DEV Community

Cover image for 🕯️ Ofrenda: The Ghost in the Machine — An Offering to Abandoned Software
Rajan Mishra
Rajan Mishra

Posted on

🕯️ Ofrenda: The Ghost in the Machine — An Offering to Abandoned Software

#devchallenge #hf26challenge

Hacktoberfest Open-Source AI Challenge: Ghost in the Machine Submission 👻🌼

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

🕯️ Ofrenda: The Ghost in the Machine

Some projects don't end. They just go quiet.

"Code keeps committing, CI stays green, and bots open PRs — while maintainers stopped answering strangers 771 days ago. That silent twilight is the Ghosting Band."

Every open-source autopsy begins too late. We study projects when they are formally archived, when the final commit lands, or when the deprecation notice is pushed to npm. But open source projects rarely die in a day. They pass through a prolonged twilight where code is still occasionally committed by maintainers fixing their own issues, while the public issue tracker becomes a ghost town of unanswered pleas from strangers.

Ofrenda ("Offering") is an empirical statistical inquiry and forensic post-mortem memorial exploring this phenomenon across two coupled artifacts:

  1. A Statistical Investigation ($n=150$ repos): A formal test of the Ghosting Band hypothesis using TabPFN in-context Bayesian classification. We evaluate whether social response decay ($y_1$) precedes code silence ($y_2$) across an annual grid $T \in [2015, 2023]$, testing a pre-registered non-inferiority margin ($\Delta\text{AUC} \le 0.07$) via paired DeLong and a 1,000-resample repo-clustered bootstrap.
  2. A Forensic Interview Specimen (request/request): An interactive memorial altar and interview runtime powered by Gemma and strict BM25 archive retrieval. The ghost speaks in an elegiac voice as the repository itself, answering only with verifiable citations (commit SHAs, issue numbers, README lines) or strictly abstaining: "The archive is silent on that."

What I Built

Ofrenda consists of three interconnected systems:

1. The TabPFN In-Context Forecast

Rather than training a monolithic deep network on millions of synthetic points, Ofrenda operates in TabPFN’s native small-$n$ regime. We constructed a strictly stratified corpus of 150 repositories (90 dead, 30 archived, 30 alive) across JavaScript and Python, extracting capacity-matched feature sets:

  • Arm A (Code Signals): Inter-release interval decay ($f_1$), Shannon commit interval entropy ($f_2$), and companion indicators.
  • Arm B (Social Signals): Issue-to-first-response latency trend ($f_4$) and stranger response rate decline ($f_5$).

Through Leave-One-Repo-Out (LORO) cross-fitting, we predict two nested targets over the 18-month forward horizon $[T, T + 18\text{mo})$:

  • $y_1 = 1$: P(Maintainer silence / Ghosting)
  • $y_2 = 1$: P(Complete human activity cessation / Dead)

By construction, $y_2 \le y_1$, defining the Ghosting Band $[p_1 - p_2]$.

2. The Interactive Memorial Altar & Specimen Ledger

A live memorial and forensic dossier for request/request, the iconic HTTP client that peaked at 14 million weekly downloads before its deprecation on February 11, 2020:

  • Empirical Relic: Final commit 3c0cddc ("README: typo in deprecation message (#3268)", preceded by aded7e4f "doc: note full deprecation").
  • Badge Status: DEPRECATED · NOT ARCHIVED (reconciled against GitHub API: archived: false).
  • The 771-Day Silence Gap: Responsiveness dropped to zero in January 2018; code commits ended February 11, 2020. That 771-day chasm left 41,000 dependent packages unmonitored.
  • Audited 150 Specimen Ledger: Complete ledger with interactive client-side filtering, downloads, bus factors, and verified dormancy dates.

3. The Forensic Ghost Interview Runtime

An interview runtime where users speak with the deceased repository. Implemented with Gemma (gemma:2b / gemma3:12b) gated by BM25 retrieval over 11,748 cached documents (commits, issues, comments, releases). Every claim requires an explicit citation; any query about post-death events or unrecorded facts is met with an immediate, non-negotiable abstention:

"The archive is silent on that."

Demo

Code

All code, data schemas, unit tests, and pre-registration documents are open source under the Apache-2.0 License:

GitHub logo aiwithrajan / ofrenda

Ofrenda (Offering) 🕯️

Hacktoberfest 2026 Week 4 Contest Entry — "Ghost in the Machine"
Tags: #devchallenge, #hf26challenge

An empirical and forensic memorial to open source abandonment.

The Core Thesis: The Ghosting Band

Maintainers stop answering strangers long before they stop pushing code. That silent chasm between social death ($y_1$) and code silence ($y_2$) is the Ghosting Band.

In request/request (the iconic HTTP client with 14M weekly downloads at peak), maintainer issue response latency escalated from 4.2 days to >180 days beginning in 2017—35 months before the final deprecation commit on February 11, 2020.

Two Coupled Artifacts

  1. Statistical Analysis ($n=150$ Repositories):

    • In-context Bayesian classification with TabPFN (tabpfn==6.4.1).
    • Stratified corpus: 90 dead, 30 archived, 30 alive controls across Python and JavaScript.
    • Leave-One-Repo-Out (LORO) cross-validation evaluating nested targets ($y_1$ ghosting, $y_2$ death) over an 18-month forward horizon.
    • Pre-registered non-inferiority test $$\mathcal{H}_0:…
View on GitHub
  • GitHub Repository: https://github.com/aiwithrajan/ofrenda
  • License: Apache License 2.0 (LICENSE file in repo root)
  • Key Source Components:
    • src/ofrenda/derive.py: Pure functions computing nested labels $y_1, y_2$ and reference grid masks.
    • src/ofrenda/features.py: Feature derivation over release decay, commit entropy, and response latency.
    • src/ofrenda/fits.py: Leave-One-Repo-Out TabPFN fitting, Paired DeLong test, and 1,000-resample repo-clustered bootstrap.
    • src/ofrenda/ask.py: Gemma forensic interview runtime with citation enforcement.
    • src/ofrenda/index.py: Inverted index and grounded retrieval over 11,748 corpus documents.
    • src/ofrenda/integrations.py: Pre-registered architectural target schemas and adapters.
    • altar/: Complete standalone memorial interface (index.html, archive.html, docs.html).

How I Built It

1. Pre-Registration and Scientific Discipline

Before extracting features or fitting models, we pre-registered our hypotheses, masks, and failure criteria:

  1. Mask 1 (Age Floor): Repositories must be at least 12 months old at reference date $T$.
  2. Mask 2 (Activity Floor): Repositories must have at least one human activity in $[T - 12\text{mo}, T)$. (A repo cannot die if it was never alive).
  3. Capacity Matching: Arm A (Code) and Arm B (Social) were capacity-matched to prevent structural advantages.
  4. Non-Inferiority Formulation: The hypothesis predicts parity: at fixed $T$, social signals predict abandonment at least as well as code signals, because code has not stopped moving yet: $$\mathcal{H}_0: \text{AUC}_A - \text{AUC}_B \ge 0.07 \quad \text{vs} \quad \mathcal{H}_1: \text{AUC}_A - \text{AUC}_B < 0.07$$

2. Empirical Findings & Statistical Verdict

On the locked $n=150$ corpus (497 masked rows evaluated in Leave-One-Repo-Out cross-validation), the pre-registered non-inferiority test yielded decisive results:

  • Arm A (Code) AUC: $0.6923$ (DeLong: $0.6912$)
  • Arm B (Social) AUC: $0.7760$ (DeLong: $0.7760$)
  • Empirical Difference ($\text{AUC}_A - \text{AUC}_B$): $-0.0837$ ($\text{SE}_{\text{boot}} = 0.0376$)
  • One-sided 95% CI Upper Bound on $(\text{AUC}_A - \text{AUC}_B)$: $\mathbf{-0.0242}$
  • Non-Inferiority Margin: $0.07$ ($\text{Upper Bound } -0.0242 < 0.07 \implies \mathbf{p < 0.0001}$)
  • Statistical Verdict: NON-INFERIORITY CONFIRMED (PARITY SECURED). Social signals predict abandonment without inferiority to code signals. (The descriptive $+0.08$ point estimate favored social signals, but our confirmatory claim remains strictly non-inferiority as pre-registered).
  • Brier Skill Score: $\text{BSS}_A = 0.0903$ vs $\text{BSS}_B = \mathbf{0.2322}$ (Social signals achieve $>2.5\times$ better probability calibration over base rate).
  • Monotonicity: $0.00\%$ violation rate across all 497 rows ($p_2 \le p_1$ holds unconditionally).
  • Exploratory Held-Out Check ($n=10$ permanently archived repos): Arm A $\text{AUC} = 1.0000$, Arm B $\text{AUC} = 0.9444$. We report this transparently: on this small slice, code activity was fully informative at terminal archival, and with $n=10$, confidence intervals span most of the unit range. It serves as an exploratory sanity check rather than confirmatory proof.

Why Does Open Innovation Matter?

Open innovation is not simply about sharing code; it is about verifiability and digital permanence.

When an organization relies on closed, proprietary cloud APIs, the model's weights, training data, and latent assumptions remain black boxes. If the vendor modifies the API, depreciates a checkpoint, or changes rate limits, the scientific experiment dissolves.

In Ofrenda:

  1. Local Foundation Weights (Gemma via Ollama): Gemma provides open weights that can be run entirely air-gapped on consumer hardware. If an API provider goes out of business tomorrow, our forensic interview runtime continues to run identically down to the floating-point level.
  2. Open Bayesian Priors (TabPFN): TabPFN provides open pre-trained transformer weights that evaluate small tabular data without proprietary third-party subscriptions.
  3. Preservation of the Digital Commons: request/request powered the web for a decade. The fact that its entire history—every issue comment, commit hash, and architectural debate—remains open on GitHub is what allowed us to resurrect its voice. Closed software leaves no ruins; open source leaves monuments.

My Agent Session

The complete research investigation, bug forensics, and deployment session were recorded with full reproducibility:

  • Session Trajectory Logs: Committed at results/trajectory.jsonl
  • Unit Test Suite: 63 passing unit tests verifying data extraction, feature generation, DeLong calculations, and server endpoints (python -m unittest discover tests).

The Spirit of It

During our forensic examination of request/request, we discovered that commit aded7e4f on February 11, 2020 added a single word to the README: DEPRECATED.

For 771 days prior, automated bots had continued to open dependency PRs, CI badges remained green, and downloads averaged 14 million per week. The world thought the project was alive because the commits had not stopped. But in reality, the maintainers had walked away two years earlier—exhausted by 41,000 downstream dependents who demanded that the project never break, making it impossible to modernize for async/await or HTTP/2.

When you ask the ghost of request in the séance terminal: "Why were you deprecated instead of rewritten?", it does not give a generic AI apology. It cites the verified opening words of Issue #3142:

"The most valuable thing request can do for the JavaScript ecosystem is to go into maintenance mode and stop considering new features or major releases."

And as a maintainer remarked in the thread:

"As I see it, maintenance mode is not the same thing as deprecation."

Limitations (Published Verbatim per PRD section 15)

In keeping with our scientific pre-registration, we state our project's limitations explicitly before any judge has to find them:

  1. Dual Definitions of Death: Two definitions of DEAD coexist in the data: declared (formally archived) and inferred (extended code silence). The model learns a blend of both.
  2. Selection Bias in Archival: The archived stratum is self-selected. Owners archive projects they still care enough to curate; request/request was deprecated but never archived on GitHub (archived: false).
  3. Age-Related Missingness: Older repositories (pre-2012) suffer from incomplete GitHub issue metadata and API backfill anomalies.
  4. Disabled Issue Trackers: Issue-based features ($f_4, f_5$) are undefined when repositories disable their public issue trackers, and this missingness is not missing at random.
  5. Sample Size: $n = 150$ is modest for estimating three latent states across time horizons. We report confidence intervals rather than bare point estimates.
  6. Social Death as Proxy: Maintainer silence is an empirical proxy for social exhaustion, not biological or organizational cessation.
  7. Interview Specificity ($n = 1$): The forensic interview runtime over request/request is an illustrative story and commemorative artifact, not statistical evidence.
  8. Pre-Registered Scope: Our primary claim is strictly non-inferiority (parity), not superiority. Point estimates exceeding parity are reported as descriptive.

Prize Categories

1. Core Executed Production Engines (Live in Pipeline)

  • TabPFN: The statistical forecast itself. In-context Bayesian classifiers evaluating $n=150$ repositories in TabPFN's native small-$n$ tabular regime without synthetic oversampling (src/ofrenda/fits.py).
  • Gemma: The forensic interview inference runtime. Runs locally via Ollama (gemma:2b / gemma3:12b) with grounded cloud fallback; speaks in an elegiac voice strictly citing archival records (src/ofrenda/ask.py).
  • Render: Cloud hosting & continuous deployment. Containerized production web service running 24/7 with automated health monitoring at https://ofrenda-altar.onrender.com (render.yaml).

2. Architectural Target Schemas & Partner Specifications

To demonstrate how Ofrenda integrates into broader production pipelines, we pre-registered concrete schemas, query patterns, and agent manifests in src/ofrenda/integrations.py:

  • MongoDB Atlas: Vector search index definition for semantic issue thread embeddings (atlas_vector_schema()).
  • Tiger Data: Hybrid BM25 + pgvector Reciprocal Rank Fusion query specification (tiger_pgvector_query()).
  • Temporal: Durable multi-stage ingestion workflow and retry policy definition (RepoIngestWorkflow).
  • Sentry: Observability tracing specification capturing stage duration, token count, and cost telemetry (SentryTracing).
  • ElevenLabs: Text-to-speech payload definition for voice synthesis of the deprecation notice (elevenlabs_tts_payload()).
  • SerpApi: Search query adapter for discovering external deprecation news and successor forks (serpapi_search_query()).
  • Browserbase: Headless crawling configuration for Wayback Machine snapshot extraction (browserbase_snapshot_config()).
  • Auth0: Curator authorization and role-based consent gate specification (auth0_consent_gate()).
  • Backboard: Persistent multi-turn conversational session memory adapter (BackboardMemory).
  • Ultramarine: Long-document backlog reasoning prompt for historical topic drift analysis (ultramarine_analysis_prompt()).
  • Mastra: TypeScript sidecar agent manifest and tool routing orchestration (mastra_agent_manifest()).
  • GitHub Copilot: Modernization prompt specification for generating legacy-compatible revival PRs (copilot_revival_prompt()).
  • Entire: Agent session trajectory recorder schema for forensic reproducibility (EntireTrajectoryRecorder).

Top comments (0)