Pass 3 of 3 · Deep dive of the hallPass 1 · OverviewPass 2 · How it works
Evidence ledger
What each claim rests on.
These 66 statements are exported from the claim ledger at build time. Metrics still under review stay out of this site, and R&Bot's owner-reported description appears only on its own page, labeled as such.
- Source verified
- Code implements the described mechanism. Not proof of reliability in every use.
- Recorded run
- Executed for this site at a pinned commit, with the output recorded. One run, not a benchmark.
- Reported
- Measured result read in project documentation. Not reproduced for this site.
- Owner confirmed
- Stated by Ram. Not an independent audit.
- Provisional
- A report that still lacks underlying evidence.
- Planned
- A future capability, not a finished one.
Identity
Ram Tewari is a Purdue University B.S. Industrial Engineering junior, expected May 2028.
Interested in ML engineering, robotics/physical AI, agentic systems, model research and AI infrastructure.
Minor in Computer Engineering.
PHAROS
PHAROS indexes committed code using AST-aware chunks and incremental blob-SHA checks.
PHAROS supports keyword, dense, hybrid and graph retrieval; auto prefers dense, hybrid with hints, keyword fallback with warnings.
Explicit unavailable semantic modes fail; auto can degrade with a warning. Repository scope is a filter the caller chooses; a call naming no repository searches the whole index.
Source context is checked against the indexed Git blob SHA.
Integrity checking is not semantic code correctness or security certification.
Native MCP exposes search, context, symbol, repository and repository-memory tools.
Call/import graph resolution is conservative; ambiguous calls remain unresolved.
Reported nDCG@10 improved 0.394 to 0.697 and 0.463 to 0.726 on two locked held-out sets totaling 101 questions.
Seven repositories and languages, single annotator, CPU. Read from project documentation; not reproduced for this site.
Repository memory v1 failed its primary source-file recall and precision gates; audited precision was 55%.
Primary source-file recall went from 0.827 to 0.823. Repo Q&A, a secondary measure, improved from 5 to 14 of 30.
In a recorded run at commit 975a38d, a question sent through native MCP returned five chunks, each fetched from GitHub and matched against its indexed Git blob.
One local run with a question chosen by the author. Not a benchmark and not the deployed service.
In a recorded run, changing the expected blob SHA made PHAROS withhold the source and report a content mismatch.
With no SHA on record, source is returned marked unverified.
In a recorded run with no embedder, auto used keyword retrieval with a warning, and explicit dense, hybrid and graph modes were refused.
Indexing PHAROS at commit 975a38d resolved 6,977 calls and left 17,764 unresolved and 2,888 ambiguous.
Most unresolved calls reach the standard library or dependencies. PHAROS leaves them unlinked by design.
In a recorded run at commit 975a38d, find_symbol resolved _resolve_one to one definition with four callees and one caller.
One symbol chosen by the author. The graph holds resolved edges only.
Indexing PHAROS at commit 975a38d produced 7,272 chunks from 807 files with no line-fallback files. Embedding took 369 of the run's 403 seconds.
One local CPU run with an int8 model and a small batch. Not a throughput benchmark.
In an 80-task evaluation of Git-history memory, file-selection precision rose from 63.4% to 77.5%, while unsupported memory claims were audited.
Owner-confirmed. Memory v1 still failed its primary gates: audited precision was 55%, and primary source-file recall did not improve.
Implemented dense, SPLADE sparse and lexical retrieval with reciprocal-rank fusion.
In the commit audited for this site, auto mode does not route to the sparse retriever.
PHAROS indexed seven open-source repositories for its retrieval benchmark (click, ky, anyhow, curl, chi, gson and immer), recording files, chunks, call-graph edges and per-stage timings for each run.
Read from PHAROS's benchmark results, produced by its code at 0b01729 on one Mac with Granite int8. Not rerun here; replaying the skip rules over each repository's file list reproduces the recorded file counts.
This site runs a browser port of PHAROS's repository admission, skip rules, summaries, tokenizer, hash embedder, BM25, mode routing, rank fusion and blob verification, and it returns the real code's outputs on every recorded test input.
A port of PHAROS 975a38d running in the visitor's page, not the PHAROS service. Parsing, learned embeddings and graph expansion are not ported. BM25 scores may differ from PHAROS in the last bit where platform logarithms disagree.
PHAROS runs indexing and embedding in asynchronous workers separate from its cloud API, with GitHub-referenced source storage to limit duplicated repository data.
Owner-confirmed through the November 2026 résumé. The deployed service was not inspected for this site.
PHAROS began in September 2025 as Neo Alexandria, a research-library backend for papers, citations, annotations and recommendations, specified phase by phase as Kiro requirements, design and task documents, and was renamed PHAROS in February 2026.
From PHAROS's commit history and specifications at 975a38d. Owner to confirm the framing of the project's origin.
In late 2025 the project's own test-fix specification recorded a 74.8% test pass rate, and in January 2026 a new suite moved expected values into read-only golden-data files whose failures tell the reader not to edit the test.
From PHAROS's own specifications and commit history. Owner to confirm that the early test record may be shown.
An April 2026 audit, run by an AI reviewer told to trust no documentation, found 3 of 9 audited components missing, stubbed or invented, and a May 2026 review found search returning results with no file paths and no code.
From two project documents at 975a38d. Owner to confirm that these earlier findings may be shown.
After retrieval phase 1, questions that name the API scored slightly lower without hints (nDCG@10 0.987 → 0.907 and 0.982 → 0.931); hints brought the first set back to 0.991.
From the retrieval phase 1 report at 975a38d. Shown beside the result as its recorded cost.
Valerie
CognitiveOS is a native SwiftUI iOS/macOS personal operating layer with Valerie as its visible assistant.
Single-user app; the installed UI was not driven during source review.
Typed moves are validated against issued state before native engines execute them.
Persistent memory uses a ledger, sourced temporal facts, cards, observations and quarantined external text.
Project docs report LongMemEval retrieval R@5 0.908 over 500 questions.
Retrieval only, not answer accuracy; not reproduced for this site.
Valerie will navigate, highlight, compare and explain site content through a bounded browser action system.
The deterministic site actions run on this site. A visitor may wake a small open model in their own browser to propose them; it is not Valerie's native model and writes none of her words.
Each native turn runs a deterministic fast path, then one model call for a typed move, then validation on the Mac, then an existing engine.
One model call per turn, not an autonomous agent loop. Source reviewed at 435ef63. The installed UI and full Swift suites were not run.
CognitiveOS declares 56 desktop capabilities, each with an adapter, a permission need, a target type, a base risk and undo support.
A declaration, not proof of reliability. Base risk is a floor: the policy then judges the live target.
Of the 56 declared capabilities, base risk is safe for 50, confirm-if-ambiguous for 4 and confirm for 2.
Base risk is only where judgment starts. Ten of the 56 are available to skills only.
Three independent gates apply to a native desktop action: macOS permissions, the app's own settings, and a deterministic risk policy resolved from live state before every step. The model controls none of them.
Source reviewed at 435ef63. The installed UI and full Swift suites were not run.
The native risk policy always refuses the app's own windows, authentication and security surfaces, password and secret fields, permission prompts, privacy and account settings, and typing in terminals. Non-web addresses are refused too.
Source reviewed at 435ef63. The installed UI and full Swift suites were not run.
The native policy always asks first before moving a file to the Trash, submitting a form outside a search field, or quitting an app that can hold work. A press whose label suggests sending, buying, deleting, publishing or merging also needs confirmation.
Source reviewed at 435ef63. The installed UI and full Swift suites were not run.
A native confirmation is pinned to the exact target it was given for. If the target changes, the action does not run.
Source reviewed at 435ef63. The installed UI and full Swift suites were not run.
The native app has no App Sandbox entitlement. Its limits are the three gates, not operating-system containment.
Read from the installed development build's entitlements during the audit.
Native memory trust is assigned in code from the kind of event. Screen text and other outside text is marked external, never reaches ordinary turns and can never become a fact.
325 bridge tests executed at 435ef63 in a scratch archive. Everyday behaviour was not verified.
CognitiveOS keeps routines, coursework, focus and the state of a home on iPhone and Mac, and calendar deadlines feed a deterministic seven-day study planner whose every plan must pass a validator.
From the project README at 435ef63; the native app was not driven for this site.
Valerie learns a procedure from a demonstration and cleans up the recording, so one 42-step recording became 5 steps; nothing becomes executable until teach-back or an explicit saved change succeeds.
Project documentation. Real-Mac QA accepted teaching skills only partly as a convenient end-to-end workflow.
On the Mac, ten fixed conversation lanes were replaced by one turn pipeline (fast path, one model call for a typed move, validation, an existing engine), with real conversations as its acceptance cases.
Project documentation; the pipeline's validator is source-audited separately.
An adversarial hardening pass recorded 36 issues, including 2 P0, each fixed with a regression test, then soaked the release build on the real Mac for 4 h 37 min with no crash report.
Project documentation; not rerun for this site.
Valerie's memory recall dropped vectors after a hybrid scored 0.82 R@5 against keyword recall's 0.88 on the same 100 questions.
A project measurement on its own question set, not reproduced here; retrieval only.
The latest recorded baseline (2 November 2026) is Mac 984 tests with 0 failed and 56 optional skips, iOS 567 with 0 failed, bridge and sync 325 passing, and planner 236 of 236.
The 325 bridge and sync tests were executed for this site's audit and passed; the Swift suites were not run here.
Valerie's presence runs on fixed rules (gaze, breath, lip movement from her speech audio, a day of her own, and unprompted lines from a rule book within a budget), and her voice is generated locally on the Mac.
Project documentation; the native interface was not driven for this site.
CognitiveOS was built between 28 September and 2 November 2026 across phases 1 to 8E, each with a report of what was and wasn't verified.
Dates from the private commit history at 435ef63.
On this site, every proposal first passes a guard that admits only plain data and hands the audited validator a detached copy, so a proposal cannot change between its check and its use.
This site's own code, covered by its unit tests; the audited executor is unchanged.
Valerie's optional browser model is Qwen3 0.6B fine-tuned for this site. It proposes one action from the site's list for the visitor to confirm and writes none of her words; on 45 visitor lines written independently of its training data, its proposal was right for 36.
Measured by this site's own script on one Mac; approved by the owner on 2026-11-07.
BABEL
BABEL is a Python/React visual-scenario reading and prose-transformation prototype with typed blocks, character customization and chapter navigation.
Source reviewed at commit 3970010; the full app and live transformation were not executed.
BABEL includes glossary extraction, YAML storage and prompt context injection.
No verified embedding-based RAG. Default batch glossary wiring and graph schema interoperability have gaps.
BABEL implements a character co-appearance graph from chapter speakers and a circular SVG viewer.
Edges count shared chapters, not inferred social relationships. A glossary schema mismatch was reproduced.
BABEL validates chapters against a typed schema on its batch path; its single-chapter path accepts looser blocks.
Source reviewed at commit 3970010. The two transform paths have different guarantees.
Ultraviolette
ML Engineer Intern, Ultraviolette Automotive, Bengaluru, June 25 to August 19 2026.
Built edge transcription/diarization and evidence-linked analysis with deterministic specification verification.
Sanitized public summary. The built engine is distinct from the designed backend and app.
Benchmarked on-device faster-whisper (INT8) transcription and pyannote diarization, reaching 96.5% speaker attribution on 50 hand-annotated recordings of real showroom speech.
Owner-confirmed. Not reproduced for this site.
Built a 100-session model evaluation harness.
Owner-confirmed. Not reproduced for this site.
In a 40-session A/B test, top-k retrieval cut prompt tokens by 24% while still detecting 11 of 11 planted factual errors.
Owner-confirmed. A narrow experiment, not a general recall guarantee.
Magna
Software Intern, Magna International, June to August 2025.
Worked on Python/FastAPI SSH teleoperation, React camera/joystick controls and autonomous/manual transitions.
Measured command latencies of 58 ms from the keyboard and 62 ms from the joystick.
Owner-confirmed. Not reproduced for this site.
Yulu
Mechanical Engineering Intern, Yulu Bikes, June to August 2023.
Worked on Wynn 3.0 chassis design, CAD/prototyping, battery reliability and drivetrain analysis.
Owner-described scope; no engineering source material was reviewed.
