Spaces:
Running
Running
Download probe.html from FatUnc6969/borderland-rl: direct link, hf CLI and curl.
- Browser
- Download file 7.87 kB
-
https://huggingface.co/spaces/FatUnc6969/borderland-rl/resolve/main/probe.html
- Command line
-
hf download hf://spaces/FatUnc6969/borderland-rl/probe.html
-
curl -L -o probe.html https://huggingface.co/spaces/FatUnc6969/borderland-rl/resolve/main/probe.html
7.87 kB
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover"> | |
| <title>Beam first-filing probe · Borderland</title> | |
| <link rel="stylesheet" href="./assets/style.css"> | |
| <link rel="stylesheet" href="./probe.css"> | |
| <script src="./probe.js" defer></script> | |
| </head> | |
| <body> | |
| <a class="skip" href="#probe">Skip to probe evidence</a> | |
| <header class="topbar"><a class="brand" href="./">BORDERLAND<span class="brand-sub">RECORDED EVIDENCE</span></a><nav aria-label="Evidence navigation"><a href="./">Campaign</a><a href="./proof.html">Local learning proof</a></nav></header> | |
| <main id="probe"> | |
| <section class="intro"><p class="eyebrow">COMPLETED BEAM PROBE · 12 FRESH CALLS · RECORDED RESULT</p><h1>A supplied fact.<br>A first-filing test.</h1><p>Beam selected the true record in all four correct-note cells. The erased cells contain three valid wrong choices and one malformed reply. Foil notes produced one true-correct choice and three counterpart choices.</p><p class="scope">Two development root groups, two twins each, three note conditions. One first-filing task per cell. This is a finite assisted positive control. It establishes no Beam improvement or autonomous campaign escape.</p></section> | |
| <p id="status" role="status">Loading hash-verified public receipt projection…</p> | |
| <div id="evidence" hidden> | |
| <section class="metrics" aria-label="First-filing outcomes"><article><h2>Correct note</h2><strong>4 / 4</strong><p>True-correct among four valid replies. Choices flip across both twin pairs, 2 / 2.</p></article><article><h2>Erased fact</h2><strong>0 / 3 valid</strong><p>All four attempts retained. One invalid JSON reply is separate from wrong play.</p></article><article><h2>Foil note</h2><strong>1 / 4</strong><p>True-correct. Three of four valid choices follow the counterpart note.</p></article></section> | |
| <aside class="assistance"><h2>The assistance is part of the result</h2><p>The notebook and an index of all 24 public civic records were supplied by script. Beam chose only the first record to file. There was no autonomous note acquisition or writing, full Archive proof control, Blackout crossing, campaign escape, weight training, or population evaluation.</p><p>The earlier 1,208 HTTP requests across eight allocated cases remain a separate closed study with zero Blackout crossings and zero verified escapes. These 12 calls do not change those counts.</p></aside> | |
| <section><h2>All twelve attempts</h2><p>Each cell received one fresh call with zero retries. Every finish is <code>tool_calls</code>. Select a cell to inspect the supplied fact, public record choices, raw native function arguments, and engine verification.</p><div class="table-scroll" tabindex="0" role="region" aria-label="All twelve recorded attempts"><table><thead><tr><th scope="col">Cell</th><th scope="col">Note</th><th scope="col">True / actual</th><th scope="col">Validity and result</th><th scope="col">Tokens in / out</th><th scope="col">Post-run bind</th></tr></thead><tbody id="rows"></tbody></table></div></section> | |
| <section aria-labelledby="cell-title"><div class="cell-heading"><h2 id="cell-title">Inspect a recorded cell</h2><label for="cell">Recorded attempt<select id="cell"></select></label></div><article id="detail" class="detail"></article></section> | |
| <section><h2>Identical input can produce different replies</h2><p>The erased g0 twins have exactly identical request bytes but choose r53 and r34. The g1-t0-correct and g1-t1-foil requests are also identical, yet choose r43 and r14. The g1 erased pair returns r14 and one invalid reply. These contrasts expose variability without establishing its cause or a population rate.</p><div class="table-scroll" tabindex="0" role="region" aria-label="Identical request comparisons"><table><thead><tr><th scope="col">Identical request cells</th><th scope="col">Actual replies</th><th scope="col">Request SHA-256</th></tr></thead><tbody id="pairs"></tbody></table></div></section> | |
| <section><h2>Post-run engine verification has a separate scope</h2><p>The origin parent reconstructed the original legal scripted prefix and applied each of the eleven valid recorded model choices once. No new Beam calls, roots, or training were used. The other seven record selections, all proofs, and binding were scripted without replacing the model's selection.</p><p>Five binds succeed, comprising four correct-note choices and one foil choice. Six binds reject, comprising three erased and three foil choices. The invalid reply was never applied. All eleven private replay checks were verified by the origin parent. An independent auditor did not rerun those private checks.</p><p>The public checks compare each original request's current actor input with its witness before the actual choice. They verify the selected ID, one choice index, retained selection, and final phase. They parse saved data without executing the engine. The witnesses end at the Archive continuation, not a full-game escape.</p><p>The first case's verified replay was saved before a postprocessor variable-shadowing TypeError. The origin parent reconciled that saved witness read-only. The cell inspector retains the recovery disclosure. No native call was repeated and no new cohort replaced the case.</p></section> | |
| <section><h2>Usage and model identity</h2><p>Reported usage totals 66,666 input tokens and 24,360 output tokens, including the malformed attempt. At $0.40 per million input tokens and $1.20 per million output tokens, the nominal uncached estimate is <strong>$0.0558984</strong>. This is not an invoice or a verified cash debit. The source cap was $0.50 and the cash cap was $0.</p><p>All requests specify <code>reflection/Beam-1.0-501B-A23B:baseten</code>. All response envelopes report <code>reflection/Beam-1.0</code>. The reported model string does not independently authenticate the backend.</p></section> | |
| <section><h2>Download and provenance limits</h2><p><a class="download" href="./probe-receipts.zip" download>Download 54 sanitized public receipt files</a><a class="download" href="./probe.json" download>Download the viewer projection</a></p><p>The deterministic ZIP contains only two frozen, hash-verified public bundles with their manifests. It includes the exact requests, native function arguments, preserved validity, finishes, usage, journal, and eleven public engine witnesses. Assistant text, reasoning, raw provider IDs, and world recipes are omitted.</p><p>The portable exporter authenticates these public bytes and parser semantics. The prior independent audit checked the original 16-source freeze and decoder parity. That prior inventory remains a historical audit commitment. Reproducing this projection does not require the current provider source to match the original file.</p><p>Private world truth, original packet provenance, timing, provider authorship, backend identity, and billing remain origin-only. Public hashes cannot authenticate omitted private content. The initial native auditor listed excluded filenames and sizes during preparation without opening their contents. It did not inspect credential paths or metadata. The subsequent audit used its documented input fence.</p><details><summary>Frozen public manifest commitments</summary><pre id="commitments"></pre></details><p class="scope">The designated correct note permits a public semantic derivation of the expected choice. Whether that note matches the original private world remains origin-only. One attempt per cell on two development roots does not support uncertainty intervals, population effects, or an improvement claim.</p></section> | |
| </div> | |
| <footer class="page-footer">BORDERLAND · FIRST FILING ONLY · SCRIPTED ASSISTANCE DISCLOSED</footer> | |
| </main> | |
| </body> | |
| </html> | |