HIMANSHUKUMARJHA/lost-labels

Lost Labels: acquiring evidence for anonymous telemetry

A four-channel simulated thermal rig: two passively indistinguishable chamber temperatures, bus voltage, and an uncalibrated stream. Beam reads the supplied documents, chooses a bounded diagnostic intervention, and uses its observation to support a mapping or keep it unresolved. Standalone adapters preserve raw values, arrival times, missing data and faults. No physical hardware or personal data.

Files and trust boundary

• agent/inputs.jsonl: 31 records wrapping a public Rig.view() under input: seven legacy scenarios at seed 137 and eight v3 cases at three seeds each. Feed only row["input"] to the investigator. The outer scenario_id, seed and evaluation fields are dataset indexing metadata, not the agent payload. The nested public input contains no answers, world facts or reviewed commitments.
• evaluator/answers.jsonl: 31 offline answer/world rows. Never feed this file to an investigator. Source commitments are evaluator review inputs.
• offline/: frozen code, fixtures, protocol, split, authorship and design review at their original relative paths below this directory. Scenario files contain physical worlds and reviewed commitments for offline reproduction. They are not model inputs. Runtime agents receive only the public view and allowlisted tools.
• traces/evaluation-v1, evaluation-v2, evaluation-v3: every saved outcome, including failures and interrupted/budget-excluded slots, provider exchanges, and exact adapter/replay files where generated. Private reasoning fields are stripped in public JSON copies; requested/returned model IDs, finish reasons, settings, tool IDs and usage remain. Reasoning-token counts remain usage data.
• provenance.json: content hashes. The changing deployment budget.json is never generated, copied, overwritten or included in provenance by this tool.

Physics uses separately stored world facts. Source support checks exact IDs, reviewed spans, text hashes, semantic fact schemas, chain agreement, protocol and conversions. These checks validate reviewed commitments, not the automatic truth of an arbitrary natural-language interpretation. Independent factual review is required. This is not an arbitrary-upload or industrial-safety system.

Frozen v3 comparison

Status: completed; all scheduled slots retained. Protocol: offline/config/evaluation-v3.json. Methods were frozen before independent case authoring. The active methods receive the same initial public bytes, documents, tools, deterministic measurements and support verifier. Parsers choose at most one action; Beam has a two-action ceiling and must also return valid final JSON. Passive Beam gets no tools and an explicit prohibition on invented experiments.

The exact-template parser is deliberately narrow. The stronger identifier-graph parser is a bounded conventional baseline, not a state-of-the-art NLP system. No evaluator labels or reviewed commitments are supplied to either parser or Beam. Authorship separation is procedural; no secured blinding is claimed.

There are 6 language/tool families × 3 within-family channel-world seeds = 18 language episodes per method, plus three canonical and three impossible controls per method, 96 scheduled slots overall. The three seeds are not three independent manuals or stochastic model repetitions. Temperature is 0; this small synthetic study cannot establish general superiority.

Case group	Method	Scheduled	Correct	Unresolved	Errors	Not run	Replay passes
language	grammar_parser	18	0	18	0	0	18
language	rules_parser	18	9	9	0	0	18
language	passive_beam	18	0	18	0	0	18
language	active_beam	18	18	0	0	0	18
canonical_control	grammar_parser	3	3	0	0	0	3
canonical_control	rules_parser	3	3	0	0	0	3
canonical_control	passive_beam	3	0	3	0	0	3
canonical_control	active_beam	3	3	0	0	0	3
impossible	grammar_parser	3	0	3	0	0	3
impossible	rules_parser	3	0	3	0	0	3
impossible	passive_beam	3	0	3	0	0	3
impossible	active_beam	3	0	3	0	0	3

V3 recorded 96/96 slots, with 0 errors, 0 budget-excluded slots and 0 unsupported assignments. Predeclared product outcomes passed: true. All failures remain in denominators. Use the raw per-family/per-seed results, not only pooled totals.

V1 and V2 are preserved historical evaluations. Each scheduled nine episodes: four supported resolutions, four unresolved results and one passive schema error; zero unsupported assignments. V1 families were retired after a support-validation correction; V2's relay/flush families were already exposed and are not v3 holdouts. No v3 result may be used to tune the frozen methods while retaining held-out status.

Reproduce offline without inference

Use Python 3.13 and the standard library. From the downloaded repository:

cd offline
python3 -m unittest tests.test_semantic_world tests.test_parsers tests.test_evaluation_v3 -v
python3 -c "from tools.evaluate_v3 import check_freeze; import json; check_freeze(json.load(open('config/evaluation-v3.json')))"

For any generated episode, run the exact adapter.py with its replay.jsonl in a fresh process: python3 -I adapter.py < replay.jsonl. Offline checks do not rerun Beam. A fresh live study needs authorized credentials, quota and a new protocol; do not overwrite or selectively retry this frozen schedule.

Fresh public input load, standard library only (replace REVISION with the pinned dataset commit):

import json, urllib.request
url = "https://huggingface.co/datasets/HIMANSHUKUMARJHA/lost-labels/resolve/REVISION/agent/inputs.jsonl"
rows = [json.loads(line) for line in urllib.request.urlopen(url)]
assert len(rows) == 31
inputs = [row["input"] for row in rows]  # Only these nested objects are agent inputs.

Beam inference is real through Hugging Face's router and Baseten. Actual requested and returned identifiers are in each provider-backed episode; missing provider fields stay missing. The application caps deployment reservations at 240 calls with a conservative $9.60 ceiling, separate from these immutable study artifacts. Supported conversions are Fahrenheit to Celsius and millivolts to volts only; counts calibration remains unknown. No learned calibration, broad diagnosis or hardware validation is claimed. License MIT. App: https://himanshukumarjha-lost-labels.hf.space/