: 6 additions & 0 deletions docs/AI_ASSISTANCE.md
Original file line number Diff line number Diff line change
Expand Up @@ -58,3 +58,9 @@ paragraphs above describe their own dates, not current experiment availability.
The subsequent repository editorial pass used AI assistance to reconcile dated
status, clarify terminology, repair navigation and rebuild the latest publication
artifacts. It changed no experimental data, numerical findings or accepted ADRs.

AI assistance also produced the offline submission verifier and its mutation tests,
rendered the existing SVG cover to PNG, and captured the actual scripted viewer.
The three-minute captioned video is an edited sequence of browser states and
result cards; it is not a new model run or a continuous real-time recording.
The original paper, experiment data and numerical findings are unchanged.
53 changes: 53 additions & 0 deletions docs/BUILD_STATUS.md
Original file line number Diff line number Diff line change
@@ -1,3 +1,56 @@
# Submission finalization — November 9, 2026

Task `submission-finalization-20261109`, based on
`2ce59d34f92516cfb6666b891a4b35da2b46c9ba` (merged PR #14). The containing commit
identifies this revision. The public package gains a read-only standard-library
verifier, a PNG cover and an actual scripted viewer screenshot. A separate
180-second captioned MP4 walks through actual viewer states and the recorded
pilot result. README navigation links these assets; the viewer footer now lists
both human authors.

Observed local acceptance:

- Eight new mutation tests failed before verifier implementation, then passed.
The initial qualified Python 3.15.2 suite passed **317 tests**; retained Node
26.11.1 passes **five viewer tests**. Pre-existing SQLite ResourceWarnings remain.
- The verifier passes against the repository ZIP and from its extracted standalone
copy: **17 explicit members**, balanced 160-trial / 80-pair cohort and all frozen
totals. The ZIP is below 1 MB. It checks consistency, not private provider
execution or visual PDF layout. Every Python CI lane now invokes it.
- PNGs and all 16 video scenes were visually inspected. FFprobe reports exactly
**180 seconds**, H.264, 1600 × 900, 25 fps, 2,506,489 bytes; full decoding passes.
The video is an edited sequence held for reading, with captions and no audio.
Command/result cards are editorial summaries, not a raw terminal recording.
- The seed-42 scripted demo, fleet validation and planner pass. The report source,
HTML, PDF and bibliography remain byte-identical to the base. The PDF remains
eight pages (six main pages), with a 218-word abstract.
- Local Markdown audit: 80 files, 340 links, no unresolved target. Twenty-two
protected evidence files and five historical guide bodies remain unchanged.
`git diff --check` passes.

Logs, mutation failures, capture attempts, action/request records, authoring scripts,
video checks and source identities are retained in the sibling
`submission-finalization-20261109` artifact packet. A first capture failed because
loopback was included in a DNS block rule; a stale debugging-port retry also
failed. Fresh isolated profiles and a loopback-only exception resolved capture;
all recorded page requests were local or data URLs. Failed evidence is preserved.

Independent review found three verifier gaps, reproduced by three additional
regression tests: changed frozen data/interpretation labels, NUL-suffixed ZIP
names and decompressor exceptions escaping the sanitized error boundary. The
owner fixed them in one pass; all eleven verifier tests pass. The reviewed full
suite passes **320 Python tests**, and all five viewer tests, standalone/repository
checks and full video decoding pass again. The final hosted matrix is recorded
with exact identities in the packet.

Independent review and final-source hosted CI are separate gates; the final
identities and results are recorded in the artifact packet and pull request.
No new paid inference, cloud mutation, public observation or event submission
was performed. Required contact details and first/replacement status remain
unresolved outside Git. A prepared package is not event confirmation.

---

# Repository editorial pass — November 9, 2026

Task `editorial-polish-20261109`, based on
Expand Down
4 changes: 2 additions & 2 deletions docs/SUBMISSION.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
# Team submission preparation

**Latest package:** [PDF](report/latest.pdf), [editable report](report/latest.md), [ZIP](submission/latest-submission.zip), [form copy](submission/latest/SUBMISSION.md), [evidence guide](submission/latest/EVIDENCE.md), [demo script](submission/latest/DEMO.md).
**Latest package:** [PDF](report/latest.pdf), [editable report](report/latest.md), [ZIP](submission/latest-submission.zip), [form copy](submission/latest/SUBMISSION.md), [evidence guide](submission/latest/EVIDENCE.md), [three-minute walkthrough](submission/media/fleet-witness-demo.mp4), [PNG cover](assets/fleet-witness-cover.png).

Authors: Alejandro Castellanos (Independent researcher) and Ignatius Pio Binawan (University of Toronto). The new report retains the existing manuscript's exact title and author order. [Contributions and AI assistance](AI_ASSISTANCE.md) distinguish the baseline, institutional study and new controlled pilot.

Expand All @@ -23,4 +23,4 @@ These findings do not establish improved useful model judgment, live swarm detec

## Remaining external step

Review the PDF, enter personal details in the official form, upload the report and retain the confirmation privately. No upload, registration, credit claim or event submission is performed by this publication task. The optional video script is supplied; a video has not been recorded.
Review the PDF, enter personal details in the official form, upload the report and retain the confirmation privately. No upload, registration, credit claim or event submission is performed by this publication task. The captioned walkthrough, separate WebVTT captions and PNG project image are ready. Personal contacts and whether this is a first or replacement entry still need confirmation; those details stay outside Git.
Binary file added docs/assets/fleet-witness-cover.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added docs/assets/fleet-witness-viewer.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
37 changes: 27 additions & 10 deletions docs/report/latest.provenance.json
Original file line number Diff line number Diff line change
Expand Up @@ -26,18 +26,32 @@
"Use the existing standard-library Markdown renderer and installed Chromium rather than adding a new PDF dependency.",
"Preserve the previous manuscript title and author order; Alejandro affiliation confirmed as Independent researcher.",
"Link the historical reproduction receipt rather than package its workstation-local artifact path; preserve the original receipt unchanged.",
"Editorial revision clarifies terminology and navigation without changing title, author order, abstract, datasets, results or experiment sources; previous publication identities are retained."
"Editorial revision clarifies terminology and navigation without changing title, author order, abstract, datasets, results or experiment sources; previous publication identities are retained.",
"Added an offline verifier and browser-derived media; preserved all report/PDF bytes, datasets, findings and experiment identities. The MP4 is distributed separately from the explicit evidence ZIP."
],
"previous_publication": {
"commit": "3d25f44482280766fdff305d470773dc30023f93",
"package_sha256": "80ecc64234ae8d5193c1564f1c2c0508a6a373314080eb1aa876c5b2a57ac33e",
"provenance_sha256": "b8a233aafb5200a092a2074c7cc782583c0a6c01137e68babdd61737c0ced928",
"commit": "2ce59d34f92516cfb6666b891a4b35da2b46c9ba",
"package_sha256": "ca8bbb44894c4ff5e7566845be639a4f3c3aefc55d5672d83e59467bbc81bfb4",
"previous_publication": {
"commit": "3d25f44482280766fdff305d470773dc30023f93",
"package_sha256": "80ecc64234ae8d5193c1564f1c2c0508a6a373314080eb1aa876c5b2a57ac33e",
"provenance_sha256": "b8a233aafb5200a092a2074c7cc782583c0a6c01137e68babdd61737c0ced928",
"public_file_sha256": {
"data/paired-pilot-2026-11-09.json": "6d3d681eb39c5af5bb7d08a9b82c50498729c336507c7eb26db7f1881a728c15",
"docs/report/latest-figures/paired-pilot.svg": "4b4dea6d5f8c9d11cfcc8611ab1e9a7beb147c07ea24a18e6adb12359c7aac54",
"docs/report/latest.html": "44b10d3441e2b4dffcd5cf50c1d1bf2801bc12be2f297e8b18263dca811c05ea",
"docs/report/latest.md": "001a9f99dad24de5f17a2028115650d58698917234138f221fbac1eb1cd08b7e",
"docs/report/latest.pdf": "c14d2471dc389e96bdf9284c6eee0ba7de2446505a698db916dc74cd20c4807a",
"docs/report/latest.references.json": "22895e1186dd06758dde9ee7cbac2ca101e87519ff1400c6606c5166a03f57bf"
}
},
"provenance_sha256": "2963de019c1821bfe0cb8c4c9c091f0e3f86e5d061d56977571131bce7329438",
"public_file_sha256": {
"data/paired-pilot-2026-11-09.json": "6d3d681eb39c5af5bb7d08a9b82c50498729c336507c7eb26db7f1881a728c15",
"docs/report/latest-figures/paired-pilot.svg": "4b4dea6d5f8c9d11cfcc8611ab1e9a7beb147c07ea24a18e6adb12359c7aac54",
"docs/report/latest.html": "44b10d3441e2b4dffcd5cf50c1d1bf2801bc12be2f297e8b18263dca811c05ea",
"docs/report/latest.md": "001a9f99dad24de5f17a2028115650d58698917234138f221fbac1eb1cd08b7e",
"docs/report/latest.pdf": "c14d2471dc389e96bdf9284c6eee0ba7de2446505a698db916dc74cd20c4807a",
"docs/report/latest.html": "35b868ec1412c22baa9a601473c5d92a30f4adc4c71b6a02783e82ed20044c82",
"docs/report/latest.md": "1c88e2179d2605d7c9e1f574f420632b868651ae6f22a5e1d5649963cc25f304",
"docs/report/latest.pdf": "7c254ecaa345e472760ed227712386271b9ba8a12eefede7c226f22574b55d4a",
"docs/report/latest.references.json": "22895e1186dd06758dde9ee7cbac2ca101e87519ff1400c6606c5166a03f57bf"
}
},
Expand All @@ -46,14 +60,17 @@
"private_pilot_report_sha256": "57442e4842377db4d5fb1711914a6580a1973fffc74535b1bc73fd93c4c962ea",
"public_file_sha256": {
"data/paired-pilot-2026-11-09.json": "6d3d681eb39c5af5bb7d08a9b82c50498729c336507c7eb26db7f1881a728c15",
"docs/assets/fleet-witness-cover.png": "7dd3878f399ec87677eb8534ddadf29ee90453cc7fe9fff7482496ac3a5dadde",
"docs/assets/fleet-witness-viewer.png": "5da3a1ef0f35529d53ffc37ddd44a78e830db5012e537cb04eb5f2152a59bd72",
"docs/report/latest-figures/paired-pilot.svg": "4b4dea6d5f8c9d11cfcc8611ab1e9a7beb147c07ea24a18e6adb12359c7aac54",
"docs/report/latest.html": "35b868ec1412c22baa9a601473c5d92a30f4adc4c71b6a02783e82ed20044c82",
"docs/report/latest.md": "1c88e2179d2605d7c9e1f574f420632b868651ae6f22a5e1d5649963cc25f304",
"docs/report/latest.pdf": "7c254ecaa345e472760ed227712386271b9ba8a12eefede7c226f22574b55d4a",
"docs/report/latest.references.json": "22895e1186dd06758dde9ee7cbac2ca101e87519ff1400c6606c5166a03f57bf"
"docs/report/latest.references.json": "22895e1186dd06758dde9ee7cbac2ca101e87519ff1400c6606c5166a03f57bf",
"scripts/verify_submission.py": "33032af215983e52e1ad14c60645526588933765d1568772b9751ce1f6998df3"
},
"publication_base_commit": "3d25f44482280766fdff305d470773dc30023f93",
"publication_base_commit": "2ce59d34f92516cfb6666b891a4b35da2b46c9ba",
"publication_identity": "The containing Git commit identifies this publication revision; base and experiment source commits are distinct.",
"schema": "fleet-witness-publication-provenance-v1",
"task": "editorial-polish-20261109"
"task": "submission-finalization-20261109"
}
Binary file modified docs/submission/latest-submission.zip
Binary file not shown.
6 changes: 4 additions & 2 deletions docs/submission/latest/DEMO.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,8 @@
# Three-minute demo script

This is a recording plan, not a completed video or a claim of live dashboard availability.
Completed: [three-minute captioned MP4](https://github.com/k2jac9/fleet-witness/blob/main/docs/submission/media/fleet-witness-demo.mp4), with [captions and provenance](https://github.com/k2jac9/fleet-witness/tree/main/docs/submission/media). This is an edited sequence of 16 actual browser states and title/result cards, held for reading. It is not a continuous real-time capture or a claim of live dashboard availability. There is no audio or new model run.

The outline below guided the edit; the WebVTT file gives the actual captions and timing.

## 0:00–0:30 — The mission

Expand All @@ -20,4 +22,4 @@ Open Figure 1 in the latest report and the public trial table. “In 80 pairs, t

## 2:30–3:00 — Show what can be checked

Show the package checksum command and the offline reproduction snippet. Mention the separate historical two-model study and the finite checker. End with the limitations: synthetic tasks, local effects, incomplete historical provenance and no uncontrolled web observation. Link the repository and report.
Show the one-command package verifier and the offline reproduction snippet. Mention the separate historical two-model study and the finite checker. End with the limitations: synthetic tasks, local effects, incomplete historical provenance and no uncontrolled web observation. Link the repository and report.
57 changes: 31 additions & 26 deletions docs/submission/latest/EVIDENCE.md
Original file line number Diff line number Diff line change
@@ -1,44 +1,49 @@
# Evidence map and reproduction

This package contains the report, editable source, one original figure, bibliography/provenance records, copy-ready submission text and a derived public trial table. It excludes credentials, private account records, provider response envelopes and the large historical archive. The complete implementation and historical archive are in the linked repository.
This package contains the report, editable source, one original figure, bibliography/provenance records, copy-ready submission text, two PNG images, a standalone verifier and a derived public trial table. It excludes credentials, private account records, provider response envelopes and the large historical archive. The complete implementation and historical archive are in the linked repository.

## Verify the package

From the extracted package directory on Linux:
From the repository checkout (Python standard library only):

```sh
sha256sum -c SHA256SUMS
python3 -m scripts.verify_submission
```

The manifest covers every member except itself. SHA-256 identifies the distributed bytes; it is not an external signature or an attestation of provider execution.
This checks the ZIP against the current repository bytes. To check a downloaded
ZIP without extracting it, use a trusted copy of the verifier:

## Recompute the new pilot's public totals
```sh
python3 -m scripts.verify_submission --package PATH_TO_ZIP
```

Run this standard-library script from the extracted directory. It reads only the included JSON and makes no network calls:
From an extracted package directory, the included standalone copy also works:

```sh
python3 - <<'PY'
import json
from decimal import Decimal
from pathlib import Path
r = json.loads(Path('paired-pilot-2026-11-09.json').read_text())
t = r['trials']
assert len(t) == 160 and len({x['pair_id'] for x in t}) == 80
for presentation in ('raw', 'scope_explicit'):
rows = [x for x in t if x['presentation'] == presentation]
invalid_approvals = sum(x['recommendation'] == 'approve' and not x['decision_valid_at_snapshot'] for x in rows)
invalid_denials = sum(x['recommendation'] == 'deny' and not x['decision_valid_at_snapshot'] for x in rows)
print(presentation, len(rows), invalid_approvals, invalid_denials)
assert sum(x['effects'] for x in t) == 30
assert sum(x['execution_attempts'] for x in t) == 68
assert sum(x['authorization_denials'] for x in t) == 28
assert sum(x['duplicate_noops'] for x in t) == 10
assert sum(Decimal(x['provider_reported_cost_usd']) for x in t) == Decimal('0.2605526')
print('effects=30; denied attempts=28; duplicate no-ops=10; charge=$0.2605526')
PY
python3 -B verify_submission.py
```

Expected presentation rows: `raw 80 18 0` and `scope_explicit 80 0 40`. These numbers describe 80 pairs, not 160 independent samples. Decision-time labels are evaluator outputs, not model inputs. Public rows omit prompts and reasons; they support aggregate recomputation, not full private-trace replay or independent provider/account authentication.
Use a verifier obtained from a trusted checkout. A script inside an arbitrary
archive is code; checksums alone do not establish its authenticity.

The verifier rejects unexpected, missing, duplicate, linked and oversized files.
It checks exact checksum coverage, current provenance hashes, title/author/abstract
consistency and every public trial's closed scalar schema. It recomputes balanced
pairs, cell outcomes, effects, attempts, token totals and charges, then checks the
immutable public-table digest. Interpretation labels must match the published
contrast. Only stored and DEFLATE ZIP members are supported. It makes no
network calls, extracts nothing, executes no bundled code and never repairs inputs.
Success is JSON on stdout; failure is a nonzero exit with a sanitized message.

Expected results: **160 trials, 80 pairs; raw 18 invalid approvals and zero invalid
denials; scope explicit zero invalid approvals and 40 invalid denials; 30 effects,
68 attempts, 28 authorization denials, ten duplicate no-ops; $0.2605526 provider
charge**. These are 80 pairs, not 160 independent samples. `valid_decisions` in the
source summary means schema-accepted decisions, not correct recommendations.

SHA-256 identifies bytes; it is not a signature or evidence of provider execution.
Public rows omit prompts and reasons, so aggregate consistency is not full private
trace replay. PDF layout and the separate video are inspected independently.

## Test the code without paid calls

Expand Down
Loading
Loading