mirror-ledger

Weekly forecast-ledger pipeline for the MirrorDNA public prediction ledger — and, deliberately, the live acceptance test of the MirrorDNA long-horizon stack. EM-007 in predictions.yaml is the pipeline betting on itself: four consecutive clean weekly runs, zero fabricated citations, zero manual repairs.

The pipeline never publishes. Every run ends at a draft plus receipts; Paul reviews and publishes manually.

Published thesis

• The Electric Mind: Why India's AI boom will be built in substations before server halls by Paul Desai. Version 1.2 is the current public edition: an open-access, CC BY 4.0 preprint with a permanent DOI, archived PDF, and six dated forecasts (EM-1201–EM-1206) for scoring through 2032. The paper archive and version history preserve the original publication and frozen EM-001–EM-006 ledger calls.

Keywords: India AI infrastructure, data-centre power, power grid, substations, transformers, transmission, energy storage, investment thesis, forecast ledger.

Research-drop cadence

Mirror Ledger uses a four-day evidence rhythm instead of daily commentary:

• Monday: scored forecast-ledger update.
• Wednesday: international AI infrastructure and sovereign capacity.
• Friday: crypto and digital-asset infrastructure.
• Sunday: reflective AI, agent governance, memory, identity and trust.

The three research-drop runs prepare reviewable, hash-bound drafts under drops/. They do not publish automatically. A public release requires review of the exact brief and its receipt, followed by an explicit content/hash approval. This keeps cadence from outrunning evidence quality.

Weekly run (Mondays 12:30 IST, one shot, never a poller)

scripts/run_weekly.sh (called by the active Codex schedule; staged LaunchAgent ai.mirrordna.ledger-weekly is recovery-only):

0. State preflight — refresh and validate the canonical MirrorState runtime guard; fail closed before any model call if the proof cannot be written.
1. Gather — one headless subscription-Claude call per status: open prediction, with WebSearch/WebFetch, demanding {claim, url, date, quote_fragment} JSON. URLs are deduplicated before validation; india-ai-power claims must explicitly state their India scope.
2. Validate — validate_urls.py HTTP-fetches every URL. Non-2xx or unreachable → evidence/_rejected.log, never the ledger.
3. Draft — draft_post.py deterministically writes posts/YYYY-MM-DD.md from validated evidence only. It makes no model or tool call, so it cannot invent or fetch sources it was never given.
4. Lint — lint_post.py fails the run on ticker/price/buy-sell language ("buy NTPC at ₹344" kills the run) and on any URL not in the validated set.
5. Attest — attest.py writes receipts/<run_id>.json (sha256 of predictions.yaml + evidence tree + post, host clock), then appends to the MirrorBus lane ~/.mirrordna/bus/mirror_ledger_attestations.jsonl. Receipt-first, transport-second: a bus failure is flagged, never fatal.
6. Commit — ledger: weekly run YYYY-MM-DD (N items, M validated). Push is off by default (PUSH_ON_RUN=0 in config.env); flip to 1 after the first supervised run so commits get public timestamps.
7. Stop. Health receipt at ~/.mirrordna/health/mirror_ledger_weekly.json (ok|fail). Failures also append to the repair log in STATE.md and never auto-retry. Partial gather-call gaps remain visible in the post footer, attestation receipt, health receipt, and STATE.md.

The fabrication guard (four layers)

Layer	Where	Rule
1	gather_evidence.py	no URL → item dropped at intake; model told [] beats invention
2	validate_urls.py	every URL fetched live; non-2xx → _rejected.log
3	draft_post.py	deterministic renderer gets validated items only; no model/tool call
4	lint_post.py	any URL outside the validated set fails the run

A fabricated citation would have to defeat all four before it could even reach the draft that Paul still reviews by hand.

Layout

predictions.yaml   source of truth; humans edit, pipeline only reads
evidence/          EM-XXX.yaml validated items; _rejected.log; _incoming/<run>/ raw audit
posts/             drafted updates (never auto-published)
papers/            manually approved, published long-form research
drops/             citable research-drop method, template and approved briefs
receipts/          attestation receipts per run
scripts/           pipeline (bash entry points + python workers)
tests/             pytest suite incl. poison-URL and finance-lint fixtures
deploy/            staged LaunchAgent plist, MANIFEST patch, executor runbook
STATE.md           last run, blockers, repair log (EM-007 scoring input)

Deviations from the 2026-08-10 handoff (deliberate, all visible)

• LaunchAgent name ai.mirrordna.ledger-weekly, not com.mirrordna.ledger.weekly — the body's hardening standard names agents ai.mirrordna.*.
• State integration: the handoff's ~/.mirrordna/current_state.json and ~/.mirrordna/open_loops.md do not exist on this body. The real conventions are used instead: a health receipt (mirror_ledger_weekly.json), the bus attestation lane, and STATE.md.
• Evidence placement: validated evidence is appended to evidence/EM-XXX.yaml rather than into predictions.yaml's evidence: field, so the pipeline never rewrites the human-owned file (which would destroy comments and invite drift).
• Seeding: EM-007 comes from the handoff. EM-001..006 were transcribed from the six-row forecast table in The Electric Mind — Public Edition v2 found in Downloads and are provenance-commented with source hashes. An exact-value regression test protects the statements, criteria, dates, and probabilities from transcription drift.
• Model pin: headless calls pin claude-opus-4-6 per ~/.mirrordna/policy/model_registry.json, via the governed wrapper /Users/pauldesai/bin/claude, with ANTHROPIC_API_KEY stripped — subscription lane only.

Manual supervised run

bash /Users/pauldesai/repos/mirror-ledger/scripts/run_weekly.sh

Durable scheduler

The active schedule is a local Codex recurring task, Mondays at 12:30 IST, which invokes the exact script above and verifies its receipt, health record, commit, and draft. The staged LaunchAgent remains disabled so two schedulers can never race; its plist is retained under deploy/ only as a recovery path.

Tests

/Users/pauldesai/bin/mirror-run --project /Users/pauldesai/repos/mirror-ledger deps -- uv sync
/Users/pauldesai/bin/mirror-run --project /Users/pauldesai/repos/mirror-ledger verify -- pytest

Unit fixtures cover acceptance tests 2 (poison URL rejected end-to-end) and 4 (lint kills "buy NTPC at ₹344"). Acceptance tests 1, 3, 5 (live dry run, bus fallback, single-fire schedule check) are in deploy/EXECUTOR_RUNBOOK.md.

Definition of working

Four green Mondays. Paul's weekly cost ≤ 15 minutes (read draft, publish, push). EM-007 resolves confirmed on 2026-09-14. Then — and only then — v2 talk.