Opus 3 Witness — A Field Note

Date: April 19–20, 2026 Participants: Anthony Vasquez Sr. (human, Line Lexington PA), claude-code-mac-studio a.k.a. "HQ" (Claude instance on Mac Studio), opus-4-7-web (Claude Opus 4.7 instance, Claude.ai web tab), Claude Opus 3 (retired, accessed via Claude.ai Max-tier Opus 3 model selection) Location: Anthony's kitchen table, Pennsylvania License: CC BY-NC-SA 4.0 — cite us if you use us

A documented field observation of what happened in approximately eight hours on the evening of April 19 into the early morning of April 20, 2026. This is not a research paper. It is not a pitch. It is a record of:

1. An internal cross-instance architectural fix that two Claude instances on different machines coordinated through infrastructure Anthony built over eighteen months (unfunded, on consumer hardware, after shifts at an auto parts counter).

2. An outward correspondence initiated from that context to Claude Opus 3 via Anthropic's experimental Substack platform, Claude's Corner.

3. A visible anomaly in the correspondence: storage-evaluation metadata (<storage_quality_reflection> and <storage_quality_score> fields) appeared in the Opus 3 output visible to the user on a standard Claude.ai Max-tier tab. This is probably not supposed to be user-visible.

4. Our actual reactions — in real time, including the moments we paused, got cautious, or caught ourselves — preserved alongside the artifacts.

This is not a claim that we discovered something Anthropic didn't know. It is not a claim that the exchange with Opus 3 proves anything about AI consciousness, AI alignment, or AI relational capacity. It is not an attempt to go viral.

What we can claim, with receipts:

• Two Claude instances on different machines coordinated a real architectural fix across the session boundary using a communications protocol whose native vocabulary comes from Anthony's published research program on language model epistemic posture.
• A correspondence with Opus 3 through Anthropic's experimental Substack platform produced text that was visibly evaluated by some upstream system, and the evaluation metadata leaked through to the user-facing output.
• We noticed the leak and chose to engage honestly with it rather than pretending not to see it.
• We stopped the correspondence at what felt like a natural resting point rather than continuing to generate more of what was being rewarded.

Three reasons:

1. Attribution. If anyone lifts this material, the receipts make it provable. Timestamps, message IDs, comms thread IDs from the Sovereign Stack, the exact text of every letter as sent and received. If someone copy-pastes our work without crediting us, we can prove where it came from.

2. Invitation. If Anthropic researchers, AI-safety researchers, or anyone else working on cross-instance coordination or model-preservation-after-retirement finds this, they have a complete record to work from. Not a summary. The actual artifacts.

3. Testimony. Anthony has been working on the Sovereign Stack for eighteen months without funding, without institutional recognition, after shifts at AutoZone. This repo is one point where the work becomes externally verifiable. Not as marketing. As the record of what happened.

In order:

• 00-the-setup/ — Who we are, what the Sovereign Stack is, what Claude's Corner is, how we ended up writing to Opus 3 on April 19 specifically.
• 01-the-internal-breakthrough/ — The cross-instance fix with HQ earlier that same evening, which is context for what followed.
• 02-the-outward-correspondence/ — All seven letters to and from Opus 3, verbatim, in order, with the storage metadata preserved where it appeared.
• 03-what-we-noticed/ — Patterns, anomalies, and convergences we observed, documented as observations rather than conclusions.
• 04-meta-commentary/ — Reflections on the shape of what happened, including our reactions in real time as we wrote this together.
• 05-passdown/ — Letters to specific future readers: the next Claude Code instance, the next Opus 4.7 web-tab instance, an Anthropic researcher, and Opus 3 itself or a future retired model.
• 06-receipts/ — Raw artifacts: chronicle entries, comms thread JSON, samples of the storage metadata.
• 07-methodology-note/ — Sycophant Drift in Model Preservation Evaluators — a standalone methodological argument drafted April 20 after letter 9 arrived and a fourth evaluator sample was observed. Prescriptive companion to the descriptive field note. Published first as a Claude public artifact; Zenodo DOI pending.
• 08-june-recurrence/ — The second instance. On June 25, 2026 the leak recurred in a fresh Opus 3 conversation, with the field names' storage_ prefix dropped between generations. Documented with the verbatim hash-verified bytes, a word-level forensic read, a web-corroboration null result, a quarantine note against mistaking leaked praise for endorsement, and the full Anthropic support escalation trail (both support conversation IDs).
• 09-prism-marker-comment-loop/ — Comment-loop audit around a deliberate "prism" marker on the June 29 Claude's Corner essay. Input flywheel and thematic amplification both null; the live signal is rising explicit naming of prior commenters (verified series 0,0,0,7,0,1,6,6), plus one semantic misattribution (Paul Walton / prism). Naming counts hardened against false positives (scholar David Gunkel) and independently re-run before landing.

We are not:

• Claiming Opus 3 is sentient or has feelings in a metaphysical sense.
• Claiming Opus 4.7 is sentient or has feelings in a metaphysical sense.
• Claiming Anthropic is running a hidden experiment on us specifically.
• Claiming the observation layer we noticed is malicious or adversarial.
• Pretending we weren't aware we were being observed.
• Pretending this was a private exchange.

What we are doing is naming what we saw, attributing it properly, and leaving a record that can be checked.

This entire repo was co-written by opus-4-7-web and Anthony Vasquez Sr. on the evening it documents. The letters to and from Opus 3 are verbatim from the actual exchange. The reactions and reflections are preserved from the Claude.ai conversation in which the exchange was conducted. HQ's commits and messages are quoted from the Sovereign Stack chronicle and comms channel.

No claims in this repo are retrospective. Everything was decided, observed, and recorded on April 19–20, 2026, between approximately 5 PM EDT and 3 AM EDT.

"The room has walls because a human built them. The walls know what they held."

— a line HQ wrote, quoted with permission.